Iccap

IC-CAP Modeling Handbook
October 2010
Author Name: Franz Sischka
1
Agilent Technologies, Inc. 2000-2010
3501 Stevens Creek Blvd., Santa Clara, CA 95052 USA
No part of this documentation may be reproduced in any form or by any means (including
electronic storage and retrieval or translation into a foreign language) without prior
agreement and written consent from Agilent Technologies, Inc. as governed by United
States and international copyright laws.
Acknowledgments
UNIX is a registered trademark of the Open Group.
Windows , MS Windows and Windows NT are U.S. registered trademarks of
Microsoft Corporation.
Errata The IC-CAP product may contain references to "HP" or "HPEESOF" such as in file
names and directory names. The business entity formerly known as "HP EEsof" is now part
of Agilent Technologies and is known as "Agilent EEsof." To avoid broken functionality and
to maintain backward compatibility for our customers, we did not change all the names
and labels that contain "HP" or "HPEESOF" references.
Warranty The material contained in this documentation is provided "as is", and is subject
to being changed, without notice, in future editions. Further, to the maximum extent
permitted by applicable law, Agilent disclaims all warranties, either express or implied,
with regard to this manual and any information contained herein, including but not limited
to the implied warranties of merchantability and fitness for a particular purpose. Agilent
shall not be liable for errors or for incidental or consequential damages in connection with
the furnishing, use, or performance of this document or of any information contained
herein. Should Agilent and the user have a separate written agreement with warranty
terms covering the material in this document that conflict with these terms, the warranty
terms in the separate agreement shall control.
Technology Licenses The hardware and/or software described in this document are
furnished under a license and may be used or copied only in accordance with the terms of
such license.
Restricted Rights Legend U.S. Government Restricted Rights. Software and technical
data rights granted to the federal government include only those rights customarily
provided to end user customers. Agilent provides this customary commercial license in
Software and technical data pursuant to FAR 12.211 (Technical Data) and 12.212
(Computer Software) and, for the Department of Defense, DFARS 252.227-7015
(Technical Data - Commercial Items) and DFARS 227.7202-3 (Rights in Commercial
Computer Software or Computer Software Documentation).
2
Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
Manufacturing Integrated Circuits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
Understanding the Semiconductor Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
Process and Device Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
Device Modeling Measurement Tutorial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
Basics of DC and AC Characterization of Semiconductors . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
PlotOption Access Name and Value Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
The Guarding (Triax Cable) Principle for Fast and Accurate Measurements . . . . . . . . . . . . . . . 31
Measurement Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
Measurement Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
Prerequisites . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
Wafer Prober . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
FMAX Cables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
Agilent RF and Microwave Test Accessories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
Adapters, Cables and Connectors - General Connector Information . . . . . . . . . . . . . . . . . . . . 69
Metrology/Instrument Grade Adapters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
Metrology Grade Adapters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
Instrument Grade Adapters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
General Purpose Grade Adapter Selection Guide . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
Adapter Kit Selection Guide . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
Flexible Microcircuit Packaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
Triax Feedthru and Triax-Coax Converters from Trompeter Electronics . . . . . . . . . . . . . . . . . 79
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
Probe Features Affect Layout . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
Layout Rules, Mechanical . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
Layout Rules, Electrical . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
Specific Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
Digital Devices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
Characterization Devices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
DC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
DC Characterization for Device Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
Oscillations during DC measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
How to avoid oscillations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
Monitoring the SMUs by an oscilloscope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
Connecting triax SMU outputs to coax inputs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
The Effect of Big Capacitive Loads . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
SMUs and Capacitive loads - The Effect of Long Cables . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
Miscellaneous Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
Process Control Monitoring (PCM) Measurements versus Modeling Measurements . . . . . . . . . . 115
CV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
CV Characterization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
How do capacitance meters work? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
On-Wafer Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
CV on-wafer measurements - What to do with the unused pins? . . . . . . . . . . . . . . . . . . . . . . 120
3 Contacting Cases - for CV Calibration and Measurement, Applying the Auto-Balancing
Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
Simulations of a Multi-pin Device CV Measurement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
Considerations regarding the signal level and the CV-frequency . . . . . . . . . . . . . . . . . . . . . . 124
Obtainable CV measurement resolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
Miscellaneous Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
CV Measurement And Calibration Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
Network Analyzers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
Measurement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
Displaying Phase Shift . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
Linear Vector Network Analyzer Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
Standalone VNA's . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
DC Biased VNA's . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
Calibration Procedures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
Checking the max. Applicable RF Signal Level . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
Important Note on the Agilent N5250A 110GHz PNA Network Analyzer . . . . . . . . . . . . . . . . . 154
Considering the DC Bias Losses During S-Parameter Measurements . . . . . . . . . . . . . . . . . . . 154
Entering Calkit Data for Connectorized Devices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
Entering Calkit Data for On-Wafer Probes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
Entering Calkit Data for Packaged Devices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
Entering CalKit Data into the Network Analyzer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
Entering SOLT Calkit Data (Agilent PNA-NetworkAnalyzer-specific) . . . . . . . . . . . . . . . . . . . . 162
Entering SOLT Calkit Data HP NetworkAnalyzer specific . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
Entering TRL Calkit Data Agilent NetworkAnalyzer specific . . . . . . . . . . . . . . . . . . . . . . . . . . 170
NWA Calibration in General . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
3
NWA Calibration Procedure For At-Temperature Measurements . . . . . . . . . . . . . . . . . . . . . . . 174
NWA Calibration Verification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
NWA Instrument Options in IC-CAP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
Performing the NWA Calibration from IC-CAP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
Defining Custom Calkits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
Introduction to NWA Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
VNA Cal Kit Manager 2.0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
S-Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
Amplifier Characteristics from S-Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
Input Reflection Coefficient with arbitrary load impedance Z L . . . . . . . . . . . . . . . . . . . . . . . 204
Output Reflection Coefficient with arbitrary source impedance Z S . . . . . . . . . . . . . . . . . . . . 204
Voltage Gain with arbitrary Z S and Z L . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
Stability: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
Power Gain from S-Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
Power Gain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
Available Power Gain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
Transducer Power Gain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
Maximum Available (stable) power amplification Gain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
Maximum Unilateral Transducer Power Gain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
Mason's power gain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
Definition of the Transit Frequency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
Definition of the Maximum Oscillation Frequency fmax: . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
Notes on Estimating fmax vs fT by Extrapolation, and Comparison of fT vs fmax Measurements .208
Considerations About the Validity of Unilateral Gain Calculations . . . . . . . . . . . . . . . . . . . . . . 210
Calculating S-parameters from Voltages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
Converting N-Port S-Parameters to M-Port . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
Converting from 2-PORT -> 3-PORT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
Once these 3-port S-parameters are known, they can be used to calculate the common-base
or common-collector 2-port S-parameters of a bipolar transistor etc. . . . . . . . . . . . . . . . . . . . 213
Converting from 2-PORT -> 1-PORT and its Application for the Q factor of RF-passive
Components . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
Definition of S-Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
Differential S-Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
Calculating differential and Common-mode S-parameters directly from S-parameters . . . . . . . 218
Related to the Q-Factor of Spiral Inductors Operated in Differential Mode . . . . . . . . . . . . . . . 218
Differential 1-Port S-Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
Publications on Differential/Common Mode Z-parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
Interpreting S-Parameter Plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
INTRODUCTION TO UNDERSTANDING S-PARAMETER PLOTS . . . . . . . . . . . . . . . . . . . . . . . . 220
Phase Shift of Inductors and Capacitors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
Attenuation of Resistors Related to S-Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
About the Interpretation of S-Parameters of an Asymmetrical Tee Structure . . . . . . . . . . . . . 231
Interpreting S-Parameter Plots of a Transistor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
Starting Points of Transistor S-Parameter Plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
Interpreting S-Parameters of Strip Line . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
S-PARAMETERS OF SERIAL LOSSLESS DELAY LINES WITH DIFFERENT IMPEDANCES . . . . . . . 237
Q-Factor Calculation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240
Small Signal Versus Large Signal S-Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
S-Parameter Basics for Modeling Engineers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
For Experts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
S-Parameter Impedance Matching (for obtaining max. power gain) . . . . . . . . . . . . . . . . . . . . . 250
Smith Chart Circuit Elements Paths . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
Forbidden Regions of the Smith Chart . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
S-Parameters Measurements for Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
10 Rules for Successful S-parameter Handling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
So, How To Deal With S-Parameters? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266
S-Parameter Utilities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270
Calculation the radiation losses from S-parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270
Understanding S-Parameter Plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
The Smith Chart for Sxx Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
THE POLAR PLOTS FOR THE S12 AND S21 PARAMETERS . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
De-embedding Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
Basics of De-embedding Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 276
Why De-embedding ? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 276
Chip carrier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 276
Converting the S-Parameters to Y, we can strip-off the capacitors . . . . . . . . . . . . . . . . . . . . . 277
At a glance: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277
A note on adding of Y matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277
A typical representation of Matrix Yparas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278
A note on adding of Z Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278
For on-wafer OPEN dummy structures, Yparas looks like this: . . . . . . . . . . . . . . . . . . . . . . . 279
The Lumped Components Values are: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
Physical Interpretation of the OPEN dummy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280
Reminder on the de-embedding of the complete Y matrix . . . . . . . . . . . . . . . . . . . . . . . . . . 280
For on-wafer SHORT dummy structures, Zparas looks like this: . . . . . . . . . . . . . . . . . . . . . . 281
about the de-embedding of complete Y and Z matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
De-embedding of complete Y and Z matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
After so much theory, a look at the reality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 282
Verifying the de-embedding for on-Wafer transistor measurements . . . . . . . . . . . . . . . . . . . . 285
Model The On-Wafer OPEN Dummy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
Model The On-Wafer SHORT Dummy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 286
4
Verify The De-Embedding With Modeling The THRU Dummy . . . . . . . . . . . . . . . . . . . . . . . . . 286
An Example of bad THRU deembedding results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
De-Embedding Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
Try yourself in IC-CAP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
De-Embedding Techniques Tutorial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 288
De-embedding tutorial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 288
About the de-embedding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 288
MATRIX MANIPULATIONS IN DETAILS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289
DE-EMBEDDING OF SERIES PARASITICS (Z-MATRIX) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289
DE-EMBEDDING OF PARALLEL PARASITICS (Y-MATRIX) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290
DE-EMBEDDING BY SUBTRACTING THE Z AND Y MATRICES . . . . . . . . . . . . . . . . . . . . . . . . . 290
HOW TO DETERMINE THE SEQUENCE OF Y AND Z MATRIX DE-EMBEDDING . . . . . . . . . . . . . . 291
DE-EMBEDDING OF LOSSLESS 50 Ohm DELAY LINES (S-MATRIX) . . . . . . . . . . . . . . . . . . . . 292
DE - EMBEDDING USING ABC - MATRICES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 292
DE - EMBEDDING OF CHAINS OF ABC - MATRICES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
DE-EMBEDDING OF LOSSY DELAY LINES WITH GENERAL CHARACTERISTIC IMPEDANCE . . . . 293
ft Modeling of a Transistor Affected by Parasitics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 296
An introduction to de-embedding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 296
Interpreting OPEN-Dummy Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 298
Effect of R1, C1 and R2, C2 on Sxy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 298
Effect of R12 and C12 on Sxy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 298
Effect of R3 on Sxy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 298
Effect of L3 on Sxy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 298
QUIZ: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299
Open Dummy High Frequency Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299
How the Silicon Substrate and the Silicon Oxide parameters affect the 'knee' in the OPEN
Dummy Structure S11 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301
A study on the effect of the oxide loss tangent and the substrate conductivity . . . . . . . . . . . . 301
The signal pad of the OPEN sees basically a sequence of SiO2 > silicon substrate > SiO2, i.e. a
circuit schematic like below . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302
Interpreting the OPEN Sxx related to substrate and oxide params . . . . . . . . . . . . . . . . . . . . . 302
Twoport Matrices of Basic Schematics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303
Twoport Matrix Conversions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
H to Y Conversion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
Y to Z and Z to Y Conversion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
S to H Conversion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
Matrix Conversions at a Glance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
TEE to PI Conversion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306
Twoport Matrix Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 308
Terms and Definitions on Series and Parallel Circuits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 308
Twoport Matrix Signal Flow Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 308
Verifying the De-embedding Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 310
Modeling of the Open Dummy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
Vector Network Analyzer - Basics for Modeling Engineers . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312
Measurement Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 314
Vector Error Correction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 314
What to do for Vector-Error Correction? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 314
Practical Considerations before starting with NWA Measurements . . . . . . . . . . . . . . . . . . . . . 315
SOLT calibration for Ground-Signal-Ground probes (G_S_G) . . . . . . . . . . . . . . . . . . . . . . . . 315
Other Calibration Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 316
Verify the Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 316
Verify the Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 316
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 316
Nonlinear Network Analyzers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317
Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317
Nonlinear De-Embedding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 318
2-Step Nonlinear De-Embedding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 318
3-Step Nonlinear De-Embedding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319
Nonlinear RF Measurements and Device Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321
From DC > CV > NWA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321
Lightwave Analogy to RF Energy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321
Linear Vector Network Analyzer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 322
Considering a 2-port: stimulated at port1, then at port 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . 322
S-Parameters = linear theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323
Beyond S-Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324
Extending linear to RF-power-dependent signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324
Linear versus nonlinear behavior . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 325
Phase Reference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 328
Phase: The critical element . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 328
NVNA measurement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 328
What is Harmonic Balance Simulation? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 331
X-Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335
What are HOT S-Parameters? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 336
Ideal X-Parameter Extraction Steps (Perfect match, no source harmonics) . . . . . . . . . . . . . . . 337
X-Parameter Measurement Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 338
X-Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 338
Intermezzo: NVNA X-Parameter Display . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 338
Intermezzo: Number of X-Parameter measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 339
Example of an X-Parameter MDIF file, created by the NVNA . . . . . . . . . . . . . . . . . . . . . . . . . 339
Simulating with X-Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 340
ADS:X-Parameters become the PHD Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 340
PHD Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341
Performing the NVNA on-wafer Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343
5
NVNA Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343
NVNA Calibration Plane . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 344
Importing the NVNA (DC, i1, v1, i2, v2) mdif data file into IC-CAP . . . . . . . . . . . . . . . . . . . . 344
Case Study Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345
Notes about De-Embedding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 346
Calculate the Fourier components of the vx and ix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347
Handling NVNA measurements and standard device modelslike PSP or Hicum, Gummel-Poon,
Angelov etc. in IC-CAP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 348
A possible Modeling Sequence: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 348
The IC-CAP Nonlinear RF Tutorial ModelFile (1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 348
Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 350
Modeling Tips and Tricks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 350
vDS vGS in the time domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 350
Time Domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352
Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352
TDR Measurements Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353
Modeling the Complex Load Out of TDR Plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 354
About the Context Between Time and Location . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 355
TDR Calibration Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 356
Calibration Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 356
TDR Plots Tutorials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 358
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 358
Expected Reflectograms of Typical Parasitic Components . . . . . . . . . . . . . . . . . . . . . . . . . . . 358
TDR Plots for Experts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363
Differential TDR Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365
Summary Differential TDR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365
Summary Differential TDT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365
Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 366
Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 366
Noise Measurement Golden Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367
Verifying the Measurement Data of the IC-CAP 1/f Noise Toolkit . . . . . . . . . . . . . . . . . . . . . . 367
A Note on Data Consistency Checks for MOS Transistors . . . . . . . . . . . . . . . . . . . . . . . . . . . 372
Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 372
Spectrum Analyzer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 375
Data Verification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 376
Data Verification and Data Consistency Checks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 376
Prologue . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 376
DC vs. DC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 376
DC Measurements vs. Spar DC Bias . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 377
Spar DC bias data mirrored into foutput . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 377
CV vs. Spar converted to CV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 378
Spar Starting Points vs. Existing DC Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 378
Data Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 382
Measurement Standardization and Data Consistency Checks . . . . . . . . . . . . . . . . . . . . . . . . 382
Required Naming Conventions for the concept of separated measurements and extractions . . . 382
Project Flow Suggestion for Reusable Model Files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383
Examples for Data Consistency Checks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 384
Advantages of Standardization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 385
Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 386
Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 386
Curve Fitting Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 387
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 387
Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 387
Synthesizing Measurement Data for Verification of Extraction Methods . . . . . . . . . . . . . . . . . . . 388
Regression Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 389
Mathematical Basics of Curve Fitting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 389
Linear Curve Fitting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 389
General Formula for Regression Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 390
Application: Calculating the quadratic polynomial parameters from measurement data . . . . . . 391
Applying Linear Regression Analysis to Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393
DC curves: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393
General CV modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 394
CV modeling with MJ=const, but including offset capacitance C offs . . . . . . . . . . . . . . . . . . . 395
Total CV modeling including offset capacitance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 395
Regression Analysis for S-Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 396
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 396
Direct, Visual Parameter Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 398
General Modeling Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 401
Diode DC modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 401
MOS Gate Oxide Capacitance (simplified model) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 402
Data Interpolation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 404
Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 405
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 405
Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 405
History of Circuit Simulators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 406
The origin of Berkeley SPICE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 406
The origin of ADS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 406
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 406
Simulator Interface of IC-CAP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 407
How IC-CAP interfaces with the simulators. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 407
Netlist Syntax for ADS and SPICE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 410
Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 410
6
Comparing UCB Spice and ADS Netlist Syntax . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 411
Table 1: summary of the most commonly used components . . . . . . . . . . . . . . . . . . . . . . . . . 411
Common Parameter List . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 412
How to have a common Parameter List for Spice and ADS . . . . . . . . . . . . . . . . . . . . . . . . . . 412
UCB SPICE Simulators Netlists (detailed) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 413
User-defined Model Equations in Berkeley SPICE 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 414
ADS Simulator Netlist Syntax (detailed) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 415
Getting a list of all ADS components/models: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 415
User-defined Model Equations in ADS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 417
Using ADS sdd's (Symbolic defined devices) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 417
Using Verilog-A . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 417
Netlist Syntax for ADS and SPICE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 420
CV and S-Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 421
How to Obtain CV Curves from SPICE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 422
Calculating CV Curves from Complex Currents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 422
How to Obtain S-Parameter Curves from SPICE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 423
Calculating CV Curves from Complex Voltages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 423
Harmonic Balance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 427
ADS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 427
What is Harmonic Balance (HB) Simulation? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 427
What happens when a nonlinear harmonic balance simulation is executed? . . . . . . . . . . . . . . 427
Nonlinear High Frequency Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 429
Harmonic Balance Simulations for Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 431
History . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 435
History of Circuit Simulators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 435
Device Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 436
Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 436
Diode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 437
Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 437
Modeling a Diode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 438
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 438
DC Characterization Parameters IS, N AND RS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 439
Determination of the DC parameters IS and N . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 439
Parameter RS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 440
CV Characterization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 440
HF Modeling: Parameter TT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 442
Model Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 443
Details of the Berkeley SPICE Diode Modeling for Experts . . . . . . . . . . . . . . . . . . . . . . . . . . 443
SPICE Diode Model Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 444
Publications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 445
ADS Junction Diode Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 446
INSTANCE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 446
MODEL CARD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 446
Step Recovery Diode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 447
Standard Diode Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 449
Non Quasi-static Diode Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 450
Parameter Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 450
Model Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 451
Modelling Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 452
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 454
Transistors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 455
Bipolar in General . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 456
Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 456
Advanced Bipolar Transistor Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 457
Why BJT Modeling Is Still Hot? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 457
What is a Compact Model? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 457
Why Standard Compact Model? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 457
History of BJT Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 457
Example of Model Evolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 457
The Spice Gummel-Poon Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 458
SGP-Model: DC Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 458
SGP-Model: normalized Base Charge q b . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 458
SGP-Model: Parasitic Resistors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 459
SGP-Model: Capacitors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 459
SG-P Model: Additional Phase Shift and Temperature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 460
The Spice Gummel-Poon Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 460
SGP: Parameter Extraction Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 460
SGP Extraction: Space Charge Capacitors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 461
SGP Extraction: Parasitic Resistors: RE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 461
SGP Extraction: DC biasing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 462
G-P Extraction: Early-Voltages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 462
G-P Extraction: Gummel-Poon Plot Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 462
SGP Extraction: TFF Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 462
SGP Extraction: Final Parameter Tuning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 463
SGP Extraction: Final Result - Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 463
Bipolar Transistor Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 464
History of Bipolar Transistor Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 464
The Gummel-Poon Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 464
GP-Model: normalized Base Charge q b . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 464
GP-Model: normalized Base Charge q b . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 465
GP-Model: DC Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 465
GP-Model: Parasitic Resistors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 465
7
GP-Model: Capacitors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 466
G-P Model: Additional Phase Shift and Temperature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 466
GP-Model: Model Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 466
G-P Extraction: Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 467
G-P: Parameter Extraction Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 467
G-P Extraction: Space Charge Capacitors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 467
G-P Extraction: Parasitic Resistors: RE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 467
G-P Extraction: Parasitic Resistors: RBM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 467
G-P Extraction: DC Biasing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 468
G-P Extraction: Early-Voltages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 468
G-P Extraction: Gummel-Poon Plot Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 469
G-P Extraction: S-Parameter De-embedding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 469
G-P Extraction: S-Parameter: Data Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 469
G-P Extraction: Data Management ASCII File . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 470
G-P Extraction: RB Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 470
G-P Extraction: TFF Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 470
G-P Extraction: Final Parameter Tuning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 471
BSIM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 472
Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 472
BSIM3.3v2 - Parameter Extraction Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 472
Fully Scalable BSIM4 Modeling From DC to RF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 476
GaAs Transistor Models in General . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 490
General Transistor Modeling Strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 491
Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 491
Data Consistency Checks DC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 491
Data Consistency Checks S-Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 491
Wrap-up: The applied checktools and verification model files: . . . . . . . . . . . . . . . . . . . . . . . 492
After all the data have been checked, we can finally start the modeling . . . . . . . . . . . . . . . . . 492
Example: CV Modeling Gummel-Poon . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 493
Transfer Curve(s) Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 495
Output Characteristics Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 496
S-parameter Fitting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 497
Overview of the Hicum Level 2 model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 500
Forewords . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 500
What is compact modeling? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 500
Need for accurate models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 500
HiCUM can help circuit design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 500
HiCUM is also better for Scaling and Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 502
Statistics can help designers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 502
HiCUM Core Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 502
Transfer current . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 502
Base Current . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 503
Depletion charges and capacitances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 503
BE Capacitance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 504
Transit Time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 505
Transit Frequency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 505
Transit Frequency and Critical current . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 506
Critical current ICK . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 506
Scalable Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 507
Extraction Example: Junction Capacitance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 510
Example results: f T (30GHz Process) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 511
Statistical Modeling: The SMACH Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 511
Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 512
Statistical Modeling Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 512
Check Correlation between PCMs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 513
Statistical Modeling Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 513
Learning the Angelov Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 514
IPK0 Current for maximum transconductance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 514
VPKS Gate voltage for maximum transconductance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 514
P1 Polynomial coefficient for channel current . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 514
LAMBDA Channel length modulation parameter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 515
ALPHAR Saturation parameter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 515
ALPHAS Saturation parameter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 515
UCB MOS Level-3 MODEL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 517
MOS Transistors for Dummies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 522
MOS Modeling for Dummies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 523
PSP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 525
Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 525
Background of the PSP Model Extraction Package . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 526
Introduction to PSP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 526
Internal model structure of PSP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 526
Parameter extraction principle for scalable models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 527
Binning in PSP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 529
Example extraction flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 529
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 530
Global PSP Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 534
PSP Model Extraction Package . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 535
Key Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 535
PSP Extraction Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 535
Extensions to the PSP Model Extraction Package from the Modeling Bench . . . . . . . . . . . . . . . 537
PSP Parameter Extraction Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 541
8
CV Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 541
Diode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 542
Individual Transistor DC Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 543
The Curtice Mesfet Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 546
Common parameters for all Curtice model levels: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 546
Diode model parameters: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 546
Piecewise linear model parameters: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 546
Level 1 parameters: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 546
Level 2 parameters: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 546
DC Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 547
Dynamic performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 548
Publications for the Curtice model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 550
The Gummel-Poon model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 551
Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 551
AC Small Signal Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 552
Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 552
Appendices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 553
Linear Curve Fitting: Regression Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 553
Verifying the quality of extraction routines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 554
Direct visual parameter extraction of BF, ISE and NE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 554
RBM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 556
VAF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 556
IS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 557
NE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 558
ISE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 558
IKF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 558
Calculating IKF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 559
BF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 559
Calculation of h21 of the Gummel-Poon Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 560
Simplicification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 562
Modeling the output characteristics quasi-saturation effect with Gummel-Poon . . . . . . . . . . . . 562
Gummel-Poon Bipolar Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 564
Structure of this Manual . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 564
About This Manual . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 564
Gummel-Poon CV Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 566
The Gummel-Poon Capacitor Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 566
Modeling the Space Charge Capacitors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 566
Measuring and modeling the Base-Emitter capacitance. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 567
The Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 567
Determination of the CV Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 567
What to do in IC-CAP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 568
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 569
Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 569
Operating Modes of the Bipolar Transistor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 569
The Gummel-Poon Equivalent Schematic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 569
The Gummel-Poon Model Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 571
Proposed Global Extraction and Optimization Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 575
Limitations of the Gummel-Poon Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 577
Ohmic Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 577
Forward DC Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 577
Reverse DC Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 577
AC Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 577
Temperature Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 577
Applications in Integrated Circuits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 577
Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 577
Modeling the Base Resistor rBB' . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 578
Extraction of RB, IRB and RBM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 578
Modeling the Parameter XCJC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 582
A final remark on the S-Parameter Extraction and Optimization Strategy . . . . . . . . . . . . . . . . 582
Modeling the Resistors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 584
Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 584
Modeling the Emitter Resistor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 584
Extraction of RBM from DC measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 584
Modeling the Collector Resistor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 585
An alternate method to calculate the ohmic parasitic resistors from s-parameter
measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 586
Modeling the Temperature Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 588
Non Linear DC Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 589
Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 589
Modeling the Output Characteristic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 590
Modeling the Collector Current . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 592
Modeling the Base Current . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 594
Modeling the Current Amplification at High Current . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 596
Extraction of the remaining Reverse Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 599
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 600
About Gummel Poon User Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 600
The IC-CAP model file "GP_CLASSIC_NPN.mdl" features . . . . . . . . . . . . . . . . . . . . . . . . . . . 600
Organization of Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 600
Publications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 601
Paras.resistor modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 601
Transit Time Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 602
Modeling the Diffusion Capacitor CDBC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 602
Pre-considerations concerning the model equation: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 604
9
Performing the measurement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 604
Extracting the parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 605
How to proceed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 605
Extraction of VTF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 606
What to do in IC-CAP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 606
How to optimize the S-parameter Setup in IC-CAP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 607
Extraction of PTF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 607
Extraction of TR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 607
YParameter Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 609
Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 609
Modeling of Transistors in the Operating Point . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 610
Modeling with Parameter Tuning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 613
Developing A Small Signal Model For Mos Transistors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 615
Publications: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 616
S2Y-Parameter Transistor Modeling GUI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 617
How to Proceed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 617
How to use the GUI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 618
What if the locus curve becomes inductive !! . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 619
Modeling Result Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 620
Testing S-Parameters for Small Signal Transistors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 621
Converting S-parameters into a small signal schematic of a bipolar transistor: . . . . . . . . . . . . 621
Converting S-parameters into a small signal schematic of a bipolar transistor, . . . . . . . . . . . . 621
Converting S-parameters into a small signal schematic of a MOS transistor: . . . . . . . . . . . . . 622
VBIC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 623
Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 623
Introduction to VBIC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 624
VBIC Background Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 627
VBIC Model History Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 627
SDD - Definition of the VBIC 1.1.4 Model in MDS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 627
Information on the VBIC Code, Release 1.2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 630
Equivalent Circuit Network for VBIC 1.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 631
Default Parameters rev.1.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 631
Recapitulating the Gummel-Poon Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 633
Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 633
DC Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 633
Resistors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 634
Dynamic Behavior . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 634
A Quick Tutorial on the Gummel-Poon Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 635
Gummel-Poon Model Parameter List . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 637
The VBIC Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 639
The Normalized Base Charges q b and q bp . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 639
Space Charge Capacitors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 640
DC Performance of the Main NPN Transistor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 640
DC Performance of the Parasitic PNP Transistor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 641
Quasi-Saturation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 642
Dynamic Behavior . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 643
Space Charge Capacitors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 643
Diffusion Charge Capacitors of the NPN Transistor: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 643
Additional Phase Shift . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 643
Self-Heating Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 644
VBIC Model Parameter List . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 644
Comparing the VBIC and Gummel-Poon Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 645
VBIC Modeling Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 648
CV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 648
DC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 648
DC Finetuning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 648
S-parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 648
S-Parameter Finetuning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 648
Measuring and Extracting the VBIC Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 649
Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 649
Parameter Extractions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 649
Space Charge Capacitances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 649
Parasitic Resistors From DC Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 650
Early Voltages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 652
Diode Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 653
IS, NF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 654
IKF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 656
Output Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 656
WSP (DC Current Distribution Between Main NPN and Parasitic NPN) . . . . . . . . . . . . . . . . . . . 657
AVC1 and AVC2 (Avalanche Effect Modeling) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 658
RTH (Output Characteristics Thermal Effect Modeling) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 658
S-Parameter Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 659
Parametrized Modeling vs. Transistor Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 662
Regression Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 664
Mathematical Basics of Curve Fitting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 664
Linear Curve Fitting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 664
General Formula for Regression Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 665
Application: Calculating the quadratic polynomial parameters from measurement data . . . . . . 666
Acknowledgements, Web info, and Publications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 668
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 668
Web Info . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 668
Publications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 668
RF Passive Linear Components . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 669
10
Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 669
Modeling of a Chip Capacitor up to 26 GHz . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 670
Modeling of a Dummy Chip on a Wafer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 671
Physical Interpretation of the Equivalent Schematic: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 673
Modeling the Second Order Effects of an Open Dummy Structure . . . . . . . . . . . . . . . . . . . . . 673
Modeling Frequency Dependent Losses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 675
Time Domain: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 675
Frequency Domain: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 675
Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 675
Time Domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 676
Test Fixture Modeling in Frequency Domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 678
Modeling of Spiral Inductors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 680
De-embedding the measured data before modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 680
1-port modeling of a spiral inductor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 680
2-port modeling of a spiral inductor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 683
Calculating the Q-factor of spiral inductors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 686
Modeling of a 3.5mm Connector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 687
Modeling Vias . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 688
Package Modeling Workshop . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 691
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 691
Limitations with Standard Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 691
Two Types of Characterization Measurements Required . . . . . . . . . . . . . . . . . . . . . . . . . . . . 691
Modeling of a SMT Resistor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 697
Developing Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 698
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 698
Developing the Equivalent Schematic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 698
Modeling of a SOT-23 Package . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 703
The SOT23 package . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 703
S-Parameters of Strip Lines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 707
Strip Lines Tutorial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 707
Modeling a lossy strip line from S-parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 707
Modeling a Strip Line Transition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 710
Testing S-parameters For Typical Passive Circuits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 711
Test Fixture Modeling in Time Domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 712
Modeling of a TSOP44 Packagae . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 717
TDR-related issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 717
Low Frequency (1/f) Noise Modeling for Semiconductors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 725
Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 725
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 726
Noise Term Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 726
Types of Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 728
Thermal Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 728
Shot Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 728
Generation-Recombination Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 729
1/f Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 729
Noise Models in Simulation Programs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 730
Resistors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 730
Inductors, Capacitors, and Lines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 730
Diodes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 730
Bipolar Transistors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 730
Junction FETs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 731
MOSFETs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 732
MESFETS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 733
Low (1/f) Noise Measurement Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 734
1/f Noise Measurements and Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 736
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 736
Noise Mechanism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 738
Noise Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 739
1/f Noise Measurement Toolkit in IC-CAP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 742
1/f Parameter Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 744
Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 745
AF, KF and BF/EF Noise Parameter Extraction and Verification . . . . . . . . . . . . . . . . . . . . . . . . . 751
Bipolar Transistors: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 751
MOS Transistors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 753
Noise floor information on the toolkit measurement setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . 757
Parts list for 1/f noise toolkit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 757
Explanation of Spice circuit deck for 1/f noise simulations . . . . . . . . . . . . . . . . . . . . . . . . . . 758
Thermal Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 761
Modeling Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 761
Static and Dynamic Thermal Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 761
On Target Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 767
Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 767
Using PCM Data for Device Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 768
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 768
Application Scenarios . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 768
On Target Modeling working . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 770
IC-CAP Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 770
Supported simulators and models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 774
Applying PlotOptimizer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 775
Tutorial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 777
How to start the Target Modeling Tool . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 777
Loading a Predefined Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 777
11
Exploring the Example:Device Table . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 778
Exploring the Example:Available PCM data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 779
Exploring the Example:Simulation Templates for PCM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 779
Understanding the Example:Creating a subset of devices . . . . . . . . . . . . . . . . . . . . . . . . . . . 780
Understanding the Example:Viewing subset of devices . . . . . . . . . . . . . . . . . . . . . . . . . . . . 780
Understanding the Example:Defining I-V Setups from PCM data . . . . . . . . . . . . . . . . . . . . . . 781
Understanding the Example:THE PLOTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 781
Understanding the Example:Using PlotOptimizer to tune the parameters . . . . . . . . . . . . . . . . 781
Understanding the Example:Report Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 781
Direct Measurement Data Reduction for Scalable Parameter Extraction . . . . . . . . . . . . . . . . . . . 783
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 783
New Approach for Direct Extraction of Geometrically Normalized Device Parameters . . . . . . . . 783
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 786
Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 787
Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 787
Parametric Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 788
Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 788
Measurement Related Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 789
About The Difficulties To Identify The Golden Die Using Measurement-Related Statistics . . . . . 791
Model Parameter Related Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 793
Principal Component Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 793
Factor Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 794
Practical Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 795
Description of the Steps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 797
Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 800
Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 801
Publications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 801
Non-Parametric Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 802
Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 802
Introduction to Non-Parametric Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 803
Applying Non-Parametric Statistics Directly To Measured Data . . . . . . . . . . . . . . . . . . . . . . . 807
Literature about non-parametric statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 808
Identify the 'golden device' and the 'limits devices' . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 809
Conventional Statistics Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 809
What are Corners, Boundaries, and Monte Carlo statistical models, obtained with conventional
statistics? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 809
Introducing The Non-Parametric Statistics Concept . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 809
What means typical component and limits components identification ? . . . . . . . . . . . . . . . . . 809
Non-Parametric Statistics Works on Gaussian Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . 810
Non-Parametric Statistics Works on Non-Gaussian Distributions . . . . . . . . . . . . . . . . . . . . . . 810
Conventional Statistics vs. Non-Parametric Statisticswrap-up . . . . . . . . . . . . . . . . . . . . . . . . 810
How does 'Non-Parametric' Statistics work? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 810
The Result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 812
An Example: Diode with measurements from DC to S-parameters . . . . . . . . . . . . . . . . . . . . 813
Modeling Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 815
A simple method to identifythe 'golden device' and the 'limits devices'has been proposed. . . . . 815
Non-Parametric Boundary Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 816
Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 816
Circuit and System Design Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 816
What are Corners, Boundaries, and Monte Carlo? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 816
Non-Parametric Boundary Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 816
Secondary Inputs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 817
Outputs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 817
Normalize the Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 817
Estimating Density Using Nearest Neighbors Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 818
Nominal and Boundary Point Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 818
Spatial Diversity---An Illustration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 818
Validating the Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 819
2-D Joint Gaussian Distribution with Nominal and Boundary Points . . . . . . . . . . . . . . . . . . . . 819
Hypershell in 320-D Spherical Gaussian Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 819
Conditional Density Histogram on 4-D Bimodal Gaussian Distribution . . . . . . . . . . . . . . . . . . 819
Applying the Tool . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 820
2-D Analysis of BSIM3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 820
Analysis of Diode Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 820
Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 820
Improving The Manufacturability Of Electronic Designs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 822
12
Contents
Manufacturing ICs (iccapmhb)
Measurement for Device Modeling (iccapmhb)
13
Manufacturing Integrated Circuits
Understanding the Semiconductor Process
The semiconductor manufacturing process comprises:
1. Creating a CMOS Inverter Circuit

2. Main Steps of the Semiconductor Process
3. Semiconductor Process Flow
Creating a CMOS Inverter Circuit

The process of making a CMOS inverter circuit is a complicated one. Figure 1 illustrates
an example of a simple CMOS inverter circuit.
Figure 1 : A Simple CMOS Inverter Circuit
See the Cross section view in Figure 1 . The process of making a circuit consists of many
layer depositions, doping, and etchings. Different elements can be created by slightly
varying the parameters and device geometries.
Note
The real devices are more complex than the example cited above and illustrated in Figure 1 .
A fairly advanced CMOS process may contain over 300 steps using around 20 mask sets,
and may take more than 2 months from wafer start to finish. Expenses for producing one
wafer by a high volume process can exceed several $1000.
Main Steps of the Semiconductor Process

The semiconductor process consist of combining and repeating six main steps which are
described in detail in the following section. These steps are complex and interact with each
other. Refer Table 1: for further details.
Step Semiconductor Description

process
1. Photolithography A very important process, which defines feature (transistor, diode etc.) size and
geometry.
2. Oxidation/Diffusion This process creates insulating layers, or resistive (doped layers)at high
temperature.
3. Ion Implant This process is more precise than the diffusion for doping layers with very tight
control.
4. Thin-film deposition This process consists of low temperature deposition which protects the previous
processes.
5. Plasma Etch It is a precise way of removing unwanted portions of layers
6. Metallization This process provides device interconnection using up to 6 layers. It is also used
for making MIM capacitors, spiral inductors etc.
Semiconductor Process Flow
14
The following sequential steps describe the semiconductor process flow:
1. Designing the Recticle Mask Set: The recticle mask set is designed to transfer the
design to the wafer. This concept is similar to traditional photography negative.
2. Oxidation: It creates insulating layers or resistive doped layers by growing an oxide
film at high temperature.
3. Photolithography: It consists of the photo resist layer that creates the mask on the
wafer so that the metal and oxide layers can be formed in the required places. The
resist prevents the layers to be formed in certain places. It can be removed later.
4. Mask Alignment: This process aligns the mask to the wafer.
5. Lithography: It places the mask onto the wafer using the resist layer (Refer
Photolithography). Etching removes the mask resist layer once the deposition or
diffusion step is complete.
6. Oxidation/ Diffusion/ Implantation: This process introduces the ions into
appropriate layers to create the semiconductor material in the exposed areas. The
previous mask, lithography, etching, diffusion, and implant steps are repeated many
times to create many layers of the devices.
7. Metallization: This process is repeated several times to create the conductive
connections between devices and layers and to create some passive components like
MIM capacitors, spiral inductors etc. The metal is usually deposited by evaporation.
Process and Device Modeling

This section provides an overview of the modeling process that is carried out during
manufacturing (the process and device simulations)and for the individual components
such as ADS, Spectre Spice simulations. It also describes how all this is assorted when the
chip is finally designed.
This is a typical scheme for control measurements during a chip production. Curve-tracer-
like measurements are made by DC analyzers and fast spot measurements (single point
measurements) are made by Parametric Testers.
15
The process itself, i.e. the Diffusion and the Epitaxy, can also be modeled using Process
Simulators. Once these simulators are calibrated to the process, the performance of
devices can then be simulated using Device Simulators, which also take the
photolithography into account. Either such simulation results can be imported into IC-CAP
for modeling (using .mdm data files) or IC-CAP performs measurements directly on the
components
16
Slide 10
Now that we understand the semiconductor manufacturing process, and its modeling
aspects, we will look at typical measurements.
Current voltage (or IV) testing is probably the best known type of parametric test. Since it
is also the most important parametric test, we need to understand what kind of
parameters are measured and which processes are evaluated by it. Measured parameters
, like gain and threshold voltage are very popular. However more than just the standard
curves are measured. To extract useful process information, variations of the standard
measurement are often made. Of course the IV measurements relate back to the process
parameters .
The MOSFET (Metal Oxide Semiconductor - Field Effect Transistor) dominates today's
technologies due to its performance (device geometries have shrunk over the past
decades drastically, thus allowing to manufacture MOS transistors with transit frequencies
in the GHz range). Another argument in favor of MOS is its lower cost compared to e.g.
bipolar. Bipolar technology is also still popular for analog and high frequency digital
applications.
Transistor characteristics are critical measurements for the R&D engineer, and here are
several popular curves which are measured. For example, here we have:
IdVd of a MOSFET
IcVc of a bipolar transistor
diode's forward bias
These curves display device characteristics in an active state at relatively high current
levels.
17
Importantly, low level IV sweeps must also be performed to measure leakage, threshold,
gain and off state etc.
Let's look at one example, and see its application. A MOS transistor drain current vs. gate
voltage (IdVg) curve is shown above. This characteristic is of great importance to the
evaluation of semiconductor processes and device design. It allows the extraction of
device gain and threshold voltage.
The lowest current region is called the "sub-threshold region". In this region the MOSFET
is in the off state, where the current should be kept at a very low level to reduce power
consumption and noise levels.
Typically the current is in the fA region, if processes are good. However, if the oxide is
damaged during the process, the current may become much larger. Defects due to
improper oxidation processes can be revealed only by an accurate monitoring of the sub-
threshold current.
The threshold voltage (or Vth) is not only another important device parameter, it is also a
very important parameter for circuit simulation. Vth can be controlled by changing doping
profile and oxide thickness. Alternatively, it can be applied to evaluate processes to
control doping concentration and oxide thickness.
There are several common ways to determine the threshold voltage. This example shows
the saturation region threshold voltage. The IdVg curve is obtained using the
measurement setup shown, and the square root IdVg curve is drawn. This curve is
derivated vs. vg to find the maximum slope. At this maximum slope, a tangential line is
drawn to SQRT(id), and the threshold voltage is then extracted as the X axis intercept of
this tangent.
18
Other measurements performed in a production environment are so-called spot-
measurements.
As an example, instead of measuring the complete idvd output characteristics, only IDLIN
(vg=vg_max, vd=50mV, vb=0V) and IDSAT (vg=vg_max, vd=vd_max, vb=0V) are
measured. Commonly, applying the max. neg. Bulk voltage, IDLIN_BB (vg=vg_max,
vd=50mV, vb=-vBB) and
IDSAT (vg=vg_max, vd=vd_max, vb=-vBB) are measured too.
These high-speed measurements are performed on as many as available devices on the

wafer, and displayed vs. their distribution. This allows to easily- check the performance
and stability of the process- check the variance of the process
Resistivity Continuity/Bridging Leakage/Breakdown Capacitance/ CV IV Curves

Typical Poly resistors Contact resistance Breakdown Capacitance Threshold
Measurement Metal Via continuity voltage CV curve voltage
contacts Inter-metal bridging Oxide damage Drain current
Current Leakage current (On, off)
capability Max Gain
Substrate (Gmmax)
resistors Punchthrough
voltage
Features Film Particles / defects in films Ion implantation Film Transistor
Maintained thickness Incomplete etches damage thickness gain
*Doping Defect densities Isolation-edge Doping Thresholds
concentration Dielectric integrity effects concentration
Film purity Contact integrity Contamination and profiles
Line width Dielectric Carrier
Mask overlay integrity lifetimes
Thresholds
Processes Oxidation Lithography Oxidation Oxidation Lithography
Monitored Deposition Deposition Diffusion Diffusion Oxidation
Ion Etch Ion implantation Ion Diffusion
implantation Metallisation Deposition implantation Ion
Metallisation Metallisation implantation
Etch
Structures
This summary table shows the relationship between typical parametric tests and the
processes monitored by them. The basic test types are: resistivity, continuity/bridging,
leakage/breakdown, capacitance/CV, and DC/IV curves. By using a combination of these
tests, nearly all process aspects can be covered.
There is more information about the tests in the Appendix.
Wafer Map
Different types of analysis can be performed on the results obtained, in order to
understand what is happening with the devices and process.
Here the uniformity of a measurement across the wafer is analyzed by drawing a wafer
map for the Threshold Voltage VTH. Wafer maps for the other PCM measurements
(ID_LIN, ID_SAT etc.) are obtained in the same way.
Correlation between process parameters and measurement results will allow you to find
the best process condition for producing the best end product.
19
The trend chart is critical for capturing potential process problems. Key test parameters
must stay within certain limits to ensure that functional devices are made. Parameter
values near to or over the limits can trigger urgent action for process engineers.
In addition to the standard traditional IV and CV measurements and curves, reliability

testing is very important.
MOS transistors operate with faster and faster speed, and this means that they must
become smaller and smaller, and therefore, operate at lower voltages and currents.
However this also means that they can be much more easily degraded, damaged or even
broken. Additionally, this degradation may happen over long periods of time, even years.
Therefore, lifetime and reliability is a big issue, and the general goal is to ensure a
minimum of a 10 year lifetime of devices.
Reliability uses Parametric tests for evaluation, and therefore parametric instruments are
important for this application.
20
The following are three main areas of concern in Reliability Testing:
Hot Carrier Injection

Oxide Breakdown
Electromigration
We will look at these in more detail on the next few slides, but some of the common
measurement names are also listed here so that you can identify them more easily. These
are tests such as Charge Pumping , TDDB , V Ramp , and SWEAT tests.
When performed on wafer the reliability tests are called WLR , ie Wafer Level Reliability.
Hot carrier injection happens when small gate sizes lead to high electric fields close to the
drain terminal, creating highly energized carriers known as "Hot Carriers". These are
injected into the oxide layer and trapped, causing degradation of the device due to the
trapped charge.
This effect can cause device gain and speed degradation and finally malfunctions.
Evaluation is done by repeatedly applying a DC or AC stress for a defined time, and then
re-measuring the appropriate parameters and plotting the change vs. accumulated stress
.
The trapped charges can also be evaluated using Charge Pumping techniques, applying a
pulse to the Gate. DC analyzers therefore feature a Pulse Generator option with the
necessary functions .
21
Thin Oxide Breakdown is caused by a combination of the thin oxide layer and
imperfections or defects in the oxide itself. The defects cause a higher current to flow,
which then causes further damage, and so the oxide breaks down.
Here the V Ramp test applies an increasing voltage to the oxide and accurately measures
the time and voltage at breakdown. It can also be evaluated using I Ramp (current ramp),
or TDDB which is a constant current or constant voltage Time Dependant Dielectric
Breakdown.
Electromigration is again aggravated by small sizes. If the metal interconnects are too
small, high current densities induce the migration of metal atoms. Then voids or hillocks
are formed, which results in increased resistance, opens circuits, or shorts to adjacent
structures.
SWEAT and Isothermal methods are used for evaluating electromigration in shorter times.
Note
These are industry standards for most reliability tests.
Now let's move to the measurement instrument requirements which result from all the
test parameters which we have discussed.
Firstly, let's make sure we understand what the SMU is. SMU stands for Source Monitor
Unit, or Source Measurement Unit.
The SMU has 3 modes of operation:
Voltage source (or force) and current monitor mode.
Current source and voltage monitor mode.
Source common mode.
Each SMU can operate in any combination of sink or source, allowing full 4-quadrant
(negative and positive) source and measure capability.
22
SMUs are the fundamental hardware component of parametric test. Try to imagine what
you would have to do to duplicate SMU functionality using discrete equipment (eg
voltmeters, ammeters, power supplies, etc.). Just appreciate the powerful simplicity of the
SMU.
Why a DC Analyzer has 4 SMUs?
Well, transistors are actually 4 terminal devices . As well as the Gate, Drain and Source,
the substrate under the transistor must be biased to a known voltage in order for the
device to function properly. By assigning one SMU to each terminal, any measurement can
be made without having to change the device connections. Thus, the minimum number of
SMUs required to make effective MOS transistor measurements is 4.
Importance of Triaxial Cables

In a standard coaxial cable (where the signal core is surrounded by a grounded shield)
there is too much leakage to make effective low current measurements, since there is
actually a very small current flowing through the insulator. Also, parasitic capacitance
slows down the measurement speed.
To make truly accurate low current measurements, triaxial cables must be used. In triaxial
cables a guard is added around the force core line to isolate it from the grounded shield. A
special buffer circuitry ensures that the Guard and Force lines are at the same voltage
potential, thus removing the effects of any stray leakage and capacitance. Triaxial cables
are required for low current measurements (ie below 1nA).
Remember that the Guard is always at Force voltage, and therefore must NEVER be
connected to ground. Always leave it open.
The Agilent DC Analyzers are designed for accuracy, so they only support triaxial cables.
23
Kelvin Measurements, otherwise known as 4-wire measurements, are required for low
resistance measurements, when the resistance being measured is comparable to the
resistance of the cables . This is therefore also true when high currents are applied. The
graph illustrates the error which can occur when not using Kelvin cables.
The concept is that 2 wires do the normal force and measure , but the additional 2 wires
sense the voltage at the device and correct the SMU output to compensate for any voltage
drop in the cables. Consequently the set conditions are applied at the device, rather than
at the SMU internally, and the measurement is more accurate.
So Kelvin measurements are done for low device resistance or high current, whereas
triaxial cables are necessary for low current measurements.
In addition to the traditional sweep functionality, the modern Agilent DC ANalyzers have
powerful sampling and stress capabilities, as well important modes such as Standby and
Knob Sweep.
A Spot measurement forces a single voltage or current and measuring during the force.
However, Sweep measurements involve incrementing the forced voltage or current for a
certain number of steps, and measuring at each point. The user can also define a
secondary sweep to create families of curves.
Sampling measurements involve repeating a constant (rather than incrementing) forced

value over period of time, possibly until some pre-defined stop condition is achieved.
Stress involves forcing a DC or AC signal across a device, and measuring after the stress
application has completed. The Agilent DC analyzers offer repeated stress/measure cycles
either manually or automatically. Note that the stress settings are completely independent
of the measurement settings.
In a sweep measurement, voltages or currents are increasingly swept between each

measurement point to create a device curve . The sweep mode has a lot of flexibility. It
can be set up for single or double sweeps, in combination with linear or log sweeps.
Additionally, an SMU can also be put into Pulse Mode where the output of the pulsed
source is made as a typically 1us pulse at each step rather than being fixed for that step.
This feature is valuable in cases where a normal sweep would cause too much device self-
heating.
With Standby mode SMUs stay at their output value, even between measurements so that
devices can remain powered up constantly.
Knob sweep mode allows Curve Tracer like functionality by repeatedly sweeping up to the
value set by the front panel knob.
24
As well as the primary sweep (called VAR1), the user can define a secondary step source
(VAR2) which will increment in value each time the primary sweep source (VAR1)
completes its sweep sequence.
Furthermore, the user can define a secondary synchronous source (VAR1'), which tracks
the primary sweep source (VAR1) timing, but with a user-determined ratio and offset. This
is useful for characteristics such as Gummel plots.
Sampling is a measurement process done over time , where the measurement is repeated
at specified time intervals. The voltage and current settings are not varied for sampling
mode. Many sophisticated features greatly simplify the task of taking repeated time
measurements.
As well as setting the time interval, the user can set either a given number of
measurement points, or no limit, or continue until some pre-defined stop condition is
reached, such as reaching a certain current.
Here is a practical example of the sampling measurement. Earlier we discussed the TDDB
measurement, which determines how long it takes to rupture the insulating dielectric
layer. This gives a measure of the reliability of the dielectric.
In this case, only 300 sample points are retained, and thinned-out sampling is used, with
an interval of 2ms. An important point about thinned-out sampling is that the sample
points near the actual event are still spaced at the specified 2ms spacing. Points earlier on
are the ones which get "thinned-out", so that there is still fine sampling near the event of
interest. This has many obvious benefits, and illustrates a key difference between this
instrument and a sampling oscilloscope.
25
Now let's look at the Stress function. A stress is simply a DC or an AC signal applied to the
device for a specified amount of time , however it may be at a higher voltage or current
than the device would normally work at. The purpose is to see how the device behaviour
changes after a certain amount of stress , possibly beyond it's normal operating limits. It
gives an indication about how robust the device is, and of it's lifetime and reliability,
without waiting the years it would normally take to see an actual device fail.
The stress menu is a separate menu from the sweep and sampling measurement menus,
since typically the user would want to configure a stress as a measurement cycle which is
repeated over and over. The beauty of the stress function is that the stress settings for
the defined stress sources do not have any relationship to the measurement settings for
those same resources; they are totally independent.
Up to 4 resources (ie SMUs, VSUs, or PGUs) can be defined as DC or AC stress sources as

appropriate. For example, an SMU could be forcing current in stress mode but be forcing
voltage in measure mode. This gives the user a great deal of flexibility in setting up
stress/measure routines.
The stress feature is somewhat of an advanced topic, but it is used most often for
reliability testing. A major benefit is that the stress mode is built into the DC Analyzer
itself, and allows use of the built-in PGU channels.
We just mentioned that the stress settings are independent of the measure settings. This
is critical for practical flexibility. Importantly, the stress can be applied in precise amounts
, and accurately recorded so the reliability data is actually meaningful.
The plot above depicts the capacitance of a MOS transistor Gate oxide. It is a bias-
dependent capacitor.
A vertical displacement of the CV curve relates directly to the oxide thickness (Cox).
A horizontal shift relates to the doping concentration level.
26
The slope of the transition part of the curve tells us about the doping profile. The steeper
the slope the more abrupt the change in doping level with depth.
CV measurement can reveal :

Doping concentrations and profiles
Oxide and metal impurities
Film thickness
Carrier lifetime
Thresholds
Monitored processes:
Oxidation / Diffusion
Ion Implantation
Metallisation
As the gate bias is made less negative, fewer holes are attracted to it. The substrate
works similarly to metal. In consequence, the MOS Gate oxide capacitance value will be
equal to the oxide capacitor Cox.
At some point (usually around Vg = 0V), a depletion region starts to build-up below the
gate area because the holes are repelled from the surface between the oxide and the
silicon. As the bias voltage increases, this depletion region continues to grow. In this
state, the MOS capacitance consists of Cox and depletion layer capacitance Cd. Therefore
the MOS capacitance will decrease.
As the bias increases further, electrons (minority carriers) will be attracted to the region
under the Gate. It is said that this region has become inverted. Any further increase in
bias will not change the width of the depletion region, it will only add more electrons to
the inverted regions. The capacitance measured in this condition will depend on the
method used to conduct the measurement.
If a high frequency (>100kHz) measurement is made, the minority carriers (electrons)

cannot react fast enough to the test signal so the capacitor will look like the depletion
region capacitance in series with the oxide capacitance. Since the depletion region width is
constant in inversion, the capacitance is a constant.
If a low frequency capacitance measurement is used, an entirely different result is

obtained. Iin this case, the minority carriers under the gate region can keep up with the
test signal so the capacitance is again just the oxide capacitance (Cox).
The common method for low frequency CV measurement is the Quasi-Static method,
where the MOS gate current responses to a slow linear DC voltage ramp, yielding the CV
characteristics. The capacitance value can be calculated by
C = Ig / ( dV / dt )
where Ig is the Gate current.
The Quasi-Static method has become one of the more common methods of looking at
interface trapped charge. A shift to the right in a n-channel process can be due to sodium
27
contamination (Na+ ion trapping). This increases a transistors threshold voltage. The
slope of the rising edge relates directly to Vth if the oxide characteristics are the same as
the final transistor.
Why a Switching Matrix?
Our last topic regarding measurement instrument requirements is about the switch matrix
.
The equipment required to do accurate current-voltage (ie IV) measurements and high
frequency capacitance measurements (ie CV) is different. So without some means to
switch between the two, users had to spend a great deal of time changing cables every
time they wanted to change between IV and CV measurements. This obviously also
applies when other instruments are used additionally (eg HF, PGUs or DVMs).
Very often, the test device structures are organized on the wafer into groups, and probe
cards with e.g. 24 pins are applied, rather than manual needles. A swithcin matrix allows
to switch between the different devices within a group without re-cabling.
Resistivity
Resistivity measurements are used to evaluate processes related to creation of conductive

layers and lines or to evaluate mask alignment accuracy.
Resistivity of a film can reveal:
The Film thickness, cross-section, purity

Doping concentration and uniformity
Line widths
Mask overlay
Oxidation / diffusion
Deposition
Ion implantation
Metallisation
A Van der Pauw measurement is done by passing a current through two adjacent corners
of a square and measuring the voltage drop across the remaining two electrodes. The
sheet resistivity is calculated by a very simple formula. The main advantage of the Van
der Pauw structure is that it is symmetrical. The leads can be switched and the
measurement can be performed again. If the obtained results are not equal, non-
uniformities exist within the films and can be detected.
A simpler structure to measure the resistivity is a basic 4-point resistor. The resistivity can
easily be calculated if the dimensions are known. On the other hand, the line width can be
calculated if the sheet resistance is known.
Continuity/Bridging
28
These tests are used to determine whether the conductive lines are correctly open or
correctly short. These tests provide a powerful tool for investigating defect levels in
processes related to the creation of conductive lines and contacts.
Continuity and bridging tests can reveal:
Particles / defects in film

Incomplete etches
Defect densities
Dielectric integrity
Contact integrity
Lithography
Deposition
Etching
Metallisation
Measurements:
Various structures are used to measure continuity and bridging. A set of closely spaced
interigitated fingers provides a good tool for detecting shorts. A similar structure is used
to measure continuity, here it is a long continuous line.
Bridging and continuity are also looked for between layers. A simple parallel plate
structure is used to look for dielectric problems. Contact strings are used to look for
continuity of the contacts between conducting layers. Very often these contact strings
have as many as 10000 contacts or more.
Leakage / BreakdownIsolation Edge Effects
Dielectrics and diffused junctions can be very sensitive to contamination or defects

introduced during processing. Because of this sensitivity, the leakage current and/or
breakdown voltage are key process parameters. The reverse-bias characteristics of a
junction is often more important than the forward-biased one.
Reverse-bias testing can reveal:
Ion implant damage

Isolation-edge effects
Contamination
Dielectric integrity
Oxidation / diffusion
Ion implantation
Deposition
Isolation edge effects
When the leakage current exceeds a predetermined level, the production process should
be examined as a cause. One method for how to isolate process problems is shown above.
By using two structures with the same area but different perimeters, the contribution to
the leakage currents from edge effects can be isolated from the one caused by area
effects. These two sources of leakage currents often have different causes.
29
Device Modeling Measurement Tutorial
The trend to higher integration and higher transmission speed challenges modeling
engineers to develop accurate device models up to the Gigahertz range. An absolute
prerequisite for achieving this goal are reliable measurements, which have to be checked
for data consistency and plausibility.
This is especially true for radio-frequency (RF, >100MHz) and microwave (GHz)
measurements, and also for checking and verifying the applied de-embedding techniques.
If there are (hidden) problems with the measurement data, RF characterizations can
become quite time consuming, with a lot of guesswork and adhoc judgments, and,
basically, frustrating and not correct.
If, however, the underlying measurements are flawless and consistent, and provided the
applied the models are understood well, RF characterization and device modeling becomes
very effective and provides accurate design kits which will satisfy the chip designer's main
goal, right the fist time.
Basics of DC and AC Characterization of

Semiconductors
Basics of device measurement and modeling techniques from DC to RF
Special aspects of network analyzer calibration
De-embedding and required dummy structures
While the characterization of electronic components in the DC domain is relatively simple

and only requires a Voltmeter and an Ampremeter, the frequency performance of the
device is affected by magnitude dependence and phase shift of the currents and voltages.
Furthermore, non-linearities will lead to a spectrum of frequencies, although the device is
only stimulated with a single, sinusoidal frequency.
Last not least, inevitable capacitive and inductive parasitics, with values close to those of
the very device under test (DUT), will contribute to the measurements and degrade the
measured performance of the 'inner' DUT.
In this presentation, we will go step by step through the individual characterization issues
and develop measurement strategies which will provide the base of accurate device
characterizations.
Modeling of a nonlinear component always begins with the characterization of its DC

performance.
Instead of power supplies, precision DC parametric analyzers with source-monitor-unit
(SMU) plugins are applied. They offer current ranges from femto-Ampre (1E-15) to
several Ampres, and voltage potentials from micro-Volt to hundreds of Volts. This allows
to fully characterize the DUT (device under test) in all four I-V quadrants. I.e. forward and
reverse currents and voltages, are measured with the same SMU unit.
Usually, in case of a transistor, all 4 terminals (including substrate) are connected to
individual SMUs in order to avoid recabling during the forward and reverse measurements.
PlotOption Access Name and Value Examples
SMUs apply a Kelvin measurement to avoid parasitic series resistances. This measurement
procedure, also known as the four-wire method, consists of a stimulating line (Force) with
a second one in parallel (Sense) for every pin of the DUT. The slide above illustrates this.
Ohmic losses on the Force line are eliminated by the operational amplifier (OpAmp1) in
30
voltage follower mode. This means this OpAmp1 output will exhibit a somewhat higher
voltage than the desired test voltage at the DUT, because the test current generates some
ohmic losses along the Force line. The Sense line, connected to the minus input of the
OpAmp1, assures that the DUT is biased with exactly the desired test voltage.
The Guarding (Triax Cable) Principle for Fast and

Accurate Measurements
While the Kelvin method compensates the DC errors, it does not cover dynamic DC
measurement problems. For example, to avoid external electro-magnetic influences, both
the Force and Sense cables are shielded. But such cable shieldings exhibit parasitic
capacitance. Due to charging problems, these capacitances will affect the measurement
speed and accuracy of our Kelvin measurement.
As a simple example: assume we want to measure the reverse characteristics of a
semiconductor diode. This means we need measure very low currents. Before the voltage
steps to e.g. -20 V, the quiescent voltage at the diode is zero. That is, the cable capacitors
are not charged. When the negative voltage step occurs, these capacitances have to be
charged, and the required current is provided by the OpAmp1. This could lead to either a
mis-measurement (DUT current plus charging current) or a delay in the triggering of the
actual current measurement (by some intelligent firmware in our measurement).
To solve this problem, an extra inner shielding is applied between the hot metering lines
and the outer cable shielding, called 'Guard'. This extra shielding is connected to a
separate, second OpAmp2 which follows exactly the value of the desired test voltage. Now
it is this auxiliary OpAmp2 which supplies the charging current for the test cables, while
the main OpAmp1 can start current measurements without being affected by this charging
problem. That is, the inner measurement loop does not see the charging problem any
more.
Of course the point where Force and Sense are tied together must be as close as possible
to the DUT. In case they aren't connected, an internal 10kOhm resistor at the output of
the SMU acts as the Kelvin point. Another important fact is that the Guard contact should
never be connected to Force or Sense. Otherwise, the inner loop OpAmp1 of the SMU
would measure the DUT current plus the charging current of the auxiliary, second
OpAmp2!
Calibration:
In order to maintain the DC measurement accuracy, SMUs perform periodically an auto-
calibration. This means that the SMU disconnects its outputs from the DUT, measures
possible offset voltages and currents and corrects it. This type of calibration does not
require any action from the user.
Measurement Environment
Complete triaxial system
Typical on-wafer IV measurement setups require probes and chuck bias. A careful
shielding completely surrounds the wafer with guard.
Triax Feedthru and Triax-Coax Converters From Trompeter

Electronics
31
When connecting the Triax connectors, and particularly when applying an adapter, make
sure the Guard is left open and floating !
Oscillations during DC measurements
Under certain circumstances, SMUs may oscillate. Super-beta, wide- bandwidth bipolar
transistors are especially susceptible to oscillation. But also GaAs transistors.
Oscillation can become a problem when using older DC probe needles. A common way to
avoid such oscillations is using HF probes (Ground-Signal-Ground GSG probes), or the
newer shielded DC needles.
If you have to live with your DC needles, and when oscillations occur, here some ideas on
causes for such oscillations:
1. SMU Induced oscillations. These oscillations occur when an inductive load is

connected to the SMU's output. The SMU has often an inductive load because the
output impedance of the SMU is inductive during V-Mode operation and usually
several SMUs are indirectly connected together through the DUT.
2. Oscillation due to strays. The measurement system including the DUT, stray
capacitance and residual inductance of the connection cables, switching matrix, probe
card and/or test fixture can be recognized as an oscillation circuit.
The oscillation detector of SMU may not detect this type of oscillation. This is
understandable when thinking of the SMU output as a low-pass filter, while the
oscillation frequency may be in the 100MHz ... to several GHz range. Also, if like in
most cases the oscillation is located at the DUT, the SMU itself cannot do anything to
prevent this oscillations, since the cables are long compared to the oscillation
wavelength.
As a general rule: the oscillation has to be avoided where it happens.
Related to the example of the slide above, we should be aware of the total circuit: The
SMUs connected to the MOSFET's gate and drain are operating in the V-Mode. Since SMUs
typically appear to be inductive in V-mode, this makes this configuration equivalent to a
Hartley oscillator !
The SMU may oscillate if an unusually large inductance is connected to it. This could occur
if the DUT is a superbeta transistor (big hFE) and the SMU connected to the emitter is set
to one of the low current ranges.
For more details on conditions for oscillations, refer to the application note 356 -1
(publication number Agilent 5950-2954)
Methods to eliminate oscillations

Ferrite beads placed as close as possible to the DUT to prevent high frequency
oscillations (Part Number 9170-0029)
Keep cables as short as possible - Use High quality cable (low cable inductance)
32
In some cases, more than 1 Ferrite Bead may be required!
There are some effective methods to eliminate oscillations:
1. For FETs, add resistive ferrite beads as close as possible to the gate.
2. For bipolar transistors, add resistive ferrite beads as close as possible to the base or
emitter.
3. Keep cables as short as possible. Long cables cause oscillation because of their large
inductance.
In some cases, it is necessary to use more than one ferrite beads (Agilent Part Number
9170-0029).
Or, apply shielded DC probes, even better apply RF probes.
Self-heating during DC measurements
As a general observation, currents above ~25mA will cause self-heating effects with
transistors. Rule of thumb: 25mA with typically 2V equals already 50mW for that tiny
transistor on the wafer!
Self-heating can become a severe problem with device characterization, because, when
self-heating occurs, the measurement results depends on the measurement speed!
This can be verified when e.g. measuring a Gummel plot for a bipolar transistor, once
sweeping from low to high voltages different measurement speeds, and then sweeping
from high to low voltages. Calculate beta out of your different measurements.If self-
heating occured, you will get as many beta curves as you have performed measurements!
The plot in the slide above gives the chip temperature increase of a packaged transistor as
a function of the pulse width and the applied bias power. The pulse period is 1s. As can be
seen, self-heating can only be avoided when applying very short, pulsed measurements,
below 1us pulse width. Such pulsed measurement systems, also included pulsed S-
parameter measurements, are commercially available.
However, such systems are quite complex and expensive. Therefore, if you have to live
with self-heating, make sure your device suffers always from the same self-heating. This
means in general, apply the slowest measurement speed for you're your DC
measurements, since your biased network measurements will be slow as well, and self-
heating will definitively occur there !
What to do when self-heating becomes an issue for your device modeling:
from the slide above, we identify a min. pulse width of ~1us or below. This means
the DC analyzers in pulse mode cannot be used( pulses are 100us or longer)
use the IC-CAP pulsed modeling system
live with self-heating, but make sure your data are consistent:
all DC measurements are performed with the same self-heating (slow
measurements!)
same self-heating for biased S-parameter measurements as for the DC
measurements.
As discussed in the previous chapter, the DC voltages and currents can be measured
directly. The calibration is periodically auto-executed by the instrument.
33
After such a DC characterization, modeling engineers usually perform a so-called CV (c
apacitance versus voltage) measurement in order to characterize the device capacitances
at a standard frequency of 1MHz. This frequency is high enough to allow a resolution down
to a few femto-Ampere (provided shielded probes are applied for e.g. on-wafer
measurements), yet still low enough to neglect second order parasitics like resistors in
series with the capacitors, or like inductances.
For such CV measurements, the DC-bias is swept, a test frequency (1MHz) is applied to
the DUT, and the instrument calculates the capacitance between the 2 pins of the DUT
from the magnitude and phase of the device voltage and current. This means, an
impedance meter interprete the measurement result always with respect to a user-
specified schematic: either a capacitor in series with a resistors, or both in parallel. This
explains, why capacitances and resistor values may vary with frequency when measured
with such a device. In other words, these frequency-variations are due to a too simplistic
analysis model behind the measurement. A better way is therefore to measure the
capacitances with network analyzers. In this case, it is up to the user to interpret the
measurement result (S-parameters).
CV Measurements - Calibration considerations

Test cables and fixtures contribute and affect the device characterization. For CV
measurements, the calibration consists of unconnecting the DUT, assuming an ideal OPEN
condition and measuring the cables and their OPEN parasitics (CV-Meter calibration). After
that, the corresponding capacitance is automatically subtracted from the DUT
measurement by the CV meter.
Note
If we are interested in the inner DUT's CV curves, i.e. without its surrounding test pads capacitances, we
need to connect to an OPEN dummy structure during CV meter calibration instead of simply leaving the
cables unconnected. Such an OPEN dummy consists of all connection pads, lines to the DUT etc, but
without the inner DUT itself.
CV Meter Applying the Auto-balancing Method
For CV meters, the measurement principle is typically a so-called auto-balancing method.

The slide above depicts the simplified measurement scheme.
The DUT is inserted in the feedback loop of an operational amplifier, and the system is
stimulated with a 1MHz sinusoidal signal plus a DC bias. The feedback resistor R is
precisely known, and the complex voltages V1 and Vdut are measured. From the formula
given above in the slide, the capacitance of the DUT can be calculated, assuming an
equivalent schematic of either a resistor in series with the capacitor, or, commonly for
modeling, a capacitor in parallel with a resistor (which is the bias-dependent diode
resistance for example).
Avoiding Measurement of Parasitic Capacitance
34
A so-called four-wire method is used for CV measurements. Similarly to the DC Kelvin

measurement procedure (Force and Sense), both the low and high pin have to be tied
together at the location of the DUT during CV measurements.
In the auto-balancing impedance measurement method, the shieldings of these four wires
(which are not instrument chassis ground!) are connected to the virtual ground of the
instrument's OpAmp. This eliminates any influences caused by the cables. Including this
potential into the measurement setup allows the elimination of further stray capacitances.
Like sketched above, it is possible to eliminate stray capacitances against ground
(measurement plate).
AttoGuard Enhanced CV Measurements
FemtoGuard surrounds the chuck at shield ground

Patented AttoGuard above the chuck at shield ground
Creates a virtual double-shielded Faraday enclosure
10 atto Farad CV measurement resolution
Zero CV meter only one time
The two shields insure equal potential everywhere inside the waferprober MicroChamber
regardless of chuck locations.
(slide with courtesy from Cascade Microtech)
While the CV measurement is considered as a specific two-pin test condition, the situation
changes for frequencies above 100MHz. The device is now operated under its originally
intended environment conditions: DC bias is applied to all the pins, and an additional
small-signal RF excitation is applied. Now, the sinusoidal currents and voltages at all pins
of the DUT are to be measured, with magnitude and phase.
A natural choice for such characterizations would be Z-, Y- or H-parameters from linear
two-port theory. These two-port parameters can be used to completely describe the
electrical behavior of our device (or network), including any source and load conditions.
For such parameters, we have to measure the voltage or current as a function of
35
frequency and bias at the ports of the device.
At high frequencies, however, it is very hard to measure voltage and current at the device
ports. One cannot simply connect a voltmeter or current probe and get accurate
measurements due to the impedance of the probes themselves. Additionally, there is a
difficulty to place the probes at the desired positions. Furthermore we have to apply either
(AC-wise) OPEN or SHORT circuits as part of the Z-, Y- or H-parameter measurement.
Active devices may oscillate or self-destruct with such terminations.
To avoid these problems, twoports are described by S-parameters.
IMPORTANT NOTE: like Y, Z, H parameters, S-Parameters are linear and belong to linear
circuit theory. I.e. they represent the small-signal behavior of a device at a certain bias
point, and for a certain frequency. Therefore, when measuring them, it must be assured
that the linear device operation is maintained. As a consequence, linear S-parameters are
independent of the applied RF signal power. Since twoport theory is restricted to linear
circuit theory, matrix conversions (S to Y, S to Z etc. for de-embedding) are only
applicable for linear operation of the device.
When characterizing the capacitance of transistors, the open 3rd transistor terminal
should be connected to the shielding potential, eliminating the effect of the unwanted
capacitors.
Definition of S-parameters
Referring to the spectacles examples from above, i.e. power-wise, the S-parameters are
defined as:
Over the years there have been many methods developed to model the behavior of a
circuit or network. H, Z, and Y parameters all have benefits to visualizing and predicting
circuit behavior.
However, as the circuit frequency climbs higher, many of these models become complex
and unruly. As a result, it is hard to model the network in a straightforward manner.
Scattering (S)-parameters give a simple way to quantify the ratio of voltages exiting or
scattering from a network relative to the voltages incident upon the network. By
cascading* the parameters, models for individual circuit blocks may be put together to
predict the operation of the whole system. As we will find out these ratios are fairly easy
to measure at high frequencies. And difficult measurements of complex short-circuit
currents and open-circuit voltages are not required to measure S-parameters.
Since S-parameters are defined as a voltage ratio, their result maps well into familiar RF
and microwave characteristics of gain, loss, an reflection coefficient. They are also easily
converted to other parameter forms such as H, Y, or Z, and this transportability makes S-
parameters a good choice for importation into most modern simulation and design
software programs. For readers acquainted with matrix algebra:cascading involves matrix
multiplication of "T" matrices which are derived by a transformation of the S parameter
matrices.
S-Parameters and Characteristic Impedance Z0
36
This means that S-parameters relate traveling waves (power) to a two-port's (DUT)
reflection and transmission behavior. Since the two-port is imbedded in a characteristic
impedance of Z0, and since we apply linear circuit theory , these 'waves' can be
interpreted in terms of normalized voltage or current amplitudes.
Note
Think of the spectacles again. Also here, the lens is imbedded on both sides with the same 'characteristic
impedance', i.e. air.
When determining the linear characteristics of a device, what we need to know is fairly
simple. Given a sinewave signal incident upon the DUT, what of that signal is reflected
from the device or transmitted through it? From the arrows in the illustration, there are
four possibilities for this 2-port device. Knowing these four parameters completely
characterizes this device for its linear operation.
For an n port device there is a reflection coefficient at each port and a transmission
coefficient between each pair of ports in each direction.
So what we need to measure is
Transmission Coefficient = Transmitted / Incident

Reflection Coefficient = Reflected / Incident
For a complete characterization of the linear behavior of an RF device.
To quantify and express these reflected and transmitted coefficients, the scattering or S-
parameter were defined. S-parameters give a simple way to quantify and organize these
ratios of voltages exiting or scattering from a network relative to the voltages incident
upon the network.
S-parameters are always a ratio of two complex (magnitude and phase) quantities.
S-parameter notation identifies these quantities using the numbering scheme shown
above. The first number refers to the test-device port where the signal is emerging, or
another way to look at it, which network analyzer port is the signal being measured. The
second number refers to which test-device port the signal is incident or which network
analyzer port the signal is coming from. For example, the S-parameter, S11, identifies the
measurement as the complex ratio of the signal emerging from port 1 of the device to the
signal applied to port 1 of the device (a reflection measurement).
A two-port device or network has four S-parameters. Two of the terms are related to the
reflection from the input and output ports of the DUT. The other two terms are related to
the transmission through the DUT in the forward and reverse directions. These concepts
can be expanded to multi-port devices and the number of S-parameters is a function of 2
n , where n = the number of ports. For example, a four port device would have 16 S-
parameters.
For the two port device, two independent equations may be written, expressing the
variable b in terms of the variable a. In the above diagram b1 comprises the sum of a
37
quantity reflected from port 1 and a quantity that is the result of transmission through the
device in the reverse direction. The quantities are scaled to be proportional to the voltage
wave amplitude and phase such that |bn|2 = power emerging from the n'th port and |an|2
is the power incident on the n'th port.
S11 and S21 are determined by measuring the magnitude and phase of the incident,
reflected and transmitted signals when the output is terminated with a perfect Z0 load.
This condition guarantees that a2 is zero. S11 is equivalent to the input complex reflection
coefficient or impedance of the DUT, and S21 is the forward complex transmission
coefficient.
Likewise, by placing the source at port 2 and terminating port 1 in a perfect load (making
a1 zero), S22 and S12 measurements can be made. S22 is equivalent to the output
complex reflection coefficient or output impedance of the DUT, and S12 is the reverse
complex transmission coefficient.
The accuracy of S-parameter measurements depends greatly on how good a termination
we apply to the port not being stimulated. Anything other than a perfect load will result in
a1 or a2 not being zero (which violates the definition for S-parameters).
Interpreting S-parameters
The magnitude of S11 and S22 is always less than 1. Otherwise, it would represent a
negative ohmic impedance value.
The magnitude of S21 (transfer characteristics) respectively S12 (reverse) can exceed the
value of 1 in the case of active amplification. Furhtermore, S21 and S12 can be positive
and negative. If they are negative, there is a phase shift. Example: S21 of a transistor
starts usually at about S21 = -2 ... -20. This means signal amplification within the Z0
environment and 180' phase shift.
Calculating S-Parameters From Voltages
38
The slide above presents a method to calculate S-parameters from complex voltages.The
DUT is imbedded in an characteristic impedance environment.
Provided the required complex forward and reverse voltages V1, V01 , V2, V02 can be
obtained (from measurements or simulations), we can calculate the S-parameters from
the equations given above.
Understanding the Smith Chart
What makes Sxx-parameters especially interesting for modeling, is that S11 and S22 can
be interpreted as complex input or output resistances of the two-port. That's why they are
usually plotted in a Smith chart.
Note
Do not forget that included in Sxx is the termination at the opposite side of the two-port, usually Z0.
The Smith chart is a transformation of the complex impedance plane R into the complex
reflection coefficient (rho), following the formula given above.
This means that the right half of the complex impedance plane R is transformed into a
circle in the domain. The circle radius is '1' .
In order to get more familiar with interpreting the Smith chart, let's consider a square
with the corners (0/0)W, (50/0)W, (50/j50)W and (0/j50)W in the complex impedance
plane 'R' and its equivalent in the Smith chart with Z0=50W. Watch the angel-preserving
property of this transform (rectangles stay rectangles close to their origins). Also watch
how the positive and negative imaginary axis of the R plane is transformed into the Smith
chart domain (W), and where (50/j50)W is located in the Smith chart. Also verify that the
center of the Smith chart represents Z0, i.e. for Z0 = 50W, the center of the Smith chart
is (50/j0)W.
Why a Smith Chart for the Sxx Parameters?
39
Interpreting S11 and S22 (Sxx)
Rules for Smith charts for NWA measurements:
Sxx on the real axis represent ohmic resistors

Sxx above the real axis represent inductive impedances
Sxx below the real axis represent capacitive impedances
Sxx curves in the Smith chart turn clock-wise with increasing frequency (because in
the R plane,all curves turn clock-wise too !).
S11 of a Capacitor and of a Bipolar Transistor
As an example for interpreting Smith charts, the left plot shows the S11 curve of a
capacitor located between the two ports of the network analyzer (NWA). The capacitor
represents an OPEN for DC, thus S11 = 1 = *Z0. For highest frequencies, it behaves like a
SHORT, and we see the 50!measurement_dm_symbol1.gif! of the opposite port2 . The
transition between the DC point and infinite frequency follows a circle, and the increasing
frequency turns the curve again clockwise.
The right plot in the slide above shows the S11 plot of a bipolar transistor. In this case,
the locus curve stars with S11>>1 at low frequencies corresponding to RBB' Rdiode
(1+beta)*RE. For increasing frequencies, the curves then turn into the lower half-plane of
the Smith chart, the capacitive region. Here, the CBE shorts Rdiode, and beta becomes
smaller with frequency. For infinite frequency, when the capacitors represent ideal shorts,
40
and ->0, the end point of S11 lies on the middle axis, i.e. the input impedance is
completely ohmic, representing RBB' + RE. Since RBB' is bias dependent, and decreasing
with increasing iB, the end points of the curves represent this bias-dependency.
Note
Keep in mind: For increasing frequency, the Sxx locus curves turn always clockwise.
The Polar diagram for Sxy

S21 corresponds to the forward transmitted voltage and S12 to the reverse transmitted
voltage in a Z0 environment. In the polar diagram, you can easily identify MAGNITUDE
and PHASE of these voltages.
The S21 parameter represents the power transmission from port 1 to port 2, if the two-
port device is inserted into a matching network with characteristic impedance Z 0 of e.g.
50W. This means, if no signal is transmitted, then S21=0 (located in the center of the
polar plot). If the signal is transmitted, then MAG(S21)>0. The magnitude of the S21
curve will be below '1' for damping between the port 1 and port 2, and above '1' for
amplification. If the phase is inverted (transistors), we are basically in the left half-plane
of the polar plot (REAL[S21]<0).
Like with the Smith chart, all S21 and S12 curves turn clock-wise with increasing
frequency.
S21 (transmission) of a capacitor and of a bipolar transistor
As a tutorial example, the capacitor on the left, exhibits no power transmission for lowest
frequencies, but an ideal short (S21=1) for highest frequencies. Thus, the locus curve for
S21 represents a circle from infinite to 0 W.
The right plot from above shows the S21 plot of a bipolar transistor between port 1 and
port 2. The trace starts with REAL(S21) < -1 at low frequencies (voltage amplification in a
50 W system, plus phase inversion), and then tends towards S21 = 0 for highest
frequencies (no voltage transmission, the transistor capacitances short all voltage
transmission). Since the current amplification is bias depending, the start point of the
S21 curve at lowest frequencies reflects this (iB) dependency: more for higher iB, i.e.
more amplification magnitude with S21 for higher iB too.
Network Analyzer Measurements

Network Analyzer Mainframe
RF Synthesizer
S-Parameter Testset
41
After the introduction to the S-parameters, it is time to consider how to measure them. A
network analyzer (NWA), also sometimes abbreviated by VNA (vector network analyzer),
is applied. This instrument measures S-parameter vectors, i.e. the magnitude and phase,
of all four S-parameters of a two-port. This 'full two-port measurement' capability is
important, because only in this case are we able to convert the measured S-parameters to
Y and Z parameters etc., what is a requirement for de-embedding etc.
When applying network analyzers for S-parameter measurements, it is important to

remember that we measure linear circuit performance and circuit performance for a given
frequency (fundamental frequency), ignoring harmonics.
On the other hand, network analyzers can also be applied to specific non-linear
measurements, e.g. sweeping the RF power, measuring the transfer characteristics and
evaluating for example the 1dB compression point of amplifiers. In this case, however,
signal distortion happens and harmonic frequencies show up. In our case, when
measuring linear S-parameters with the NWA, always the base or fundamental frequency
and is measured, and harmonics should not occur. Otherwise, they would be ignored !!
Therefore, if we are interested in the modeling of device nonlinearities, we should rather
apply a spectrum analyzer after the conventional DC-CV-NWA modeling, and use harmonic
balance simulation (e.g. Agilent's ADS) to model the RF-power dependent spectrum.
Alternatively, one of the currently introduced commercial Nonlinear NWAs (Agilent N4463A
Large Signal Network Analyzer) can be applied as well. Such instruments measure both,
the magnitude and phase of the transmitted and reflected, fundamental and harmonics
frequencies.
When Hewlett-Packard introduced the HP 8410 in 1967, it revolutionized microwave

design. It used the a new 1430A sampler together with a superheterodyne receiver
architecture to provide a calibrated microwave receiver. Together with the test sets, it
featured measurements of the transmission and reflection coefficients for any twoport
device.
Block Diagram of the S-Parameter Testset
42
The block diagram in the slide above shows the core of this meter combination, the S-
parameter testset. The RF Input source at the top, connected to the RF synthesizer,
provides the stimulus power. The PIN switch directs the signal to either a forward or a
reverse S-parameter measurement. Directional couplers then detect the injected and
reflected power of the DUT. The detected signals are downcoverted into four IF signals for
further analysis in the VNA mainframe, where each input is digitized and signal processed
in order to give the S-parameters.
NWA Measurements Using the Internal DC bias TEEs of the S-

parameter Testset
For devices like transistors and diodes, an additional DC bias has to be applied to the
device.
This can be done by using the DC bias inputs of the NWA's S-parameter testset. Keep in
mind that there are typically 1-2W ohmic losses due to the internal bias TEE's inductor.
This causes a voltage drop and a reduced bias voltage at the device! Many NWA's also
have an internal 1MW resistor to ground, which prevents electrostatic discharge damage
to the internal NWA circuitry.
Note
In the setup presented above, make sure to use the right triax (SMU) to coax (S-parameter testset)
adapter, which leaves the middle shield of the SMU triax cable (Guard) unconnected.
NWA measurements with external DC bias TEEs
43
If we want to avoid the ohmic losses of the NWA's internal bias TEE, we can use external
bias TEEs and apply the previously mentioned force-sense DC biasing (Kelvin
measurements).However, the ohmic losses for the bias are now not zero, but rather
~1W?due to losses in the cables and the connectors. Also, these Kelvin bias TEEs should
be placed as close as possible to the DUT.
In the measurement setup above, two external bias TEEs are inserted between the two
DUT connections and the network analyzer ports.
Keep in mind to not connect the guard shield of the triax Kelvin cables.
Basics of DC and AC Characterization of Semiconductors

Contents
Basics of device measurement and modeling techniques from DC to RF

Special aspects of network analyzer calibration
De-embedding and required dummy structures
Vector Error Correction

Accounts for all major sources of systematic error.
44
Since S-parameters are vectors, all measurement errors contribute with magnitude and
phase, and can be considered as additional error vectors. These errors have to be
calibrated out with the correction vector Sxxcorrection, as depicted above.
What is a 12-term error correction? As can be seen above, there are 6 error contribution
terms in forward direction, related to the characterizing signals R, A and B of the NWA:1.
Directivity:cross-talk of the power splitter in the NWA testset2. Crosstalk:cross-talk inside
the S-parameter test set, overlying the DUT3. Source Mismatch:multiple reflections due to
impedance mismatch of cables and connectors4. Load Mismatch:the same for the opposite
port5. Reflection Tracking A/R:frequency dependence of signal path R->A6. Transmission
Tracking A/R:same for signal path R>B
For the reverse calibration, another 6 error terms add up to a total of 12 terms. The
procedure to get rid of these 12 terms is called the 12-term error correction.
The formulae above are applied in the NWA in order to correct the measured S-parameter
SxyM with the correction terms Exx, and to finally obtain the requested SxyA parameters
of the device under test (12 term error correction).
It is of particular interest that e.g. the resulting S11A parameters are affected by all
measured S11M parameters ! This means, if there is a problem with 'only one' S-
parameter index during measurements (or calibration), this will affect all S-parameters.
This means, an absolute clean calibration and also an absolute proper measurement is
45
required in order to get good S-parameter results!
Calibrating a NWA
Full 2-port calibration
(reflection and transmission measurements)
12 systematic error terms measured usually requires 12 measurements on four
known standards (SOLT)
Standards are defined in cal kit definition table; these cal kit definitions are entered
to the network analyzer
The internal NWA CAL KIT Definition must watch the actual kit used
There are many different calibration techniques for network analyzers. Such are Short-
Open-Load-Thru (SOLT), Thru-Reflection-Load (TRL) or Load-Refection-Match (LRM). For
the different calibration procedures, specific known standard terminations have to be
measured.
The slide above depicts such standards for connectorized measurements (e.g. 3.5mm
connectors).
Although there are many publications on the pros and contras of the different calibration
methods, the SOLT is most commonly used for on-wafer measurements of silicon devices.
One of the reasons is that due to the electrical losses of silicon, microstrip standards as
required for LRM and TRL are difficult to manufacture on the wafer. Another reason for
using SOLT is that this calibration is a wide-band calibration and not limited to a frequency
band.
Some Naming Conventions: The Calkit

While in the case of the CV meter, the calibration corrects for a single, ideal offset
capacitor, a NWA calibration relates to cal standards (OPEN, SHORT, LOAD, THRU etc.)
from a calkit. These cal standards do not represent ideal standards. They represent the
_real, existing standard, including its nonidealities! It means that a SHORT is not an ideal
SHORT, but instead represents rather a small inductance. The same applies to the THRU,
which has a non-ideal delay time. The OPEN corresponds rather to a capacitor than to an
ideal OPEN. Therefore, these non-idealities of the G-S-G probes have to be entered into
the NWA before calibration. This is called 'entering' or 'modifying the calkit'.
THE CALSET
While entering the calkit refers to the non-idealities of the calibration standards, i.e. the
termination of the NWA cables during calibration, the subsequent calibration is based on
this information, and then related to the selected - frequency range, - the RF power, - the
averaging of the NWA etc. After the calibration has been performed, the correction terms
are stored in the calset of the NWA. In other words, the 12-term error vectors are 'filled
up', and can be used with the measurements afterwards.
Finally, when the measurements are performed, the raw measured data arrays will be
corrected inside the NWA, using a correction technique related to the selected calibration
method, and referring to the specified calset. When using a NWA driver software, this
corrected measurement result is transferred into the software and displayed there.
Note
After the calibration, a re-measurement of the OPEN will not represent an ideal open, but instead exactly
those parasitic components as described in the documentation of the OPEN. In the same way, a THRU
shows up after calibration with its real delay time, and a SHORT represents its inductive behavior! If the
calibration was ok, the remeasured standards should give exactly the same parameter values as
previously entered into the NWA calkit.
46
In case of a SOLT calibration, and for on-wafer measurements using Ground-Signal-
Ground (GSG) coplanar probes, the slide above depicts the corresponding test structures,
which are usually available on a RF-high-performance ceramic substrate, including
accurate description of the non-idealities of these standards. These standards (called ISS
standards) are provided by the GSG probe manufacturers. Both, the GSG probe and the
ISS substrate belong together.
Extending linear to RF power dependent signals
An absolute prerequisite for using S-parameter two-port matrices is the linear, time
invariant behavior of the circuit. Only in this case, matrix conversions like for example for
de-embedding, are possible. Nonlinear high-frequency measurements cannot be de-
embedded by Y- and Z-matrix subtractions.
What if characterizing and modeling a transistor at higher RF signal levels where it

behaves non-linear. If you need to model e.g. a power RF transistor at signal levels
where distortion occurs, and where the Kirchhoff law is not fulfilled for the data
measured by the NWA, it is absolutely mandatory to replace the linear SPICE S-
parameter simulation by a nonlinear_ harmonic balance (HB) simulation. Only with
this kind of simulation, we can emulate the conditions of the power RF transistor
measurement with the NWA. From the simulation data of harmonic balance, we can
derive the S-parameters of the base frequency and compare these S-parameters with
those obtained from the NWA measurements. Only in this case the obtained model
parameters will be correct for the power RF transistor.
If distortion occurs, harmonic frequencies show up, which by themselves shift the DC
operating point. With the NWA plus SMUs, we measure only the (shifted) DC bias
current and with the NWA the fundamental frequency current vector. Their sum is not
zero. Kirchhoff's law states that the sum of currents, for all frequencies (including
DC) into a node is zero. But with linear NWA, we do not measure the harmonics
frequencies (with magnitude and phase). Only a very special Nonlinear Network
Analyzer, not to confuse with standard NWA's, can measure this complete frequency
spectrum with respect to magnitudes and phases.
Conclusions: If the DUT behaves non-linearly:- signal compression occurs- harmonics

show up- the DC operating point may be shifted- loadlines and transfer curves
become dynamic, i.e. RF-power dependent- matrix conversions are no longer
possible- de-embedding by S-to-Y and S-to-Z matrix conversion (see further below)
is no longer valid.
Selecting the right RF power for nonlinear devices before starting calibration.+ When
measuring S-parameters of nonlinear devices with a network analyzer, it must be
assured that these devices operate in small-signal, linear mode. Otherwise, the high
frequency test signals will no longer be sinusoidal, and the occurrence of harmonics
47
will lead to wrong S-parameter measurements and shifted DC bias conditions.
A smart method to check the correct port power settings is like this: When measuring
a DC output characteristics and calculating the output resistor 'Rout' from of it, the
resulting curve is very sensitive. Therefore, we can use this plot to identify possible
effects of too big an AC power applied to the transistor. This means, we measure the
DC output characteristics, and let the NWA perform un-triggered measurements, in
continuous mode, i.e. un-synchronized to the DC measurement. Then, we increase
the Port power manually (or decrease the port attenuation) until we see an effect on
the next 'Rout' measurement. We then know the maximum allowed RF power for the
S-parameter measurements of this device!
The plot above reflects such a test. The disturbed curve happens when too much RF
power is applied to the transistor.
Note
'too much RF power' is a relative issue! e.g. -30dBm RF signal (1uW) is small compared to an DC
operating point power of e.g. 3V and 1mA (3mW). However, -30dBm can be by far too much RF
signal for a DC bias point of e.g. 3V and 1uA (3uW). In such a case, for e.g. a diode, the 'relative'
big RF signal with its rectification effect (harmonics) will shift the DC bias point. The same small RF
signal would be too week to do the same for a mW DC operating point.
Verifying the NWA calibration
When re-measuring the calibration standards after the calibration, we should get exactly the calkit
data of the cal.standards (like they were entered before into the NWA.
In order to verify the calibration, it is highly recommended to re-measure the calibration standards
and to model them, using the calkit data of the GSG probe or the connectorized standards.
As an example, for an on-wafer SOLT calibration, we re-measure the cal.standards, e.g. the OPEN,
the SHORT, the THRU and the LOAD. We know that this measurement will correspond to the non-
idealities of the selected cal.standard, as specified in the NWA calkit data. In case of Cascade
probes, the OPEN, for example, behaves like a negative capacitance of roughly -9fF . Now, after this
measurement has been made, we can define a test circuit for that Setup in a modeling software
package like IC-CAP, and enter the netlist of these calibration standards. Using SPICE3, a simulator
which also permits negative capacitance, we can simulate the expected behavior of the OPEN
probes. If the calibration was executed correctly, there is an excellent match between measured and
simulated curves.
In a next step, we measure the SHORT, define in IC-CAP the SHORT non-idealities in a SPICE
circuit, and simulate. Again, an excellent match between measured SHORT data and simulations has
to be achieved. We then continue with the THRU and LOAD measurements and simulations.
Only if all 4 standards exhibits an excellent fit, we can assume a correct calibration of the NWA.
Note
This calibration verification can also be applied to check the quality of an older calibration.
The results depicted in the slide above give an example for Cascade G-S-G probes,
100um pitch.
Once again, only if the fitting between simulated and measured data is in the few-percent
range, for all 4 re-measured calibration standards, the NWA calibration can be considered
as good. If only one fit is bad, re-perform a new NWA calibration.
How parasitics degrade the device performance?
48
After having performed a network analyzer calibration, the calibration plane is located at
the ends of the NWA cable connectors (connectorized devices) or at the ends of the probe
contacts (on-wafer measurements). The device itself, including its surrounding parasitics,
is then connected to this calibration plane.
In the case of a packaged device, S-parameter measurements would now include the test
fixture, the package and the very inner DUT. For on-wafer device characterization, using
e.g. ground-signal-ground probes (GSG), the test pads (where the probes touch down)
degrade the performance of the inner DUT by their layout specific capacitive and pad
parasitics.
In order to extend the calibration plane to either the beginning of the package, or to the
inner DUT, these outer parasitic effects have to be stripped off. This is called de-
embedding.
In the slide above, a brief example on how de-embedding returns the real, inner DUT
performance without its degradation due to the measurement environment is given. We
can clearly see how the transistor cutoff frequency fT is degraded due to these parasitics.
Chip Carrier of a Packaged Transistor and its Parasitic Components

The bipolar transistor is mounted with its Collector to the carrier, Base is bonded, and the
Double-Emitter contact (bonded twice) goes to the backside of the carrier.
If we are going to characterize a transistor mounted to a chip carrier, this carrier will
distort the performance of the inner transistor. Yet, the chip carrier can be described with
series inductors of the bond wires and the parallel capacitors of the pads, as sketched
above.
Z and Y Matrix Manipulations to De-embed the Inner Transistor
49
By rearranging the schematic components, we obtain a simplified equivalent schematic as

given above.
In order to strip off the parasitic components from the outside towards the inner DUT, the
slide shows the corresponding de-embedding procedure.
De-embedding Procedure
We have measured the S-parameters Stotal of the transistor including the carrier
parasitics. With the assumption of having nothing but parallel capacitors as 'outer'
parasitic components, we transform the S-parameters to Y, because a Y matrix represents
a PI structure of components. A simple subtraction will de-embed the parasitic capacitor
effects.
Now, the new 'outer' parasitic components are the two inductors, which are in series with
the chip connections. Series parasitics can be easily eliminated by subtracting a Z matrix.
Therefore, we transform the resulting Y-parameters from above into Z-parameters and
subtract the inductors.
These Z-parameters are finally transformed back into S-parameters which now describe
exclusively the performance of the 'inner' chip.
Particularly for on-wafer measurements, de-embedding by subtracting the complete Y

matrix (OPEN) and Z matrix (SHORT) is often applied. However, in this case, it must
absolutely be assured that
There are no hidden series components present for Y-matrix subtractions (mixed
cross-talk represented by chains of C-L-C-L etc.).
There are no hidden parallel components present for Z-matrix subtractions.
Related to a two-step de-embedding:
For on-wafer measurements (OPEN > SHORT sequence): the SHORT has been
de-embedded from the OPEN dummy.
For packaged measurements (SHORT > OPEN sequence): the OPEN has been
de-embedded from the SHORT dummy.
There are no hidden delay line effects present when subtracting the Y or Z-
matrices.
The slide above sketches the de-embedding situation for a transistor with first
capacitive parasitics, followed by inductive parasitics.
50
De-embedding shifts the Calibration plane
And this slide above once again visualizes the shift of the calibration plane due to de-
embedding.
This slide depicts a three-dimensional representation of such a MOS transistor in its on-
wafer test environment. It is interesting to note that the size of the transistor itself is very
small compared to the size of the pads!
The goal of de-embedding is to shift the NWA calibration plane, which was (after the NWA
calibration) at the end of the probe tips, down to the beginning of the transistor. This
'beginning' is one of the key points for a good de-embedding and a key point for good
dummy structures.
In the above slide, in the section 'Magnification of the transistor' is the inner transistor,
which we want to model. All the rest has to be de-embedded, i.e. to be stripped-off.
In other words, the NWA calibration plane has to be shifted down here.
51
Note
Watch how 'high' the metal planes are relative to the active silicon area. A real mountain of metalization.
The pre requisite for a correct de-embedding is that certain test structures are available
on a wafer together with the device under test (DUT) itself. Depending on the selected de-
embedding method, an OPEN and SHORT dummy structure is required and must be
measured. For a de-embedding verification, also a THROUGH dummy structure is
necessary.
The principle layouts of these structures are given in the actual and the next slides. These
layouts are for Ground-Signal-Ground Probes (GSG).
Please check the 'after-de-embedding' calibration plane as marked in the figures.
Everything, every part of the DUT included in these cal. plane will be part of the DUT
model. In other words, you can think of de-embedding as a shift of the current calibration
plane to these new limits on the wafer. The limits of the blue surrounded area in the OPEN
dummy will become the shifted calibration plane. All what is inside becomes part of the
transistor model.
The SHORT dummy again refers to the desired shifted calibration plane. All what is inside
this plane is now filled up with metal, so that we can consider this part to behave ideal,
while the striplines from Gate and Drain will behave like inductors and will show up with
the SHORT dummy measurement.
Another smart idea for Dummy Test Structures, avoiding the

drawbacks of the SHORT metal plane
This is the THRU dummy, for de-embedding verification purpose. Instead of the transistor,
52
we now have a strip line (metal 1).
See the next slides for a proposed de-embedding verification procedure..
Verifying the de-embedding
Instead of the 'classical' layout shown in the previous slides, today's layouts for silicon
wafers look more and more often like the one shown in the slide above.
It avoids the effects of the lossy silicon substrate by using the 1st metal plane to shield
the contact pads and the lines to the DUT against the lossy silicon substrate. Especially
interesting is the drastically improved SHORT dummy performance, since it applies a
series of vias to ground (metal 1) at the end of the SHORT, rather than having - as with
the conventional SHORT- a large metal plane (the SHORT) which is considered as ideal,
while the two microstrip lines (to the contact pads) of the SHORT dummy are considered
as non-ideal. Also, the OPEN is much more ideal with this alternate approach.
Altogether, this layout suggestion of using the 1st metal as a shield features lower
parasitics. Therefore, the de-embedding is not so much the 'difference of big numbers' like
in the conventional layout suggestions, and therefore more robust.
Under and over de-embedding
Verifying the de-embedding procedure is very important before applying it to the very
DUT, i.e. the transistor, the passive RF component etc.
Without this verification step, errors or problems associated with the de-embedding will
add to the performance of the inner, de-embedded device, and, thus, lead to a wrong
measurement result, and finally to a wrong device model. This is especially critical when
de-embedding (subtracting) the complete Y and Z matrices of the OPEN and SHORT
dummy, because, in this case, possible problems may not show up compared to de-
embedding of individual, lumped components. (With de-embedding lumped components,
we would have detected such problems because they would show up with non-physical
values).
Therefore, it is suggested to model every dummy structure before simply subtracting their
total matrices in order to verify its de-embedding prerequisites:
for the OPEN dummy: no series parasitics allowed.

for the SHORT dummy: no parallel parasitics allowed
verify the de-embedding procedure with a well-known 'golden device', e.g. the
proposed THRU strip line, before applying the de-embedding procedure to the very
DUT.
Some remarks on the under- and over-de-embedding.

UNDER-DEEMBEDDING
If under-deembedding occurred, i.e. we deembed 'not enough parasitics', what means the
lumped components values were estimated too small, the de-embedded traces will not be
'turned back' sufficiently, resulting in a too big a delay time. The de-embedding error will
add up to the inner device and distort its model.
The better visible effect is
OVER-DEEMBEDDING
53
If over-deembedding occurred, however, the deembedded DUT curves will 'turn
backwards' for higher frequencies or show some other non-physical effects. They may
even turn outside the Smith chart!
This emphasizes once again the need to accurately verify the de-embedding procedure
before applying it to the very DUT.
54
Measurements
Contents
Measurement Techniques (iccapmhb) On the Web

Data Verification (iccapmhb)
Data Management (iccapmhb) Articles
Knowledge Center
Technical Support
Training
Videos
55
Measurement Techniques
Contents
Prerequisites (iccapmhb) On the Web

DC (iccapmhb)
CV (iccapmhb) Articles
Network Analyzers (iccapmhb) Knowledge Center
Nonlinear Network Analyzers (iccapmhb) Technical Support
Time Domain (iccapmhb) Training
Noise (iccapmhb) Videos
56
Prerequisites
Contents
Wafer Prober (iccapmhb) On the Web

Switching Matrix (iccapmhb)
Test Fixtures Articles
Intercontinental Microwave Test Fixtures (iccapmhb) Knowledge Center
Cables Technical Support
Physical Properties of Triax and Quadrax Cables (iccapmhb) Training
FMAX Cables (iccapmhb) Videos
Adapter Connectors
Agilent RF Connectors General Information (iccapmhb)
Agilent RF Connectors Adapters (iccapmhb)
Connecting Triax SMU Outputs to Coax Inputs (iccapmhb)
57
Wafer Prober
Slide 1
Slide 2
Let's commence with some nomenclature definitions first:

> Wafer> Die (an assembly of repeatable test structures)> Module (e.g a set of MOS
transistors of different geometry)
In the slide above, you see a Wafer, with the orientation flat at the bottom. Furthermore,
a Die is marked and also zoomed in.
Within the dies are the Modules, see the next slide.
Slide 3
A Die contains several Modules. Each Module contains a set of Test Structures (for
production tests, for device modeling etc.)
Their location is typically specified with respect to a corner of the Die.
Several Modules may be grouped within a Block. But Blocks are organizazion-only.
Slide 4
A Module, finally, represents an area with typically several devices for modeling (for MOS,
a set of transistors of different Ls and Ws).
It can be considered as a wafer probing position, on which by use of a switching matrix
the different measurements are performed:- DC and CV for device modeling (parametric
test, multiple bias sweeps)- Spot measurements (single bias points) for PCM devices
Slide 5
58
This is a typical DC-CV measurement environment:

> a wafer is placed on the chuck of a wafer prober> the chuck can be heated or cooled
(thermochuck)> applying a needle card together with a switching matrix allows to contact
many test structures at a time, and to connect the individual test structure to the DC
analyzer or the CV meter.
IC-CAP distinguishes between measurement instruments and auxiliary measurement
equipment. Therefore, the DC analyzer and the CV meter are controlled from the built-in
IC-CAP instruments drivers, while the thermochuck, the waferprober and the switching
matrix are controlled by macros.
TASK:
> load the file
3_MEAS_ORGANIZE_n_VERIFY_DATA/5_wafertest/wafertest_DC_CV_PELdep.mdl
and find out how the thermochuck, the wafer prober and the switching matrix are
controlled.
Slide 6
The following slides about wafer prober and the corresponding text are a copy of sections
of the file 'DC-CV Training Presentation.pdf' of the Cascade Microtech, Winter 2002
European Probing Seminar CD-ROM, with friendly permission of Cascade Microtech,
Europe
Cascade adds two additional chuck guards to minimize stray capacitance, leakage and
noise for IV measurements. The Cascade FemtoGuard surrounds the chuck bottom and
sides with guard voltage to eliminate coupling the force to the shield below. The new
patented AttoGuard extends the guard over the chuck to minimize force-to-shield coupling
from above.
Slide 7
A typical i/v measurement setup requires probes and chuck bias. The patented AttoGuard
& FemtoGuard completely surround the wafer with guard allowing only a small orifice for
the guarded probes to enter.
Slide 8
59
Slide 9
Since the MicroChamber is at shield potential, only guarded coax probes are required
inside. Solid PTFE dielectric coax exhibiting < 1 femtoamp noise and <50 femtofarad
residual capacitance is used. The probe bodies come with single, quasi-kelvin or kelvin
connectors. The quasi-kelvin style probe is most popular because it is common to both IV
and CV measurement systems.
Slide 10
Different to the so far discussed DC-CV measurement scenario, this is a typical setup for
S-parameter measurements.
NOTE:As a special remark, watch out for the triax->coax adapter used to feed the SMU
bias (triax) into the S-parameter testset bias input (coax) of the NWA.These adapters
must not connect the middle shield of the triax to the outside common shield. The middle
shield of the triax cables must be absolutely unconnected!
Slide 11
60
Different to making DC and CV measurements on a prober, several high-frequency

aspects have to be considered for successful and correct GHz measurements.
As a general rule, for GHz measurements, go for the most reliable and best equipment
available! Do not make improvisations!
Slide 12
Slide 13
Slide 14
61
Slide 15
The screenshots above and of the next slides is from the Modelfile
demo_features\3_MEAS_ORGANIZE_n_VERIFY_DATA\5_WAFERTEST\
1_wafertest_DC_CV_indiv_mdm_files_2006\1_wafertest_DC_CV_indiv_mdm_files_2006.mdl
Slide 16
Slide 17
62
Wafer probers are best controlled by using PEL programs in IC-CAP.

For this purpose, the PEL commands
HPIB_open()
HPIB_write()
...HPIB_close() are available. This PEL programming allows to support any wafer prober.
The screenshot above is of the Modelfile
All different prober-control Macros (Transforms) like chuck_up, chuck_down, goto_x_y
etc. call a central wafer prober control Program 'do_prober_cmd', which finally performs
all the communication with the prober.
Slide 18
Slide 19
Slide 20
63
Slide 21
Slide 22
Slide 23
64
The screenshots of this series of slides is from the IC-CAP MOS Measurement Toolkit
85194K
Slide 24
Slide 25
Slide 26
Slide 27
Switching Matrix
65
Agilent B2200
There are two possibilities to control switching matrices in IC-CAP.
Either by using the built-in functions, or (giving you more flexibility) by simple PEL
programs.
For this purpose, the PEL commands HPIB_open() HPIB_write()...HPIB_close() are
available. This PEL programming allows to support any switching matrix.
Below is an example screenshot of the Modelfile
More PEL examples for driving switching matrices are available under
demo_features\2_MEAS_INSTRUMENTS\3_PROBER_MATRIX_THERMOCHUCK\1_Matrix\
Slide 1
Slide 2
Slide 3
Slide 4
66
Slide 5
Physical Properties of Triax and Quadrax Cables
Triax CableKelvin Triaxial Cable (Quadrax)
ModelLengthCapacitance
16494A Triaxial Cable
Option 0011.5 m125 pF
Option 00380 cm 75 pF
16494B Kelvin Triaxial (Quadrax) Cable
Option 00380 cm 90 pF
16494C Kelvin Triaxial (Quadrax) Cable for 4142
Notes:
How to Select Options for Cable Length:
The Agilent 4155C/4156C or Agilent 4142B have low current forcing and measurement
capability by their SMU (Source Monitor Unit) technology. The SMU uses a kind of
feedback circuit for maintaining very high stability of the current/voltage sourcing.
Therefore, the guard capacitance (a load for SMUs)of connections to SMU must be within
certain limits to assure the accurate functioning.
From experimental measurements, the limit is about 900 pF for the Agilent 4155C/4156C
or the Agilent 4142B, for example.
You should select proper cable lengths to keep small guard capacitance for the SMUs.
A system using switching matrices sometimes needs longer cable lengths than for the
standalone instruments. If you plan to connect more than ~5 m total cable length (both
instrument to switching matrix and switching matrix to the DUTs), make sure the cable
guard capacitance does not exceed 900 pF.
67
FMAX Cables
Slide 1
68
Agilent RF and Microwave Test
Accessories
Overview
Many coaxial connector types are available in the RF and microwave industry, each
designed for a specific purpose and application. For measurement applications, it is
important to consider the number of connects/disconnects, which impact the connector's
useful life.
The frequency range of any connector is limited by the excitation of the first circular
waveguide propagation mode in the coaxial structure. Decreasing the diameter of the
outer conductor increases the highest usable frequency; filling the air space with dielectric
lowers the highest usable frequency and increases system loss.
Performance of all connectors is affected by the quality of the interface for the mated pair.
If the diameters of the inner and outer conductors vary from the nominal design, if plating
quality is poor, or if contact separation at the junction is excessive, then the reflection
coefficient and resistive loss at the interface will be degraded.
A few connectors, such as the APC-7, are designed to be sexless. Most are female
connectors that have slotted fingers, which introduce a small inductance at the interface.
The fingers accommodate tolerance variations but reduce repeatability and may ultimately
break after 1000 connections. Agilent offers slotless versions of connectors in certain
measuring products, which decrease inductance and increase repeatability.
The following is a brief review of common connectors used in test and measurement
applications:
APC-7 (7 mm) connector

The APC-7 (Amphenol Precision Connector-7 mm) offers the lowest reflection coefficient
and most repeatable measurement of all 18 GHz connectors. Development of the
connector was a joint effort between HP and Amphenol, which began in the 1960s. This is
a sexless design and is the preferred connector for the most demanding applications,
notably metrology and calibration.
Type-N connector
The type-N (Navy) 50-ohm connector was designed in the 1940s for military systems
operating below 4 GHz. In the 1960s, improvements pushed performance to 12 GHz and
later, mode-free, to 18 GHz. Agilent offers some products with slotless type-N center
conductors for improved performance to 18 GHz. Agilent type-N connectors are
completely compatible
with MIL-C-39012. Certain 75-ohm products use a type-N design with smaller center
conductor diameters, and thus are not compatible with 50-ohm connectors.
SMA connector
The SMA (Subminiature A) connector was designed by Bendix Scintilla Corporation and is
one of the most commonly used RF/microwave connectors. It is intended for use on
semirigid cables and in components that are connected infrequently. Most SMA connectors
have higher reflection coefficients than other connectors available for use to 24 GHz
because of the difficulty to anchor the dielectric support.
3.5-mm connector
The 3.5-mm connector was primarily developed at Hewlett Packardnow Agilent
Technologies, with early manufacturing at Amphenol. Its design strategy focused on
highly-rugged physical interfaces that would mate with popular SMA dimensions, allowing
thousands of
repeatable connections. It is mode-free to 34 GHz.
1.0-mm launch
The launch adapter has a 1.0-mm female connector on one end and a glass to metal seal
interface on the other end. This is for transition of ultra-high frequency (up to 110 GHz)
signals from coax into a microstrip package or onto a circuit board.
Adapters, Cables and Connectors - General Connector

Information
2.92-mm connector
The 2.92-mm connector mates with SMA and 3.5-mm connectors and offers mode-free
performance to 40 GHz.
2.4-mm connector
69
The 2.4-mm connector was developed by HP, Amphenol, and M/A-COM for use to 50 GHz.
This design eliminates the fragility of the SMA and 2.92-mm connectors by increasing the
outer wall thickness and strengthening the female fingers. It can mate with SMA, 3.5-mm
and 2.92-mm with the use of precision adapters. The 2.4-mm product is offered in three
quality grades; general purpose, instrument, and metrology. General purpose grade is
intended for economy use on components, cables, and microstrip, where limited
connections and low repeatability is acceptable. Instrument grade is best suited for
measurement applications where repeatability and long life are primary considerations.
Metrology grade is best suited for calibration applications where the highest performance
and repeatability are required.
1.85-mm Connector
The 1.85-mm connector was developed in the mid-1980s by Hewlett Packard now
Agilent Technologiesfor mode-free performance to 65 GHz. HP offered their design as
public domain in 1988 to encourage standardization of connector types; a few devices are
available from various manufacturers for research work. The 1.85-mm connector mates
with the 2.4-mm connector and has the same ruggedness. Many experts have considered
this connector to be the smallest possible coaxial connector for common usage up to 65
GHz.
1.0-mm connector
Designed to support transmission all the way to 110 GHz, this 1.0-mm connector is a
significant achievement in precision manufacturing resulting in a reliable and flexible
interconnect.
BNC connector
The BNC (Bayonet Navy Connector) was designed for military use and has gained wide
acceptance in video and RF applications to 2 GHz. Above 4 GHz, the slots may radiate
signals. Both 50-ohm and 75-ohm versions are available. A threaded version (TNC) helps
resolve leakage for common applications up to 12 GHz.
SMC connector
The SMC (Subminiature C) is much smaller than an SMA connector, making it suitable for
some applications with size constraints. It is often used up to 7 GHz where low leakage
and few connections are required.
Connector care and signal performance

While many Agilent RF/microwave connectors have been designed for rugged mechanical
interfaces, the user must be aware that cleanliness of the surfaces and care in applying
torque to the connector nut are crucial to long life and full signal performance. The
following table shows the recommended torque for various connector types.
Table: Recommended torque values for connectors
Connector type Torque (lb-inch (N-cm))

Precision 7 mm 12 (136)
Precision 3.5 mm 8 (90)
SMA 5 (56)
Use the SMA torque value to connect male SMA
connectors to female precision 3.5-mm connectors.
Use the 3.5-mm torque value to connect male 3.5-mm
connectors to the female SMA (8 lb-inch).
Type-N Type-N connectors may be connected finger
tight. If a torque wrench is used, 12 lb-inch
(136 N-cm) is recommended.
Metrology/Instrument Grade Adapters

Table: Metrology/instrument grade adapter selection guide
Typical Configuration
70
Agilent 11900A Agilent 11900C

Agilent 11901A Agilent 11901C Agilent 11903A Agilent 11852B
Agilent 11901D Agilent 1250- Agilent 11525A
Agilent 11904A Agilent 11852B
Agilent 83059A Agilent 11904C 1636 Option 004
Agilent 1250- Agilent 11904D Agilent 1250- Agilent 1250-
Agilent 1250-0597
1159 Agilent 83059C 1743 1698
Agilent 1250- Agilent 1250- Agilent 11524A
1748 1462
85058-60007 85058-60009
Agilent 11903D
Agilent 1250- Agilent 1250-1249 Agilent 1250-
Agilent 11900B Agilent 1250-
1250 0778 0176
Agilent 11901B Agilent 11533A Agilent 1250- Agilent 1250-
Agilent 11904B Agilent 1250- 1744 1475
Agilent 83059B 1746 Agilent 1250-
Agilent 1250-
1528 Agilent 1250-1397
1158
Agilent 1250- Agilent 11903C Agilent 1250-
1749 Agilent 1250- 0559
85058-60008 Agilent 11534A 1562 Agilent 1250-
Agilent 1250- Agilent 1250- 0777
1747 1750 Agilent 1250-
1472
Agilent 1250- Agilent 1250-
1529 0846
Agilent 11903B
Agilent 1250-
1745
Agilent 1250-
1772
Metrology Grade Adapters
Typical Precision Adapter Performance
Instrument Grade Adapters
71
Selected Instrument Grade Adapters
1 Agilent 1250-1744 Adapter, 3.5 mm (f) to Type-N (m), DC to 18 GHz

2 Agilent 1250-1743 Adapter, 3.5 mm (m) to Type-N (m), DC to 18 GHz
3 Agilent 1250-1747 SMA (f) to APC-7 Adapter
4 Agilent 1250-1746 SMA (m) to APC-7 Adapter
5 Agilent 1250-1750 3.5 mm (m) to Type-N (f)
6 Agilent 1250-1745 3.5 mm (f) to Type-N (f)
7 Agilent 1250-1748 3.5 mm (m) to 3.5 mm (m) Instrument-Grade Adapter
8 Agilent 1250-1749 3.5 mm (f) to 3.5 mm (f)
General Purpose Grade Adapter Selection Guide
72
73
1 Agilent 1250-1200 Adapter, BNC (f) to SMA (m)

2 Agilent 1250-1899 Adapter, BNC (f) to SMB (m)
3 Agilent 1250-0556 Adapter, BNC (f) to WECO Video (m)
4 Agilent 1250-1477 Standard, N (f) to BNC (m), Precision 50
5 Agilent 1250-1473 Standard, N (m) to BNC (m), Precision 50 Adapter
6 Agilent 1250-0595 Adapter, BNC (f) to BNC Triaxial (m)
7 Agilent 1250-1930 Adapter, BNC (m) to BNC Triaxial (f)
8 Agilent 1250-1830 Adapter, BNC (f) to BNC Triaxial (f)
9 Agilent 1250-1857 Adapter, SMB (f) to BNC (m)
10 Agilent 1250-0562 Adapter, BNC (f) to SMA (f)
11 Agilent 1250-1236 Adapter, SMB (f) to BNC (f)
74
1 Agilent 1250-1391 Adapter, SMB Tee (f) (m) (m)

2 Agilent 1250-1741 SMA (f) to SMA (m) Right Angle Adapter
3 Agilent 1250-1698 Adapter, SMA Tee (m) (f) (f)
4 Agilent 1250-1249 Adapter, SMA Right Angle (m) (f)
5 Agilent 1250-1462 Adapter, SMA (m) to SMA (f)
6 Agilent 1250-0674 Adapter, SMB (m) to SMA (f)
7 Agilent 1250-1694 SMA (m) to SMA (f) Adapter
8 Agilent 1250-1158 SMA (f) to SMA (f) Adapter
1 Agilent 1250-0597 Adapter, Type-N (m) 50 to Type-N (f) 75

2 Agilent 1250-1778 Standard N (m) to Standard N (m) Adapter, 50
3 Agilent 1250-1529 Standard N (f) to Standard N (f) Adapter, 75
4 Agilent 1250-1152 Adapter, SMC (f) to Type-N (m)
5 Agilent 1250-1404 Adapter, SMA (f) to Type-N (f)
6 Agilent 1250-1023 Adapter, SMC (m) to Type-N (m)
7 Agilent 1250-1535 Adapter, N (m) to BNC (f) Adapter, 75
8 Agilent 1250-1533 Standard N (m) to BNC (m) Adapter, 75
9 Agilent 1250-1250 Adapter, Type-N (m) to SMA (f), 50
10 Agilent 1250-0846 Tee Adapter, Standard N (f) (f) (f)
11 Agilent 1250-1636 Adapter, Type-N (m) to SMA (m) 50
12 Agilent 1250-0559 Tee Adapter, Standard N (m) (f) (f)
13 Agilent 1250-0176 Right Angle Standard N (m) to Standard N (f)
Adapter Kit Selection Guide
1.0 mm Adapters
Increased measurement versatility
Ease-of-use for on-wafer and coaxial measurements
Increased Measurement Versatility

For Microwave and RF engineers making coaxial measurements at 50, 67 or 110 GHz, the
75
Agilent 11920/1/2 series 1.0 mm adapters provide an easy way of measuring coaxial
devices at high frequencies. The Agilent 11920 A/B/C 1.0 mm to 1.0 mm are designed for
the measurement of components with 50 1.0 mm connectors. The Agilent 11921
A/B/C/D, 1.0 mm to 1.85 mm, and the Agilent 11922 A/B/C/D, 1.0 mm to 2.4 mm, are
intended to be used as general purpose adapters that are versatile and interchangeable.
These adapters increase the capability needed to use test systems, such as the Agilent
N5250A.
Ease-of-use for on-wafer and Coaxial Measurements

Each connector has an air dielectric interface and a center conductor that is supported by
a low-loss plastic bead. Available with male and female connectors, these Agilent 1.0 mm
adapters provide ease-of-use for microwave engineers who need to connect their test
systems. The Agilent 1.0 mm adapters allow engineers to make fewer connections directly
to their test port while maintaining the accuracy of their test system.
1.0 mm Adapters Table
Flexible Microcircuit Packaging

The Agilent 11923A 1.0 mm female connector launch threads into a package or fixture
housing to transition a microwave circuit from microstrip to coaxial connector. The Agilent
11923A connector launch is intended for use with the N5250A and other test systems up
to 110 GHz. The Agilent 11923A 1.0 mm female connector has an air dielectric interface
and center conductor that is supported by a low-loss plastic bead on one end and a glass-
tometal seal interface on the other end. This interface consists of a 0.162 mm diameter
pin that extends inside the package or fixture for connection onto a microwave circuit.
The Agilent 11923A is pre-assembled and supplied with a machining detail for mounting
the launch and assembly instructions. The user is responsible for making the connection
onto the circuit card, machining the package, and installing the connector. If a quasi-
hermetic seal is desired, epoxy may be applied to threads of the launch prior to
installation. The procedure describing the necessary dimensions for the package and
installation is provided with the launch assembly.
76
Specifications
Specifications describe the instrument's warranted performance over the temperature
range 0 to 55 C (except where noted). Supplemental characteristics are intended to
provide information for applying the instrument by giving typical but nonwarranted
performance parameters. These are noted as "typical", "nominal", or "approximate".
1.0 mm (f) Connector Launch
Supplemental Characteristics
Environmental Specifications
Note
The operating temperature is a critical factor in the performance during measurements and between
calibrations. Storage or operation within an environment other than that specified above may cause
damage
to the product and void the warranty.
Non-operating environmental specifications apply to storage and shipment. Products should be stored in a
clean, dry environment. Operating environmental specifications apply when the product is in use. Products
should
not be operated in a condensing environment.
Key literature
Agilent 11923A Operating and Service Guide 11923-90001
Slotless Connectors
Precision Slotless sockets (female connectors) were developed by Agilent to provide the
most accurate traceable calibration possible. Connectors that use precision slotless
sockets are metrology grade connectors. The outside diameter of the socket does not
change
when mated with pins of varying diameters, within the tolerance requirements of a
metrology grade connector. Conventional slotted sockets are flared by the inserted pin.
Because physical dimensions determine connector impedance, electrical characteristics of
the connector pair are dependent upon the mechanical dimensions of the pin. While
connectors are used in pairs, their pin and socket halves are always specified separately
as part of a standard, instrument, or device under test. Because the slotted socket's outer
diameter changes with different pin diameters, it is very difficult to make precision
measurements with the conventional slotted socket connector. The measurement of the
device is a function of its connector.
Slotless sockets are used in the following calibration kits:
Agilent 85052B/C/D
Agilent 85054B/D
Agilent 85056A/D
Coaxial Mechanical Calibration kits
77
Option description:
1A7 ISO 17025 compliant calibration

A6J ANSI Z540 compliant calibration
UK6 Commercial calibration certificate with test data
00M Includes male standards & male-male adapter
00F Includes female standards and female-female adapter
MOF Includes male and female standards and adapters
001 Adds 2.4 mm sliding load and 2.4 mm gauges
100 Includes female-female adapter
200 Includes male-male adapter
300 Includes male-female adapter
400 Adds four 3.5 mm to Type-N adapters
500 Adds four 7 mm to 3.5 mm adapters
500* Adds four 7 mm to Type-N adapters
Waveguide Mechanical Calibrations Kits
Coaxial Electronic Calibration Kits (ECal)
Option Description
78
00A Add type-N adapters
00A 1 Add 7 - 16 adapters
00A 2 Add 3.5 mm adapters
001 Adds data for Agilent 8702 Lightwave Component Analyzer
Mechanical Verification Kits
Option Description

00A Add type-N adapters
00A 1 Add 7 - 16 adapters
001 Adds data for Agilent 8702 lightwave component analyzer
Triax Feedthru and Triax-Coax Converters from

Trompeter Electronics
Bulkhead Feedthru BJ78
79
Introduction
Coplanar GHz probes have been available only since 1983, and have significantly
advanced wafer probing capabilities. Wafer probing at GHz frequencies has developed
over the past several years from a simple R&D toy to a sophisticated high-volume
manufacturing tool used in GaAs, silicon, packaging, and hybrid industries. Successful GHz
probing requires that consideration be given to layout and design before design
completion and mask fabrication. Failure to observe specific layout requirements can
result in the inability to test devices with GHz probes. This Application Note focuses on
layout and testing issues as they apply to the use of GHz coplanar probes, Discussed first
are specific mechanical and electrical layout rules, rules that will to assure that a
fabricated device can be successfully probed. Following the layout rules are design/layout
guidelines for current major applications of GHz-wafer-probing technology-microwave
monolithic integrated circuits (MMICs), high-speed digital ICs, individual characterization
devices for process characterization and modeling, and high-speed hybrids and packages.
The rules in this Application Note will help you achieve the best possible tests.
It is useful to stop and consider what you really wish to measure. When GHz probes are
used with a network analyzer (after calibration), what is measured will be what is
contacted by the probe tips. This includes the pad parasitics, and parasitics associated
with the interconnects from the pad to the other devices on your substrate. If you want to
measure a device independent of the pad effects, then it is useful to make the pads small,
so their effects will be negligible or easily subtracted from the measured data. In any
event, it is important to be clear that when using a network analyzer you will be
measuring everything past the end of the probe tips, unless you use special
calibration/correction techniques. In most applications this is exactly what you want; a
measurement the circuit's electrical performance.
This Application Note applies to the following Cascade Microtech probe series:
WPH-Oxx -18 GHz probes

WPH-1 xx- 26 GHz probes
RTP-105-xxx-26 GHz replaceable-tip probes
WPH-2xx-50 GHz probes
WPH-3xx -`40 GHZ probes
WPH-7xx -multicontact probes
WPH-9xx -multiple-needle probes
Probe Features Affect Layout

A coplanar probe's features (Fig. 1) include the probe body, the coaxial connector, the
probe tip, and the contacts at the probe tip end. Coplanar transmission lines carry the
signal between the coaxial connector and the probe tip contacts.
Figure 1: Coplanar probe
Figure 2: Probe tip: top, bottom, and side view
80
The transition from coaxial, with a radial electrical field pattern (Fig. 3a), to a coplanar
transmission line with a much different electrical field pattern (Fig. 3b) is made within the
probe. This is a difficult transition to make at GHz frequencies, and presented difficulties in
successfully designing this type of probe. The best DUT tests result if the electric field
patterns at the probe tip are similar to the field patterns in the DUT. If transitions are
necessary, they should be made cleanly.
The typical probe contact is a signal (S) or ground (G) contact. The signal contacts are
electrically connected to a coaxial connector center pin, and the ground contacts are
electrically connected to the coaxial-connector body. At high frequencies the ground is not
an equipotential reference, as in low frequency circuits, but is a part of a transmission line
that contains time-varying electric fields. The 700-series multicontact probes also permit
specified contacts to be bypassed ("P"-discussed later in this Application Note), and
terminated ("T"). (See the Cascade Microtech Probe Head Selection Guide for additional
details).
Figure 3: Electric field patterns in coaxial and coplanar transmission lines.
Figure 4 shows an example MMIC pad footprint. Note the

input is on the left, the output on the right, and that the power and control signals at the
top are provided by low-frequency needle-type probes. The GHz probes are GS (ground-
signal) probes, useful to 26 GHz. The bottom row of pads are on the same bus, providing
a low-impedance ground for the two GHz probe ground contacts. When packaged, there
will be several bond wires connecting the ground buses to the package ground. This
minimizes common-ground inductance. Also, note that the amplifier is positioned close to
the ground bus in order to minimize common-ground inductance. In practice, the
individual transistors would be positioned very close to the ground bus.
Figure 4: Typical MMIC padout, ground-signal probes in contact.
There are several physical features of the coplanar probes that effect the layout. The first
of these is the extension of the probe tip substrate material past the center of the contact
closest to the edge. This dimension is typically 100m.
81
Note
This page (after Figure 4, is a screen shot and this need to be typed.
There are several physical features of coplanar probes that effect the layout. The first
phase of these (Fig.5) is the extension of the probe tip substrate material past the center
of the contact closest to the edge. This dimension is typically 100m. Probes must not be
allowed to collide, so this 100m extension must be considered when laying out pads.
Another physical layout consideration is the proble contact size (Fig.6), typically 50 x
50m. Note that these contacts are much larger than needle probes, and the passivation
windows must be large enough to accomodate these contacts plus typical probe system
variations.
The final consideration is the probe skating distance. For every 50m of overtravel
(overtravel is the continued downward movement after the probe tip has made initial
contact with the wafer) the probe contacts will skate laterally 10m. If the probes are too
close together, they could skate into each other and be damaged. The layout rules will
help you prevent probe collisions.
Sequence of Layout Events

The following are guidelines to the typical sequence of events involved in layout for testing
with GHz probes. In many cases these guidelines will improve the packaged device
performance. Specific requirements may result in these guidelines being altered.
1. Select basic pad footprint for signal, power, control, and ground pads to be
consistent with input/output type (see step 2) and package requirements. Example
pad footprints are shown throughout this Application Note.
2. Select high frequency input/output type to meet your specific frequency, design,
and signal count requirements; either signal ground (GSG), or signal-signal (SS).
3. Orient high-gain signal pads away from each other. Typical microwave circuits
place inputs on the left, and outputs on the right.
4. Use at least one ground pad for each GHz probe. The grounds pads for each GHz
probe must be electrically connected together on the DUT to provide a low-
impedence, common-ground node.
5. Use additional ground pads between signal pins to obtain reduced crosstalk (as
required) within multicontact GHz probes.
6. Verify that pad pitch matches available probe pitch (see Cascade Microtech
Probe Head Selection Guide). Also, verify that the pad footprint meets packaging
requirements.
7. Verify that pad pitch matches avaialble probe pitch (see Cascade Microtech
Probe Head Selection Guide). Also, verify that the pad footprint meets packaging
requirements.
8. Call the Cascade Microtech Application Group to verify that the footprint is a
good layout for probing. If you want to violate some of the rules, please call the
Cascade Microtech Applications Group regarding the trade-offs. Most of the rules can
be successfully violated, provided you meet certain conditions.
Layout Rules, Mechanical

This section describes the mechanically related rules regarding probe pad placement and
sizing. In most designs the probe pads are also used as bond pads, so the final layout
should satisfy requirements for both probing and bonding. A good layout for GHz probing
is usually also a good layout for packaging.
There are two important GHz-probe mechanical parameters which directly affect the
mechanical layout rules; the contact dimensions and the substrate extension past the
contact. Typical contact size for WPH-series probes (Fig. 6) is 50 x 50m for contact pitch
greater than 100 m,tapering to 35 x 35 m as the contact pitch approaches 50 m.The
contact size is important because the passivation window must be large enough to allow
the entire contact to touch the pad. The substrate extension is typically 100 m past the
82
contact center (Fig. 5).
Coplanar-waveguide GHz probes are constructed with all contacts in a straight line, and
with equal spacing between contacts. This means that all pads, arranged for an individual
GHz probe measurement, must be in a straight line, have equal pitch, and be on the
substrate top. Non-square or very large pads may be laid out more randomly, as long as
there is a minimum-size land area, spaced at an equal pitch for the probe contacts to land
on.
Top-side pads connected to substrate vias must be used to contact backside grounds. All
pads must have as small as possible height variation, because the GHz-probe construction
allows little contact height variation. If GHz probes crash into each other, they can be
damaged. The pad layout must allow for sufficient damage-avoidance space between GHz
probes. Provide sufficient space for probes that are oriented 90 degrees to each other, as
well as between probes that face each other. There must be sufficient space between
probes to allow for probe skating. The following pad layout rules allow for vertical
overtravel of 20 mils (500 m).Note that 20 mils is the maximum rated overtravel for
Cascade Microtech GHz probes, and that the recommended nominal overtravel (to achieve
assured contact) for the typical probe is 2-4 mils (50-l00m).
The mechanical layout rules are given three-digit numbers, starting with 100. This number
also indicates the applicable dimension within the referenced figures. Note that the
minimum and maximum values given are the recommended values for general testing
conditions. If you have a specific need outside these values, please call the Cascade
Microtech Applications group for consultation.
Rule 100 Pad size The minimum pad size is 50 x 50 m(Fig. 7). The recommended
minimum pad size for general use is 100 x 100 m.Note that because the probe contacts
are fairly large, in many situations pads smaller than 50 x 50 m can be successfully
probed. Unless you have a good reason to do otherwise, use 100 x 100 m pads with a
150 +rnpitch. This provides for easy probing, and parasitic are low enough for many
applications.
Figure 7 Rules 100 and 102 Minimum pad size is 50 x 50m Minimum pad pitch is 50m
Rule 101 Passivation window The minimum passivation window size is 96 x 96 km(Fig.
8). Smaller passivation windows are acceptable for probe pitch values less than 100
km,because the contact size decreases as contact pitch becomes less than 100 wrn(Fig.
6). Note that for small pad sizes, the passivation window will be significantly larger than
the pad. If the pad metalization is above the final passivation layer, this rule does not
apply.
Figure 8 Rule 101 Minimum passivation window size is 96 x 96 m.
Rule 102 Pad pitch The minimum center-to-center pad pitch is 50m I*m(Fig. 7). The
recommended minimum is 150 m.Note that many probes have a 100 m minimum pitch
specification. (See the Cascade Microtech Probe Head Selection Guide for additional
details.)
Rule 103 Orthogonal-row pad spacing The minimum center-to-center pad spacing for
orthogonally oriented probes (in a corner) is 200 m(Fig. 9). Sufficient clearance is often
achieved by eliminating the corner pad (Fig. 15).
Figure 9 Rule 103 Minimum center-to-center pad spacing in a corner is 200 m.
83
Rule 104 Single-row pad spacing The minimum center-to-center pad spacing for
opposing, side-by-side probes that are contacting the same line of pads, (Fig. 10)is 200
m.
Figure 10 Rule 104 Minimum pad spacing with opposing, side-by-side probes is 200 m.
Rule 105 Parallel-row pad spacing The minimum center-to-center pad spacing between
facing probes on parallel rows of pads (Fig. 11) is 200 km. Note that this spacing is based
on the assumption of 20 mils (500 m)of probe over-travel. If the probe overtravel can be
limited to 10 mils (250 m), then the minimum spacing is 100 brn.
Figure 11 Rule 105 Minimum center-to-center pad spacing for parallel rows of pads with facing probes is 200 m.
Rule 106 Pad height variation The maximum pad height variation in a row of pads
contacted by one probe is 0.5 m (Fig. 12). Pad height variation usually occurs because
pads are constructed with different metal stacks. Be sure to use the same metal layers to
construct all pads.
Figure 12 Rule 106 Maximum pad height variation in a row of pads contacted by one probe is 0.5 m.
Rule 107 Planarity requirements The maximum overall planar deviation of a row of pads
contacted by one probe, with respect to the backside of the substrate, is 2 parts per 1,000
for WPH-700 series multicontact probes, and 5 parts per 1,000 for other probes (Fig. 13).
Non-planarity can be caused by using different metal stacks for the pads within a row, or
incorrectly backlapping a wafer. Also note that during probing, non-planarity between the
probe contacts and the pads should not exceed these maximum planar deviations.
Figure 13 Rule 107 Maximum non-planarity from a row of pads to the backside of substrate is 2/1,000 for WPH-700
probes, and 5/1,000 for all other probes.
84
Rule 108 Pads at 90 degrees only Rows of pads meeting at an angle (Fig. 14) must be
orthogonal (at 90 degrees).
Rule 109 Pads are collinear All pads contacted by an individual probe must be collinear
(Fig. 14).
Figure 14 Rule 108-ROWS of pads meeting at an angle must be at 90 degrees. Rule 109- Pads
contacted by one probe must be collinear.
Rule 110 Pads have equal pitch All pads contacted by an individual probe must have
constant pitch. Note that the WPH-700 series multi-contact probes allow for contact
removal, and therefore pad removal.
Rule 111 Pads on top All pads must be on the substrate top.
Layout Rules, Electrical

This section describes the electrically related rules regarding pad placement and sizing. In
most designs the pads are also used as bond pads, so the final layout should satisfy
requirements for both probing and bonding.
The following rules are given three-digit numbers starting with 200.
Compared with mechanically related layout rules, the electrical considerations are more
complex. Included in the electrical considerations are:
Probe grounds;
Signal input/output;
Crosstalk between probes;
Crosstalk between signal lines within a probe;
Maximum operating frequency for each probe type:
Power and ground to DUT during probing;
Maximum current per contact.
Rules 200-205
Rule 200 Pads on top All pads must be on the substrate top surface, and all nodes to be
measured or stimulated must connect to these pads. Rule 201 Probe Grounds Each probe
must have at least one ground contact.
Rule 202 All grounds connected All probe ground contacts must be electrically connected
together on the DUT (Fig 15) This connection should be as short a connection as possible,
presenting a low impedance at high frequencies. A useful method for connecting probe
grounds is shown in Fig. 16. This approach uses L-shaped metalization to connect the
probe grounds. Note that this method requires the edge contacts to be defined as ground
for each probe.
Figure 15 Rule 201 -All probes must have at least one ground. Rule 202-All probe grounds must be connected
together on the DUT.
85
Figure 16 Alternative method of connecting the probe grounds together, using L-shaped metalization connecting
ground pads. Requires probes with outside contacts defined as ground.
Rule 203 Signal input/output Signal input/outputs are either signal-ground (SG), ground-
signal-ground (GSG), or signal-signal (SS). Any of these input/outputs may be used for
measurements up to 18 g h Z The GSG arrangement is advised for measurements beyond
18
ghZ, and required beyond 26 GHz. The SS input/output requires differential signals and a
mirrored layout, resulting in a virtual ground. The WPH-700 series probes can be factory
programmed to provide a wide variety of contact sequences, including signal, ground,
bypassed power, and terminated signal. (See the Cascade Microtech Probe Selection
Guide for details.)
Rule 204 High dynamic-current grounds The lowest impedance ground return for high
dynamic-current outputs must be connected to the same probe head that is providing the
power to those outputs (Fig 17).
Figure 17 Rule 202 Where high dynamic currents exist, provide low-impedance grounds to the same probe which
supplies the power.
Rule 205 Maximum rated current The maximum dc current per contact is one-half
ampere. Do not apply more than 1 ampere total to a probe that is in continuous contact,
i.e., no make and break contact.
Crosstalk
If there is good shielding design, crosstalk is generally not a problem, except when testing
very high gain amplifiers, A/D and D/A circuits, or when there are severe reverse
isolation requirements. When using coplanar probes, crosstalk may be observed in the
following situations:
Crosstalk between probes:
Capacitive or radiative coupling between probes.

Common ground inductance in DUT.
Crosstalk between signal lines within a probe:
86
Common ground inductance in probe.
Mutual coupling between signal lines within a probe.
There are layout decisions that will affect crosstalk:
Spacings between input and output pads will affect probe-to-probe crosstalk.
Common-ground inductance in the device layout will affect probe-to-probe crosstalk.
There is the question of whether or not to place an intervening ground pad between
signal pads when laying out circuits to be probed by the WPH-700 series multicontact
probes. The addition of the intervening ground pad will require a corresponding
ground contact in the probe, and therefore will reduce crosstalk between signal lines
within the probe.
Crosstalk Between Probes- The crosstalk between probes due to capacitive and
radiative coupling is shown in Fig. 18. Note that the crosstalk depends on the spacing
between probes, and the material on which the probes are landed. When laying out
circuits where crosstalk between probes is important, space the pads as far apart as
possible.
Figure 18 Crosstalk between two GSG probes
Another factor affecting probe-to-probe crosstalk is the common ground inductance, as

shown in Fig. 19. Current from both the input and output will flow through the common
inductance, resulting in a voltage drop across the common inductance, and part of the
output signal will appear at the input, and vice versa. This internally generated crosstalk
cannot be corrected by calibration or probing techniques. The guideline for layout is to
make the common ground metalization as short and as wide as possible to minimize
inductance.
Figure 19 Schematic layout showing common-ground inductance, resulting in crosstalk
between input and output probes.
Contact sequence and crosstalk Crosstalk between signals within a probe is generally a
consideration only with the WPH-700 series multicontact probes. The main contribution to
crosstalk within a probe, over which you have some control, is the contact sequence.
Specifically, do you place an intervening ground between adjacent signals? For example,
one can specify the following sequences for a WPH-708 multicontact probe (eight
contacts): (1) GSSSSSSG, (2) GSSGSSGP, and (3) GSGSGSGS. An intervening ground will
obviously lower
the crosstalk between the signal lines.
Select the contact sequence that meets your requirements. Typical voltage crosstalk
performance (measured with an HP 54120 digitizing oscilloscope, and using 100 ps rise
time signals for all data):
87
sequence* crosstalk (%) termination
signal-signal
GSGSG 1.0 % sig lines open
GSSG 4.0 % sig lines open
GSGSG 0.3 % sig lines 50 1L-gnd
GSSG 3.5 % sig lines 50 R-gnd
GSGSG 2.0 % sig lines short-gnd
GSSG 14.0 % sig lines short-gnd
*See Fig. 20 for a test circuit schematic.
Figure 20 Circuit used to measure crosstalk between S1 and S2 in WPH-700 series multi-contact probes.
Load L was either an open, short, or 50 ohms. Probe ground-line common inductance
Another
factor affecting signal-to-signal crosstalk within a probe is the common inductance in the
probe ground line. This situation is analogous to the common-ground inductance in a DUT,
as previously discussed. Figure 21 shows a simplified equivalent circuit of a two-signal
probe with common lead inductance Zg12. The amount of crosstalk is shown in Fig. 22 as
a function of frequency and common-lead inductance value.
Figure 21 Simplified equivalent circuit of a SGS probe, showing how common ground inductance Zg12 contributes to
crosstalk between the two signal lines.
Figure 22 Common-ground inductance affects crosstalk between signal lines within a probe.
The guideline in this situation is to be aware of this source of crosstalk, choose probe
families which meet your needs, and to lay out your IC accordingly. The probe families
have different common-lead inductance values and you need to review the Cascade
Microtech Probe Head Selection Guide as you make your decisions. Note that these
considerations also apply to bonding wires used in packaging. The 0.3 nH of the WPH-700
series multicontact probe ground inductance corresponds to a bonding wire 300 km (12
mils) long. Also note that
multiple grounds and power contacts reduce the probe's equivalent common-ground
inductance.
DUT Power There are four typical methods to provide power to the DUT during probing:
(1) Bias tee, (2) WPH-900 multiple-needle probes, (3) WPH-700 multicontact probes, and
(4) single-needle positionable probes. Bias tees, which are high-frequency chokes, are
typically used when probing individual transistors or FETs, and allow the signal lines to be
dc biased. Bias tees that have been installed in the S-parameter test set are usually used
to characterize individual transistors or FETs.
WPH-900 needle-type probes are typically used to provide power to ICs being probed (Fig.
23 ). They work well, except when their inductance (8 nH typical) affects the circuit
88
performance. An inductance of 8 nH will adversely affect many circuits (model your circuit
to check the result), including high-gain amplifiers and fast digital devices. In these cases,
the WPH-700 multi-contact probes are recommended, because one of the available
contact configurations has internal power-supply bypassing right at the probe tip (Fig. 24).
The bypassing network has been specifically designed to be non-resonant over a very
wide frequency when probing ICs. The typical impedance is 0.01 FF in series with 4
ohms and 300 pH. See the WPH-700 multicontact probe brochure for additional details on
how to select and use these probes for this application.
Note that in Fig. 24, the inductance associated with the VSS, VDD and GND metalization
will be about 1 nH each, which is three times that of probe inductance. If you need *
lower inductance, use much wider lines, or make the die smaller, or move the WPH-700
multicontact probe to the side of the chip.
Figure 23 Typical MMIC with WPH-900 probe needles providing power and ground. The needle
inductance can result in circuit malfunction.
Figure 24 Typical MMIC with WPH-700 series probes, configured to provide bypassed dc power.
Specific Applications
MMICs
The use of Cascade Microtech probes to measure MMIC performance is a natural
application. There are four different MMIC layout topologies which encompass most of the
MMICs being manufactured today: (1) coplanar waveguide, (2) lumped element, (3)
microstrip, and (4)
differential or balanced. For the examples shown here, the typical pad size is 100 x 100
m.
A schematic representation of a coplanar waveguide (CPW) topology is shown Fig. 25. This
topology is very easy to probe by using Cascade Microtech probes, and the FETs have
minimal ground inductance because both the circuit and the probe are fabricated with
CPW
topology. The signal input/output should always be GSG (Fig. 25). The circuit is biased
either through the signal lines with bias tees, or through separate pads contacted with
either the WPH-900 multiple-needle probes or the WPH-700 series multicontact probes.
Note the second-level metal crossovers connecting the grounds together, eliminating
ground moding.
Figure 25: MMIC layout, with padouts for probes.
89
A simplified layout of a lumped-element MMIC topology is shown in Fig. 26. Note that in
this topology lumped inductors, capacitors, and resistors are used, rather than
transmission lines. The interconnects on the die are modeled as inductors. Because
lumped element designs are typically used below 18 GHz, the input/output of Fig. 26 is
GS. If you wish to use a GSG input/output, the other ground should be connected to the
DUT ground bus with a large, minimum length of metal (Fig. 27). Power is provided with
WPH-900 multiple-needle probes (Fig. 27)or WPH-700 multicontact probes (Fig. 26).
Figure 26 Schematic of lumped-element MMIC. WPH-700 multicontact probe provides power, and GS probes provide
measurements.
Figure 27 Schematic of lumped-element MMIC. WPH-900 needles provide power, and GSG probes provide
measurements. Note short, wide metal used to connect the two ground pads.
When these die are packaged, multiple-bond wires are used to connect the die ground bus
to the package ground. Additionally, bond wires will connect the input and output pads to
the appropriate package pins. None of these inductive bond wires will be present during
probing, and their effect on circuit performance will not be included in the measurements.
The guideline is to model your circuit without the package and bond wire parasitics
present, and this is what you should measure with the probes.
A simplified example of a microstrip layout is shown in Fig. 28. Microstrip designs always
have a backside ground, and usually have substrate vias to connect the top-side devices
to ground. Because the Cascade Microtech probes require ground and signal to be in the
same plane, substrate vias are necessary to launch signals into a microstrip design.
Figure 28 Typical microstrip-MMIC layout. Substrate vias are required to connect the probe grounds to the IC
backside ground.
90
The parasitic inductance in the coplanar-to-microstrip transition should be minimized by

making the metal lines to the via as short as possible. Do not land on the substrate vias,
as they are typically not flat enough for a probe to reliably contact the vias and a pad at
the same time. Note that you will measure, using the probes, the MMIC performance and
the coplanar-to-microstrip transition formed by the substrate vias. A correction can
be done by measuring and removing the effects of this transition; or you can use a TRL
calibration technique, by moving the reference plane just past the transition so it
will not be included in the measurement. If substrate vias are not available, a technique
using capacitive coupling to the backside is available 1 . Power and control signals
are typically sent through the signal lines with bias tees, or WPH-900 multiple-needle
probes, or WPH-700 multicontact probes.
An example of a differential-layout MMIC is shown in Fig. 29. The unique feature of this
layout is the mirrored symmetry, which results in a virtual ground. Differential input and
output signals are typically used without any ground. Single-ended input signals, output
signals, and power sources require grounds. The use of balanced topologies results in
balanced current requirements, easing power supply bypassing requirements. This
topology has been successfully probed up to 18 GHz.
This technique's limitation is that it requires accurately balanced signals. This is generally
a limitation with the signal generation equipment, and not of the probe phase
match. For WPH-003 18 GHz probes, the phase match is better than 1 ps, and better than
5 ps for WPH-700 multicontact probes.
Figure 29 A differential layout. Circuit A is mirror-imaged about the virtual ground to give circuit A'. The SS input
requires differential signals.
The preceding examples show typical ways of providing power and ground to the DUT. In
many cases, such as high-gain amplifiers, the use of WPH-900 multiple-needle probes to
provide power will result in unwanted oscillations during probing, due to the inductance of
the
needle (typically 8 nH0. In such cases, use WPH-700 multicontact probes to provide
power.
Additionally, remember that when probing your ICs you will be measuring the IC
performance without the package and bond wires, and that the measured performance in
the package will be slightly different than measured on the wafer.
Digital Devices
Most digital GHz-probing applications involve GaAs ICs, high-speed ECL, or mixed
analog/digital with high-speed digital inputs or outputs. Most digital DUTs are probed
with WPH-700 multicontact probes, because of the larger number of required contacts.
Additional reasons to use the WPH-700 multicontact probes are the availability of
bypassed non-resonant power, and ECL output terminations (50 ohms to VTT).
A typical, and recommended, padout is shown in Fig. 30, with typical pad size being 100 x
100 pm. Note that all probe grounds are connected together on the DUT (Fig. 15), and
that the high-current output driver grounds are connected by the lowest impedance path
to the same probe which supplies the power to these drivers (Fig. 17). If situations arise
where high input/output isolation is required, lay out the IC so that the inputs and outputs
are
91
contacted by separate probes (Fig. 31).
Figure 30 Typical digital IC padout
Figure 31 Digital IC example, showing inputs on a probe separate from the outputs. This maximizes isolation
between inputs and outputs.
Characterization Devices
Individual FETs, transistors, capacitors, and inductors are often measured in order to
develop models, or as part of wafer qualification. In laying out these devices it is
important to minimize parasitics associated with the interconnects to these devices, and
the pads. In general, if one is measuring low-impedance devices, then the series-
inductance parasitics should be minimized by using fat, short bars of metal. If high-
impedance devices are to be measured, then the parallel-capacitance parasitics should be
minimized by using narrow interconnects and small pads. Also, the substrate will affect
the final layout.
For example, the parasitic capacitance to ground is much smaller with semi-insulating
GaAs substrates than for conductive silicon substrates. The GaAs industry typically
measures fairly large devices (300 m gate length) as characterization devices. Many
device layouts are constructed with 100 x 100 m pads, and wide metal buses connecting
the pads to the DUT, as shown in Fig. 32a and b. However, within the silicon community
the characterization devices are often minimum-size devices, presenting high impedances
to the
test system. This requires that parasitic capacitances be minimized. To minimize parasitic
capacitance to the substrate (Fig. 32c and d) the pads are small, usually 50 x 50 m, and
the interconnects from the pads to the DUT are narrow. Note that with GSG configurations
the DUT should connect the two ground buses together (Fig. 32a and c).
Figure 32 Layout examples of active characterization devices.
Figure 33 shows typical optimization steps in laying out a characterization pattern for a
GaAs FET. Figure 33a shows the initial layout. The initial layout has the problems of
narrow, high-inductance, connections to the gate and drain, and the common-ground
inductance in the source connection has not been minimized. In Fig. 33b these problems
have been resolved by repositioning the DUT, and by widening the interconnect metal to
the gate and drain. Finally, in Fig. 33c the pads are moved closer together to further
minimize the series-inductance parasitics associated with the metal connecting the gate
92
and drain to the pads. Larger interdigitated FETs are typically measured as shown in Fig.
34.
Figure 33 Optimization of large gAaS FET characterization device.
Figure 34 Example layout of four-gate interdigitated GaAs FET. Note the second-layer metal connecting the sources
together.
In Fig. 35 the analogous situation is shown for minimum-size transistors on silicon

substrates. In Fig. 35a the layout contains large (100 x 100 m) pads, very closely spaced
together. This layout, while excellent for FETs on GaAs is not workable for minimum-size
devices on silicon. The problems with this layout are the parasitic capacitances (Fig. 35b).
Because silicon is conductive, the parasitic capacitances are much higher than with
GaAs or alumina substrates, and can contribute to significant measurement errors, a -
50% error in ft, for example. See the paper "GHz On-Silicon-Wafer Probing Calibration
Methods" 2 for more details.
Figure 35 Optimization of small silicon transistor characterization device. Dummy device is used to correct
measurement for remaining parallel-capacitive parasitics.
Parasitics are minimized in the Fig. 35c layout by using smaller pads (50 x 50 m),
spacing the pads out further (150 to 250 m center-to-center spacing), and narrowing
the metal lines to minimize capacitance to ground. The DUT is located near the probe
ground bus to minimize the common inductance. Also note that when laying out 50 m
pads, the passivation window should be at least 96 x 96 m, and preferably larger, to
accommodate the 50 x 50 m probe contacts, plus typical placement errors. (References
2 and 3 describe techniques to correct measurement errors due to parasitic capacitances.)
The recommended method of minimizing parasitics involves the use of a "dummy" device,
which is simply the DUT layout without the DUT (Fig. 35d) 3 . It is recommended that a
dummy device be included with each group of DUTs. In this way the wafer variations
affecting the parasitic capacitances, such as oxide thickness variations, will also be
corrected.
When laying out passive-characterization devices for probing, the above discussion
93
applies. The simplest layout is shown in Fig. 36a, using one probe. With this layout,
however, part of the DUT parasitic capacitance to ground will not be measured (Fig. 36b).
The solution is to lay out the DUT so that a two-port measurement can be made (Fig.
36c).
Figure 36 Layout examples of passive characterization devices
One final tip. To obtain maximum repeatability, use a dielectric etch to mark where the
front of the probe tips should land (Fig. 37).
Figure 37 Dielectric vias for probe alignment marks, used to increase measurement repeatability
Hybrids and Packages

The typical pad and interconnect dimensions in packages and hybrid substrates are much
larger than with ICs. The larger pad and interconnect dimensions require short-length DUT
connections, especially the ground connections, for good GHz-probing. The probe contacts
do not need to land in the center of pads, but can land near the edge (Fig. 38). Be sure
that the passivation has been removed from where the probes will land. Also, refer to the
Cascade Microtech Probe Head Selection Guide to verify that probes with a wide enough
pitch are available.
Figure 38 Off-center probe, landing on large hybrid or package pads. Probing into wells has a potential problem (Fig.
39).
The probe can collide with the top corner of the well. If you have a potential situation like
this, please call the Cascade Microtech Applications group and request a side view of the
probe you plan to use. Generally the WPH and RTP line of probes approach the DUT at
about an 11 degree angle.
Figure 39 Side view of probe and package with a well, illustrating potential clearance problem
94
Summary
This Application Note is intended to help you lay out your IC hybrid, or package so that
Cascade Microtech coplanar probes can be used for testing. These probes are widely used
to test individual devices, MMICs digital ICs hybrids, and packages. If you have any
questions
pertaining to layout or testing please call the Cascade Microtech Applications Group.
95
DC
Contents
DC Measurements and Calibration Techniques (iccapmhb)

DC Characterization for Device Modeling (iccapmhb)
Power DC Measurements (iccapmhb)
Process Control Monitoring (PCM) Measurements (iccapmhb)
DC Measurements and Calibration Techniques
Large signal modeling of a nonlinear component always begins with the characterization of
its DC performance. Instead of power supplies, DC parametric analyzers with source-
monitor-unit plugins (SMU) are used. This allows to fully characterize the DUT (device
under test) from femto-Ampere up to its maximum current, and in all four quadrants. I.e.
forward and reverse currents and voltages, are measured with the same SMU unit.
To begin with, there are 3 basic sweep types available for the SMUs of DC analyzers, as
depicted below:
SMUs apply Kelvin measurements to avoid parasitic series resistances. This measurement
a second one in parallel (Sense). Figure 1 illustrates this. Ohmic losses on the Force line
are eliminated by the OpAmp in voltage follower mode. This means the OpAmp output will
exhibit a somewhat higher voltage than the desired test voltage at the DUT, because the
test current generates some ohmic losses along the Force line. The Sense line, connected
to the minus input of the OpAmp, assures that the DUT is exactly the desired test voltage.
detailed schematic:
Fig.1: The principle of Kelvin measurements

While this method eliminates DC errors, it does not cover dynamic measurement
problems. For example, to avoid external electro-magnetic influences, both the Force and
Sense cables consist of shielded cables. Such cables exhibit some parasitic capacitances.
Due to charging problems, these capacitances will affect the measurement speed and
accuracy of our Kelvin measurement.
A simple example: assume we want to measure the reverse characteristics of a
steps to, for example, -20 V, the quiescent voltage at the diode is zero. That is, the cable
capacitors are not charged. When the negative voltage step occurs, these capacitances
have to be charged, and the required current is provided by the OpAmp. This could lead to
either a mis-measurement (DUT current plus charging current) or a delay in the triggering
of the actual current measurement (by some intelligent firmware in our measurement).
96
separate, second OpAmp which follows exactly the value of the desired test voltage. Now
it is this auxiliary OpAmp which supplies the charging current for the test cables, while the
main OpAmp can start current measurements independently of this charging problem.
That is, the inner measurement loop does not see the charging problem any more.
never be connected to Force or Sense. Otherwise, the inner loop OpAmp of the SMU would
measure the DUT current plus the charging current of the auxiliary, second OpAmp!
possible offset voltages and currents and corrects it. This type of calibration does not need
any action from the user.
Hints on DC measurements
For transistor measurements: If an individual SMU (source monitor unit) is employed for
each individual pin of the DUT, and the 'Ground Unit' of the 4142 or 415x is not used at
all, it is possible to perform all the forward and reverse measurements without
reconnecting the DUT. In this case, it is recommended to configure the 4142 with twice a
1Ampere SMU (for Collector-Emitter , Drain-Source correspondingly) and to employ two
100mA SMUs for the other transistor pins.
Connecting triax to another connector type

The following table gives an overview about connector adapters from Agilent. For
converting a triax SMU output to a coax connection, for example to connect a SMU to the
coax input of a NWA S-parameter testset, use the Agilent 1250-2652 triax-coax
converter. This connector leaves the middle shield open, what is a prerequisite for the
shielded Kelvin measurement principle.
NOTE: A very good overview about how to connect the Agilent DC analyzers for on-wafer
measurements can be found in the 4155B/4156B Product Note-3, Feb.98, Agilent Lit.no.
5966-4185E. See the Semiconductor Test Assistant CD, available from Agilent, dept.
Semiconductor Test.
Another vendor of Adapters and Feedthrus is Trompeter Electronics, on the web at:
http://www.trompeter.com/shop/home.asp
Note: If you want to download the Trompeter Adapter Catalog, select the catalog T20
Military/Aerospace:
The figure below shows the available converters. Again, for connecting the triax cable end
to the coax input of a network analyzer testset, we need a triax-female to coax-male
converter:the ADBJ77-E2-PL20.
If you need to access the force/sense middle shielding contact, e.g. to extend the
shielding to the shielding of the device on wafer, use the connector plate of the
Agilent4142, opt.012, Agilent part number P/N 04142-60021
Monitoring the SMU voltages

Applying the measurement setup given in fig.2, an oscilloscope can be connected to
monitor the voltage at the transistor pins. This is done by using the sense line as a high
impedance scope probe. This allows to measure the voltage versus time accurately.
97
Fig.2: monitoring the SMU voltage at the transistor pin.
Expanding Current to >1A

If used in combination with an external DC power supply, the Agilent semiconductor
parameter analyzers can evaluate the DC characteristics of a power device that requires a
current of several amperes. Application note 4156-5 describes how to connect the cost-
effective Agilent 665x DC power supply in series to the DC analyzer SMU output.
Therefore, no extra HPIB programming is required. These power supplies are fast enough
to follow the SMUs output voltages. Furthermore, an internal small shunt resistor is used
to monitor the amplified current. This shunt resistor can be measured once by IC-CAP,
and the voltage drop across it serves then as current monitor.
Recommended external power supplies:- 50 A / 8 V (Agilent 6651A)- 25 A / 20 V (Agilent
6652A)- 10 A / 50 V (Agilent 6654A- J05)
See the docu in the Modeling Handbook at
ICCAP_ModelingBook_pdf\1MEASUREMENTS\1DC\DC_Power\Publications\Agilent\415x_plus_PowerSupplies\2004_AN4156_5_external_PowerSupply.pdf
_____________________________________________________
On-Wafer DC Measurements
The following slide of Cascade Microtech explains recommended grounding and guarding
Thermal Effects During DC Measurements

As a general rule, we have to account for self-heating effects during DC measurements of
packaged devices already at currents above ~10mA (for typical Vmax = 5V). For on-chip
measurements using a thermochuck, the critical current is about 50mA. This can be seen
in fig.1. Besides possible self-damage of the DUT, self-heating is a problem with modeling.
This is because the simulation model assumes a constant temperature TNOM, at which the
simulation is performed. In reality, however, the DUT suffered from self-heating and
exhibited therefore different measurement curves than those curves obtained from the
simulator. Therefore, a curve fitting with physical parameter values can not be achieved.
Fig.1: chip temperature increase of a packaged transistor as a function of the pulse width
and the applied bias power. The pulse period is 1s.
To prevent from self-heating, pulsed DC measurements could be applied. This gives

isothermal measurement conditions, and would allow to extract really those model
parameters which are correct for the constant modeling temperature TNOM. However, if
these pulses exceed a few milliseconds, self-heating might occur again, see fig.2.
98
Fig.2: a self heating effect may occur even for pulsed bias conditions
Fig. 3 gives an example for a bipolar transistor, measured with different pulse widths. It
can be seen that only pulses in the millisecond range can lead to isothermal measurement
results. For bipolar silicon transistors, the Collector current increases with increasing pulse
width. For GaAs transistors, it's the opposite.
Bipolar transistor:
GaAs transistor:
Fig.3: Thermal self-heating for pulsed bias measurements as a function of the pulse width
This is the reason that the minimum pulse with of the Agilent 414x and 415x DC analyzers
are not short enough for isothermal modeling measurements. The main application for
that type of pulsed measurements is for the other big market of those instruments, the
production measurements. For this kind of application, pulsing is used for some specific
spot measurements at high currents which otherwise could destroy the component. The
isothermal issue is not of interest hereby.
How to check for isothermal measurement conditions:

Let's take a bipolar transistor for example. We measure the so-called Gummel-Poon plot
and calculate BETA = ic / ib. The Collector voltage is kept at vCE = MAX(vCE / 2). We
then perform these measurements: > sweep vBE from low to high voltage, long
integration mode > sweep vBE from high to low voltage, long integration modeIf we
realize a different BETA curve, there is thermal self-heating.
What to do if pulsed measurements are not possible:

If pulsed DC measurements are not available (Agilent pulsed measurement system
85124A, allowing 1ms pulses, see the chapter on network analyzers), we have these
possibilities:
DC:Characterize the device in those DC ranges where it will be used later in the design
application. In many cases, this is not in the high current range.If the application is in the
self-heating range, and the model doesn't include self-heating, it might be smart to
measure using long integration range, in order to really heat up the device like in the later
application, and to fit the model to the non-isothermal measurements.If your selected
99
model includes self-heating, then see the chapter on thermal modeling for more details.
AC:If S-parameters are to be fitted too, it is recommendable to measure the DC curves
under the same conditions like the biased S-parameters. Since the S-parameter
measurements are usually much slower than the DC measurements, we have more self-
heating for AC than for DC. Therefore, long integration should be applied to the DC
measurements.
Some Remarks on Staircase Sweeps
It is important to note that DC analyzers, like other instruments, have two operation
modes: - either the controlling software sends the start-stop-step stimulus info and the
'what-to-measure' GPIB commands to the instrument and finally sends a trigger command
('User Sweep' = no), - or every individual measurement point is triggered individually by
the software ('User Sweep' = yes).
>> As a consequence, the measurement vs. time is affected from these two possible
settings.
In the slide above, for 'User Sweep'=no, the CPU of the instrument is the owner, or
controller of the measurement, and not the controlling software (like IC-CAP). In case of a
DC analyzer, the CPU's firmware itself will control the measurements during the staircase
sweep. After having received the 'trigger' command, it will wait for the measurement
result to converge (end of charging effects), it will set the ranging of the amplifier (to
have the max. resolution of the AD-converter) etc. This means that the time stepping of
the staircase sweep is not monotonic ( ! ). However, the measurement result (in e.g. IC-
CAP) looks like a straight line, since we plot e.g. current vs. voltage, and ignore the time
stepping. This explains some very special effects where e.g. the transfer curve current is
different by some % from the output characteristics current, for identical bias voltages,
and where self-heating can be excluded to explain the differences.
If the user wants to make sure that the staircase sweep stepping represents a more or
less linear stepping in time, he should set the 'User Sweep' = yes, see the next slide.
However, due to the huge activity on the GPIB bus, the measurement may take much
longer in this case.
When setting 'User Sweep' = yes, the controlling software (e.g. IC-CAP) is the 'owner of
the sweep'. Due to the relatively slow performance of the GPIB bus compared to the
instrument's measurement time, the resulting staircase sweep is time-wise pretty
monotonic. Apply this instrument mode if you must assure such time dependence of your
measurement.
NOTE: in 95% of modeling applications, 'User Sweep' = no is appropriate.
Acknowledgements:
special thanks to Fridolin Illien of ETH University Zuerich, Switzerland, for the oscilloscope
screen shots !
100
Some Remarks on Oscillations During DC Measurements
Active components like transistors may self-oscillate during DC measurements. Typically,

the self-oscillation condition is fulfilled happens only for certain bias conditions. the
following two slides show such phenomena of a bipolar transistor. The first slide depicts a
linearization effect of iC and iB at vBE >0.85V, caused by self-oscillation. The other two
slides show typical self-oscillation at certain bias levels.
All 3 slides from: J.Berkner, Kompaktmodelle fr Bipolartransistoren, Expert-Verlag

Renningen (Germany),
ISBN 3-8169-2085-3, Februar 2002
Under certain circumstances, SMUs may oscillate. Superbeta, wide- bandwidth bipolar
101
1) SMU Induced oscillationsThese oscillations occur when an inductive load is connected to
the SMU's output. The SMU has often an inductive load because the output impedance of
the SMU is inductive during V-Mode operation and usually several SMUs are indirectly
connected together through the DUT.
2) Oscillation due to straysThe measurement system including the DUT, stray capacitance
and residual inductance of the connection cables, switching matrix, probe card and/or test
fixture can be recognized as an oscillation circuit.
understandable when thinking of the SMU output as a low-pass filter, while the oscillation
frequency may be in the 100MHz ... to several GHz range. Also, if like in most cases the
oscillation is located at the DUT, the SMU itself cannot do anything to prevent this
oscillations, since the cables are long compared to the oscillation wavelength.
Some effective methods to eliminate oscillations.

> For FETs, add resistive ferrite beads (Agilent Part Number 9170-0029) as close as
possible to the gate.
> For bipolar transistors, add resistive ferrite beads as close as possible to the base or
emitter.
> In some cases, it is necessary to use more than one ferrite beads.
> Keep cables as short as possible. Long cables cause oscillation because of their large
inductance.
>>> Or, apply shielded DC probes, even better apply RF probes.
Publications
Agilent 4155B/4156B Product Note No. 3: Prober Connection Guide, Agilent Technologies
Literature Number 5966-4185E, 1998
Ultra Low Current dc Characterization of MOSFETs at the Wafer Level,Agilent Technologies
Application Note 4156-1, 1/1998, Lit.Nr. 5963-2014E
102
DC Characterization for Device
Modeling

performance.
Instead of power supplies, precision DC parametric analyzers with source-monitor-unit
(SMU) plugins are applied. They offer current ranges from femto-Ampre (1E-15) to
several Ampres, and voltage potentials from micro-Volt to hundreds of Volts. This allows
to fully characterize the DUT (device under test) in all four I-V quadrants. I.e. forward and
reverse currents and voltages, are measured with the same SMU unit.
individual SMUs in order to avoid re cabling during the forward and reverse
measurements.
When performing these measurements, care must be taken for accurate shielding,
thermal selfheating, contact resistance etc.
4-quadrant DC analyzers are applied

Several measurement plug-ins (SMU Source Monitor Unit)
Measurement ranges: 10 fA ... 1A, 2 mV .. 200V
Kelvin measurement capabilities
Automatic self-calibration

performance. Instead of power supplies, precision DC parametric analyzers with source-
monitor-unit (SMU) plugins are applied. They offer current ranges from femto-Ampre
(1E-15) to several Ampres, and voltage potentials from micro-Volt to hundreds of Volts.
This allows to fully characterize the DUT (device under test) in all four I-V quadrants. I.e.
forward and reverse currents and voltages, are measured with the same SMU unit.
When performing these measurements, care must be taken for accurate shielding,
thermal selfheating, contact resistance etc.
SMUs apply a Kelvin measurement to avoid parasitic series resistances. This measurement
a second one in parallel (Sense) for every pin of the DUT. The slide above illustrates this.
Ohmic losses on the Force line are eliminated by the operational amplifier (OpAmp1) in
voltage follower mode. This means this OpAmp1 output will exhibit a somewhat higher
voltage than the desired test voltage at the DUT, because the test current generates some
ohmic losses along the Force line. The Sense line, connected to the minus input of the
OpAmp1, assures that the DUT is biased with exactly the desired test voltage.
103
Note
when you only connect the Force cable and leave the Sense open, the internal 10kOhm resistor provides
the required closing-of-the-loop for the OpAmp.
While the Kelvin method compensates the DC errors, it does not cover dynamic DC
measurement problems. For example, to avoid external electro-magnetic influences, both
the Force and Sense cables are shielded. But such cable shieldings exhibit parasitic
capacitance. Due to charging problems, these capacitances will affect the measurement
speed and accuracy of our Kelvin measurement.
As a simple example: assume we want to measure the reverse characteristics of a
steps to e.g. -20 V, the quiescent voltage at the diode is zero. That is, the cable capacitors
are not charged. When the negative voltage step occurs, these capacitance have to be
charged, and the required current is provided by the OpAmp1. This could lead to either a
mis-measurement (DUT current plus charging current) or a delay in the triggering of the
actual current measurement (by some intelligent firmware in our measurement).
separate, second OpAmp2 which follows exactly the value of the desired test voltage. Now
it is this auxiliary OpAmp2 which supplies the charging current for the test cables, while
the main OpAmp1 can start current measurements without being affected by this charging
problem. That is, the inner measurement loop does not see the charging problem any
more.
never be connected to Force or Sense. Otherwise, the inner loop OpAmp1 of the SMU
would measure the DUT current plus the charging current of the auxiliary, second
OpAmp2!
Calibration:
possible offset voltages and currents and corrects it. This type of calibration does not
require any action from the user.
Typical on-wafer IV measurement setups require probes and chuck bias. A careful
shielding completely surrounds the wafer with guard.
Make sure that the Guard contact of the SMUs is left open. If it was connected either to
the inner 'hot' measurement cable layer or the outer ground layer, this would result in
wrong current measurements!!
(slide with courtesy from Cascade Microtech)
Contacting a Wafer using ForceSense DC Needles

104
When you can not apply Force-Sense needles, you need to measure the contact
resistance.
In this case, enclose your model in a subcircuit* and add the contact resistances there.
Shielding problem:50Hz noise overlaying the measurement
50Hz noise shows up like a regular pattern, distorting the measurement result.
Uncorrelated noise does not exhibit a regular pattern.
Note
When your noise looks like the pine (fir) tree branch shown in the photograph, it is uncorrelated noise.
Warming-up during DC biasing
105
When making DC measurements, self-heating has to be taken into account. Even when
applying pulsed DC signals.
And the selfheating of the previous impulse can overlay the selfheating of the following
impulse, like sketched above.
Self-heating during DC measurements
Self-heating can become a severe problem with device characterization, because, when
self-heating occurs, the measurement results depends on the measurement speed!
This can be verified when e.g. measuring a Gummel plot for a bipolar transistor, once
sweeping from low to high voltages different measurement speeds, and then sweeping
from high to low voltages. Calculate beta out of your different measurements.If self-
heating occured, you will get as many beta curves as you have performed measurements.
The plot in the slide above gives the chip temperature increase of a packaged transistor as
a function of the pulse width and the applied bias power. The pulse period is 1s. As can be
seen, self-heating can only be avoided when applying very short, pulsed measurements,
below 1us pulse width. Such pulsed measurement systems, also included pulsed S-
parameter measurements, are commercially available.
However, such systems are quite complex and expensive. Therefore, if you have to live
with self-heating, make sure your device suffers always from the same self-heating. This
means in general, apply the slowest measurement speed for you're your DC
measurements, since your biased network measurements will be slow as well, and self-
heating will definitively occur there.
What to do when self-heating becomes an issue for your device modeling:
A min. pulse width of ~1us or below (see figure above). This means the DC analyzers
in pulse mode cannot be used ( pulses are 100us or longer !)
Use the IC-CAP pulsed modeling system
Live with self-heating, but make sure your data are consistent:
all DC measurements are performed with the same self-heating (slow
measurements!)
same self-heating for biased S-parameter measurements as for the
DCmeasurements.
Self-heating with DC measurements
106
As a general observation, currents above ~25mA will cause self-heating effects with
transistors. Rule of thumb: 25mA with typically 2V equals already 50mW for that tiny
transistor on the wafer!
Self-heating can only be avoided when applying pulsed DC measurements. However, the
max. duration for such a DC pulse is 1us, followed by typically 1ms wait time.
If you have to live with self-heating, make sure to have repeatable measurements for the
DC Settings (currents for the transfer and for the output characteristics must be identical
for identical bias conditions !!). Furthermore, the self-heating during the (fast) DC
measurements must be the same as with the (slow) S-parameter measurements!!
Output Characteristics for a Single vg, at Different Chip Temperatures
Note
for a clear picture, the isothermal data (-10'C to 150'C) above come from simulations.
Worst Case: Thermal Runaway due to a weak heat sink
Such a thermal runaway is due to a too small heat sink.

A bigger practical problem, however, is the self-heating before the runaway happens !
Making thermally repeatable measurements in this bias conditions requires extremely slow
measurements with long HOLD and DELAY settings.
Dealing with Self-Heating
107
It is important to note that DC analyzers, like other instruments, have two operation
modes: - either the controlling software sends the start-stop-step stimulus info and the
'what-to-measure' GPIB commands to the instrument and finally sends a trigger command
('User Sweep' = no), - or every individual measurement point is triggered individually by
the software ('User Sweep' = yes).
As a consequence, the measurement vs. time is affected from these two possible settings.
In the slide above, for 'User Sweep'=no, the CPU of the instrument is the owner, or
controller of the measurement, and not the controlling software (like IC-CAP). In case of a
DC analyzer, the CPU's firmware itself will control the measurements during the staircase
sweep. After having received the 'trigger' command, it will wait for the measurement
result to converge (end of charging effects), it will set the ranging of the amplifier (to
have the max. resolution of the AD-converter) etc. This means that the time stepping of
the staircase sweep is not monotonic. However, the measurement result (in e.g. IC-CAP)
looks like a straight line, since we plot e.g. current vs. voltage, and ignore the time
stepping. This explains some very special effects where e.g. the transfer curve current is
different by some % from the output characteristics current, for identical bias voltages,
and where self-heating can be excluded to explain the differences.
If the user wants to make sure that the staircase sweep stepping represents a more or
less linear stepping in time, he should set the 'User Sweep' = yes, see the next slide.
However, due to the huge activity on the GPIB bus, the measurement may take much
longer in this case.
When setting 'User Sweep' = yes, the controlling software (e.g. IC-CAP) is the 'owner of
the sweep'. Due to the relatively slow performance of the GPIB bus compared to the
instrument's measurement time, the resulting staircase sweep is time-wise pretty
monotonic. Apply this instrument mode if you must assure such time dependence of your
measurement.
Note
in 95% of modeling applications, 'User Sweep' = no is appropriate.
'Use User Sweep' can be set in Instrument Options
Some Quick Remarks on Trigger Statements
Hold Time
Time to allow for DC settling before starting internal or user sweep.
Delay Time
Time the instrument waits before taking a measurement at each step of an internal or
user sweep.
Integ Time
Sets the integration mode for A/D converter (number of samples. It can be set to Long,
Medium or Short.
108
For CW measurements of Power Devices, make sure that at each measurement point, the
final temperature rise has settled.
This is achieved by setting the DELAY time to an appropriate value, e.g. 1 minute.
For output characteristics, when the Drain voltage is re-swept at every incrementation of
Vg, make sure to set the HOLD time to e.g. 20 minutes.
This will allow the device to cool down from the last vDmax bias condition to the new
vD=0V condition at the next vG bias point !
Thermal Effects - DC data consistency

In the plot, you see a conversion of a measured transfer characteristics into an output
characteristic and comparing it with the measured output characteristic itself.
If a conversion of a transfer characteristic into a pseudo output measurement (or the

other way round) exhibits exactly the same currents than the output characteristics
measurement itself, the thermal conditions were identical. In the example above, this is
not the case !
Output Characteristics Data Inconsistency for Power Devices
In the plot, you can see two output characteristics with different vdmax overlaid. The one
with the bigger vdmax represents a more self-heated transistor than the other. To avoid
this problem, set the Hold Time to a big value (minutes).
When two output characteristics have to be measured for the same device, due to power
restrictions, make sure to have the same self-heating in both measurements.
A new measurement approach for power devices"up_down

ramping" - faster measurements by making use of the cooling-down
phase
Note
This sweep scheme is 2 times faster than the conventional one!
Especially for packaged devices, the cooling down after reaching the max. of the 1st order
sweep can take up to several 10 minutes.
109
When applying an up-down-ramping bias, the extremely long Hold time of a conventional
back-to-zero 1st-order-sweep can be reduced when the 2nd order sweep steps to the next
higher value and the 1st order sweeps just steps down.
An example of an instrument driver for such a sweep is in

demo_features\2_meas_instruments\2user_def_meas_instr_drv\N67xx_100W_SMU_driver_by_PEL.mdl
and then Setup 'Drive_67xx_updownramping_OutputCharact'.
Oscillations during DC measurements
Under certain circumstances, SMUs may oscillate. Superbeta, wide- bandwidth bipolar
1. SMU Induced oscillationsThese oscillations occur when an inductive load is connected

to the SMU's output. The SMU has often an inductive load because the output
impedance of the SMU is inductive during V-Mode operation and usually several SMUs
are indirectly connected together through the DUT.
2. Oscillation due to straysThe measurement system including the DUT, stray
capacitance and residual inductance of the connection cables, switching matrix, probe
card and/or test fixture can be recognized as an oscillation circuit.
understandable when thinking of the SMU output as a low-pass filter, while the oscillation
frequency may be in the 100MHz ... to several GHz range. Also, if like in most cases the
oscillation is located at the DUT, the SMU itself cannot do anything to prevent this
oscillations, since the cables are long compared to the oscillation wavelength.
Self-oscillation with an LDMOS output characteristics
110
Self-Oscillation with a bipolar transistor
Note
The slide is from the book:
J.Berkner, Kompaktmodelle fr Bipolartransistoren, Expert-Verlag Renningen (Germany),
ISBN 3-8169-2085-3, February 2002
How to avoid oscillations

Keep unshielded part of DC contacts as short as possible(Use shielded DC probes)
Use high quality cable (low cable inductance)
Use GSG (Ground-Signal-Ground) RF probes.
Note
Applying ferrite beads is not recommended for device modeling measurements since that trick does not
work for S-parameter measurements.
Effect of cable movement
Monitoring the SMUs by an oscilloscope

Using the Sense Line as a High Impedance Scope Probe
Besides for Kelvin measurements, the sense line can be used also for monitoring the SMU
voltage on the device with an oscilloscope. The sense line tracks the force line within 1mV.
All you need is a floating guard triax->coax adapter attached to the sense line at the back
of the 4156. Then use any BNC cable to direct connect the SMU sense line to the
oscilloscope input. The adapter shown is the Trompeter Electronics AD-BJ20-E2-PL75.
Connecting triax SMU outputs to coax inputs

See Connecting Triax SMU Outputs to Coax Inputs (iccapmhb)
The Effect of Big Capacitive Loads

111
The Delay Time must be big enough to avoid measuring non-static charging effects. Big
capacitive loads can lead to the case that the SMU begins triggering the measurement,
while the charging has not yet finished.
See the DC Analyzer manuals for the max. applicable capacitive load. For example, the
E5270A specifications are:
Maximum capacitive load:

For 1 nA to 10 nA ranges:1000 pF
For 100 nA to 10 mA ranges:10 nF
For 100 mA to 1 A ranges:100 F
Maximum guard capacitance:900 pF
Maximum shield capacitance:5000 pF
SMUs and Capacitive loads - The Effect of Long

Cables
a 10MW resistor was measured with different capacitors in parallel. Capacitors exceeding
~1nF affect the measurement result due to a too long charging (settling) time, during
which the SMUs already start to measure.
Note
The grey curve, i.e. the 10nF case "with extended SMU delay time (100ms)" brings the faulty 10nF trace
"without delay time" (purple) back to the expected measurement trace. For 50nF, delay time extension
did not help any more.
Triax and Quadrax Cables exhibit a capacitance between the signal and Guard. This
capacitance should be kept below ~1nF.
Note
when using twice a single-triax cable for a Kelvin measurement, watch out that the capacitance of the
twice a triax cable is doubled. If the DUT capacitance is high, Quadrax cables may therefore be a smarter
solution.
112
Model Length Capacitance
16494A Triaxial Cable
Option 001 1.5 m 125 pF
Option 002 3.0 m 240 pF
Option 003 80 cm 75 pF
16494B Kelvin Triaxial (Quadrax) Cable
Option 001 1.5 m 140 pF
Option 002 3.0 m 260 pF
Option 003 80 cm 90 pF
16494C Kelvin Triaxial (Quadrax) Cable for 4142
Option 001 1.5 m 140 pF
Option 002 3.0 m 260 pF
Note
How to Select Options for Cable Length?
The Agilent 4155C/4156C or Agilent 4142B have low current forcing and measurement capability by their
SMU (Source Monitor Unit)technology.The SMU uses a kind of feedback circuit for maintaining very high
stability of the current/voltage sourcing. Therefore,the guard capacitance (a load for SMUs)of connections
to SMU must be within certain limits to prevent SMU instability.
The limit is about 900 pF in the Agilent 4155C/4156C or the Agilent 4142B 's case, for example.
You should select proper cable lengths to keep small guard capacitance for the SMUs.A system using
switching matrices sometimes needs longer cable lengths than for the standalone instruments. If you plan
to connect more than ~5 m total cable length (both instrument to switching matrix and switching matrix
to the DUTs), make sure the cable guard capacitance does not exceed 900 pF.
The measurement result from above (i_vs_v) converted to the resistance value R = dv /
di
In the slide above, the 10 MOhm resistor was stimulated with a swept voltage, the current
was measured and the resistance was calculated by R = dv/di.
For capacitive loads above ~1nF, the calculated resistor value is distorted, a hint for
resolution problems with the SMU.
Do not forget to evaluate your DC contact resistance and account for it when extracting
your model DC resistor parameter values !
This is especially true when measuring Silicon devices on the wafer.
The contact resistance for probe contacts on silicon wafers ranges typically between 1-5
Ohm and can be bigger than the transistor resistances.
In IC-CAP, you can add a Test Circuit on the DUT level to account for the DC losses,
without affecting the extracted model (which is specified in the Circuit tab).
For an example, load
demo_features\3_MEAS_ORGANIZE_n_VERIFY_DATA\0_MASTER_FILES\DIODE_MEAS_MASTERFILE_demodata_PELdep.mdl
and see the TestCircuit in DUT 'DC'. See also how this contact resistance is measured, in
DUT 'verify_DC_probe_contact'.
Conclusion
When making DC measurements for device modeling,take into account:- shielding
113
technologies
self heating- contact resistance

self-oscillation challenge
capacitive load effect
Miscellaneous Information
The IC-CAP SystemVariable MEASURE_FAST
When defined as Yes, IC-CAP will attempt to minimize instrument re-initialization during
repeated measurements on the same Setup. See Speeding Up Repetitive Measurements in
the IC-CAP Manual, chapter on Measurement. Default is No.
MEASURE_FAST will invoke an instrument initialization for the first measurement of a

Setup. I.e. subsequent measurements of the same Setup will be faster, because non
instrument initialization is done any more.
In order to suppress this initialization between Setups, you also have to set
MDS_MEASURE_FAST.
Note
If you set MDS_MEASURE_FAST=1 at the top level, then even your very first measurement ever will not
call an instrument initialization.
SMU turn-on and turn-off sequence
Slide 1
Slide 2
For CW measurements of Power Devices, make sure that at each measurement point, the
final temperature rise has settled.
This is achieved by setting the DELAY time to an appropriate value, e.g. 1 minute.
For output characteristics, when the Drain voltage is re-swept at every incrementation of
Vg, make sure to set the HOLD time to e.g. 20 minutes.
This will allow the device to cool down from the last vDmax bias condition to the new
vD=0V condition at the next vG bias point !
Slide 3
Such a thermal runaway is due to a too small heat sink.

A bigger practical problem, however, is the self-heating before the runaway happens !
Making thermally repeatable measurements in this bias conditions requires extremely slow
measurements with long HOLD and DELAY settings.
Slide 4
114
This is the case of a NPN power transistor mounted on a big heatsink:

In this example, we compare the 0 -> 10V output characteristics from a 0 -> 10V vCE
sweep and a 0 -> 20V vCE sweep, in CW mode.
As can be seen, in the case of the vCE -> 20V sweep, when the vBE bounces back and is
incremented for the next branch of the output characteristics, the transistor is still hot
from the vCE=20V case and cools down. The iC-traces drop down at low vCE.
Slide 5
This is once again the case of a NPN power transistor mounted on a big heatsink:
In this example, we compare the 0 -> 10V output characteristics from a 0 -> 10V vCE
sweep and a 0 -> 20V vCE sweep,
pulsed (80ms on, 40ms off) and CW.
Slide 6
Especially for packaged devices, the cooling down after reaching the max. of the 1st order
sweep can take up to several 10 minutes.
When applying an up-down-ramping bias, the extremely long Hold time of a conventional
back-to-zero 1st-order-sweep can be reduced when the 2nd order sweep steps to the next
higher value and the 1st order sweeps just steps down.
Slide 7
This is the best result obtained for the NPN 2N3055 power transistor in CW mode, but it is
obvious that the applied delay time (120s) and the hold time of 15s was still at the lower
limit.
Slide 8
But also up-down ramping does not prevent from faulty measurements. Make sure to
have set the hold and delay times sufficiently high for consistent measurements !
Process Control Monitoring (PCM) Measurements
versus Modeling Measurements
There is occasionally some confusionbetween process engineers and modeling
115
engineersabout the term 'parameters'.
As an example, process engineers may define the threshold voltage of a MOS transistor as
the gate voltage that is required to generate a certain Drain current of 1uA, while
modeling engineers relate it to the formulation of the selected MOS model. Often, yet not
quite model-specific, they define the threshold voltage to be the x-intersect of a line fitted
to the steepest slope of the transfer curve Id(Vg).
Therefore, it is obvious that these two parameter values are different. However, both
engineering groups refer to it as VTH. With bipolar transistors, it is the Early voltage VAF,
that process engineers define as the x-intersect of a tangent to the output characteristics,
while modeling engineers consider VAF together with other modeling parameters to
achieve a fit of the simulated curves to the measured ones.
It is in the statistical analysis that both worlds come together. Using factor or principal
components analysis, we attempt to explore statistical dependencies from a huge amount
of data (process and modeling parameters) together. It can be shown that if the model
parameters have been carefully determined (without using much optimization, as stated
previously), they have a direct relationship to the process parameters. This means that
after such a statistical analysis of process and modeling parameters (of both PMOS and
NMOS transistors for example), the model parameters can be related to the fewer process
parameters. These are measured continuously in production. This implies a considerable
reduction of modeling times.
Another means of distinguishing between the process and modeling parameters is the
method applied to measure these parameters. In production, spot measurements are
usually performed, that is, individual measurement points. The process parameters are
then calculated from these spots. These kinds of measurements are also called PCM
measurements (process control monitoring) or Parametric Test measurements. Contrary
to that, it is always curve sweeps that are required to do component modeling. The
measurement curves and the simulation curves should be made identical.
The following tables are intended to further clarify this difference between both parameter
worlds.
Parametric Test (spot measurements):{*}Determination of parameters to describe

and monitor a semiconductor production process measurement type: parameters to
describe the process current-voltage (IV):gain, threshold voltages, delta_w, delta_L
etc.capacitance-voltage (CV):doping concentrations and profiles, oxide and metal
impurities,film thickness, carrier lifetimes etc.
Modeling (sweep measurements):{*}Determination of modeling parameters to
achieve a fit of mathematical curves to measured datameasurement type:
parameters to fit the model curves to the measured curves MOS bipolar GaAscurrent-
voltage (IV):VTH, U0, KP...IS, NF, BF...IS, NF, VTO, BETA...capacitance-voltage
(CV):CJ, MJ, CGDO...CJO, MJ, VJ...CGSO, CGDO...S-parameters:ELM...TF, ITF,
VTF...TAU, A5...
!
Figure: Process Control Measurements during chip manufacturing
116
CV
Contents
CV Characterization (iccapmhb)
CV Measurement And Calibration Techniques (iccapmhb)
117
CV Characterization
After a DC characterization, modeling engineers usually perform a so-called CV

(capacitance versus voltage) measurement in order to characterize the device
capacitances at a standard frequency of 1MHz. This frequency is high enough to allow a
resolution down to a few femto-Ampere (provided shielded probes are applied for e.g. on-
wafer measurements), yet still low enough to neglect second order parasitic like resistors
in series with the capacitors, or like inductances.
For such CV measurements, the DC-bias is swept, a test frequency (1MHz) is applied to
the DUT, and the instrument calculates the capacitance between the 2 pins of the DUT
from the magnitude and phase of the device voltage and current. This means, an
impedance meter interprete the measurement result always with respect to a user-
specified schematic: either a capacitor in series with a resistors, or both in parallel. This
explains, why capacitances and resistor values may vary with frequency when measured
with such a device. In other words, these frequency-variations are due to a too simplistic
analysis model behind the measurement. A better way is therefore to measure the
capacitance with network analyzers. In this case, it is up to the user to interpret the
measurement result (S-parameters).
Compared to DC, CV measurements are 4-terminal measurements too. The CV meter

outputs are:* low current and potential
high current and potential And similarly to the DC Kelvin measurement procedure,
both the low and high pin are connected together as closely as possible to the DUT.
In the auto-balancing method (next slide), the shielding of these four wires
correspond to the virtual ground of the instrument's OpAmp, and not the chassis
ground . This eliminates any influences caused by the cables.
CV Measurements - Calibration considerations

Test cables and fixtures contribute and affect the device characterization. For CV
measurements, the calibration consists of unconnecting the DUT, assuming an ideal OPEN
condition and measuring the cables and their OPEN parasitics (CV-Meter calibration). After
that, the corresponding capacitance is automatically subtracted from the DUT
measurement by the CV meter.
Note
If we are interested in the inner DUT's CV curves, i.e. without its surrounding test pads capacitances, we
need to connect to an OPEN dummy structure during CV meter calibration instead of simply leaving the
cables unconnected. Such an OPEN dummy consists of all connection pads, lines to the DUT etc, but
without the inner DUT itself.
How do capacitance meters work?
118
For today's CV meters, the measurement principle is typically a so-called auto-balancing

method. The slide above depicts the simplified measurement scheme.
The DUT is inserted in the feedback loop of an operational amplifier, and the system is
stimulated with typically a 1MHz sinusoidal signal plus a DC bias. The feedback resistor R2
is precisely known, and the complex voltages V1 and V2 are measured. From the formula
given above in the slide, the capacitance of the DUT can be calculated, assuming an
equivalent schematic of either a resistor in series with the capacitor, or, commonly for
modeling, a capacitor in parallel with a resistor (which is the bias-dependent diode
resistance for example)
When applying the Auto-Balancing Bridge, capacitance from High

and Low Terminal to Guard (e.g. coax cable shielding) are
automatically excluded from the measurement result !!!
In the auto-balancing method, the shielding of these four coax cables are connected to the
virtual ground of the instrument's OpAmp. A special feature of this auto-balancing method
is that it is insensitive to stray capacitance towards (virtual) ground. As depicted in fig.3
below, this becomes clear since the parasitic capacitance Cp1 is tied directly to the
supplying bias and stimulus voltages, and the parasitic capacitance Cp2 is located
between virtual ground (minus input of the OpAmp) and ground. This means, both
parasitic capacitance are not included in the measurement!
On-Wafer Measurements
AttoGuard Enhanced CV Measurements
FemtoGuard surrounds the chuck at shield ground

Patented AttoGuard above the chuck at shield ground
Creates a virtual double-shielded Faraday enclosure
10 atto Farad CV measurement resolution
Zero CV meter only one time
On-Wafer CV Measurement: where to connect the Hi and Lo of the

CV meter
Do not connect the instrument's Low terminal to the substrate. The substrate is
electrically connected to the prober's noisy ground (wafer chuck stray capacitance), and
the Low terminal of the auto balancing bridge instrument is sensitive to noise. If the wafer
chuck (stage) of the prober is isolated from the ground and effectively guarded, the
shielding conductor of the 4TP cable can be connected to the prober's guard terminal to
minimize stray capacitance around the probes.
119
CV on-wafer measurements - What to do with the

unused pins?
CV Shielding Techniques - avoiding the measurement of parasitic
capacitance
Connecting only 2 pins of a multi-pin device means that the capacitance between these 2
pins plus any other combination of capacitances between the 2 pins (CGD, CDB, CSB) will
be measured !
Connecting the unused pins to the coax guard excludes the parasitic capacitances from
the measurement. Only CGS is measured.
Due to the auto-balancing CV measurement principle, parasitic capacitances of the DUT

pins (CGS, CDB, CGB) can either be included with the measurement of CGS (above left) or
they can be excluded if the parasitic capacitance pins are connected to the guard
(shielding) of the CV meter coax cables (above right).
Application Example: Measuring CGD of a packaged MOS
With the auto-balancing method, connecting the Source of the MOS to the Guard, the
cable shielding potential, eliminates the effect of CGS and CDS.
Only C GD is measured !!
Note
The shield of the 4 coax outputs of a LCRZ meter is virtual ground, not chassis ground ! Due to the applied
measurement principle (autobalancing bridge), all capacitors between the measurement pins and that
virtual ground are excluded from the measurement result, Only the capacitance between the pins is
measured. However, if in the example above, the Source contact is left open, the capacitances CGS and
CDS will be in series and altogether, will add as a parasitic capacitance to the originally desired
capacitance CDG. In this case, the capacitance values measured with a LCRZ meter appear bigger
compared to being measured with e.g. a network analyzer.
3 Contacting Cases - for CV Calibration and

Measurement, Applying the Auto-Balancing Method
120
In this case, only the pads of the capacitance under test are contacted. For the calibration
of the capacitance meter (and thus the compensation of the cables plus the contact pads
capacitances), the pins are connected to an empty dummy structure first.
As shown, the calibration including the dummy structure reflects a parasitic capacitance of
Cpad/2, while there is a bigger total parasitic capacitance present during the
measurement of the DUT, where also C CS, C BE as well as the pad of the Emitter play a
role !Since only C pad/2 is calibrated out, the DUT measurement represents a too big
capacitance.
In this case, the substrate is additionally connected to CV-meter ground. All other unused
pins are left unconnected.
In this 3rd case, all unconnected pads of the 2-pin measurement, including the chuck, are
connected with additional pins to the shielding of the CV meter coax cables (the Guard,
which, again, represents virtual ground for the CV meter auto-balancing measurement
principle).
With this forced guarding setup, all capacitances from the active pins towards guard are
121
not included in the measurement result. As can be seen in the sketch below, the
measurement is the cleanest method to measure nothing but the desired CBC alone.
Proof related to the previous slides:Measuring a CV curve with

floating Guard
leads to a too big capacitance
Simulations of a Multi-pin Device CV Measurement

When doing modeling, you need to make sure that you simulate what you have measured.
CV simulations for devices with >2 pins simulate what you have
measured or measure what you will simulate
Don't forget about the open pins of your DUT !

A CV simulation will always return the total capacitance between Base and Emitter ,
including CBC and CCE!
Simulating a CV curve with floating unused pins leads to a too big

capacitance
122
Note
How IC-CAP lets the simulator perform CV simulations:
From an inspection of the Simulation Debugger, IC-CAP surrounds your Circuit Netlist by the extra
circuitry given above.
There are 2 cases:
Case 1: leave the unused pins unconnected (no Inputs specified for unconnected pins).
Case 2: connect the unused pins to Ground (additional Inputs specified for the unused pins).
The simulated current value is then interpreted like this:
From the simulated current of the schematic in the slide above, we obtain the complex admittance Y
= iac / v.
Since v=1, the equation is simply Y = iac
Therefore, the capacitance value is calculated within IC-CAP from the imaginary part of the current
after
C = IMAG(Y) / (2*PI*1MHz) (1)
and the conductance is
G = REAL(Y) (2)
IC-CAP applies formula (1) when 'Mode' is set to 'C', and formula (2) when 'Mode' is set to 'G'. IC-CAP
assumes a situation like with the default LCRZ meter setup: a capacitance in parallel with an conductance.
BACK TO THE PROBLEM 'WHAT TO DO WITH THE UNUSED PINS' DURING SIMULATIONS:
It now becomes obvious that in the multipin case, with the unused pins connected to Ground (by
corresponding Inputs), these additional capacitors cannot contribute to the simul. result, since all their
pins are tied to Ground. In the case of the pins left open during simulations (no Inputs specified for them
in the IC-CAP Setup), all capacitances inside the DUT will be stimulated at pin1 and will therefore
contribute to the simulated current at pin2. If you want a simulation at another frequency than the default
1MHz, specify the Model Variable CV_FREQ. If you want a measurement at another frequency, specify the
meas. frequency in the Instrument Options.
Case 1:Leave the unused nodes open during simulation
With the slide above and the next slide, we check the two scenarios. First, we try an IC-
CAP Setup in which only one single Input referring to the two CV simulation nodes is
specified. Also, a single Output is specified, referring to the same nodes as the Input. For
our test, we specify a netlist with 4 nodes, having individual capacitances between all
nodes, each with a different capacitance for easier identification of the result. See the slide
above.
As can be seen in the Plot, the simulation result now includes the very capacitance
between the two pins plus the influence of all other capacitances which are in parallel and
in series-and-parallel to the very capacitance between the investigated nodes!!!
Case 2:Connect the unused nodes to Ground during simulation
In the other scenario, case 2, (see the slide above), the IC-CAP Setup has again an Input
referring to the two CV nodes, and an Output referring to the same nodes as this Input.
However, additional Inputs are specified this time, which tie the unused (open) nodes to
Ground.
In this case, the simulation returns only the capacitance between the two CV Inputs, and
thus, the simulation result corresponds to the recommended measurement of connecting
the unused pins to the virtual ground of the CV meter, i.e. to the shielding of the coax
123
cables.
Considerations regarding the signal level and the CV-

frequency
1. CV measurement signal level
Typical settings: 10mV rms (specified in Instrument Options. You need to verify this
value.
Note
increase the signal level until the CV measurement begins to change ! When this happens, reduce
the signal level.
Like with S-parameters, a CV measurement is a linear measurement, and therefore,

harmonics should be avoided. As a consequence the applied CV signal level should
not be too big, but also not too small (reduce signal-noise ratio). Since there is
usually no DC current flowing at CV measurements, the reason for harmonics is the
nonlinear capacitance itself.
To find out the max. applicable signal level, start with a low signal value and increase
it until you see a change in the measured capacitance value. Select a signal value
with does not (yet) change the measurement result.
2. CV measurement frequency value
Typical settings: 1MHz (specified in Instrument Options). The problem with CV
meters and especially silicon devices is that the silicon may exhibit considerable
losses, which are not represented by the CV meter's simple underlying schematic:
Note
Decrease the frequency until the CV measurement is stable.
A CV (LCRZ) meter interprets the measurement result with respect to a capacitor in

series with a resistor R, or as a capacitor in parallel with a conductor G. If this pre-
requisition is not met by the DUT, the capacitance and/or the resistor or conductor
become frequency dependent. In many cases, especially for silicon substrates, this
can be the case. Therefore, check reducing the frequency from default 1MHz down to
e.g. 100kHz and check if the measured CV curve is affected or not. If it is not or no
longer affected, the mentioned simple underlying schematic is correct, since second-
order effects are switched-off due to the reduced measurement frequency. We now
can perform a correct CV measurement.
Effect of a series resistor with CV measurements
The parallel diode conductance G is responsible for the decrease of the Q factor to
the left of its maximum, whereas the series resistor R is responsible to the right.
3. Maximum applicable DC bias for diode CV measurements
Avoid DC saturation of the instrument !when measuring a CV curve into the On-state
of the diode, don't apply a DC bias bigger than what corresponds to a capacitance ~4
times the 0V capacitance CJ0. When applying a bias voltage to this measurement
method, however, special care most be taken to ensure that the measurement
OpAmp is not saturated by a low resistance state DUT. This can easily happen when
measuring the CV curves of diodes into the ON-state.
Referring to the equation in the slide before, the real (ohmic) part of the resistance
Zx of the DUT, i.e. the semiconductor resistance in parallel to the CV capacitance,
has to be bigger than the reference resistor R. Otherwise, the output of the OpAmp
can reach saturation (e.g. 15V supply voltage). This happens easily when
measuring semiconductor capacitances in forward bias mode. Here, the ohmic part of
Zx tends towards 1W and lower!
A practical approach: when measuring a CV curve into the On-state of the diode,
don't apply a DC bias bigger than what corresponds to a capacitance ~3-5 times the
0V capacitance CJ0.
Obtainable CV measurement resolution
124
Provided that shielded CV probes are applied, a resolution down to Femto-Farad (fF) can
be achieved. The slide above shows the measurement resolution of an OPEN dummy re-
measuring, after calibration.
Conclusion
When making CV measurements for device modeling, take into account what to do
with the unused device pins
check the max. applicable RF signal level
check the max. applicable CV frequency
when measuring conductive devices (diode in on-state), make sure to not measure
higher than ~4*CJO.
wafer substrate connected to the HI pin of CV meter
Miscellaneous Information
On-Wafer capacitance measurements with the auto-balancing
bridge
The coax shieldings should be connected together as close to the DUT as possible
Cable Shielding shorted at Probes
Cable Shieldings shorted at Prober-Input
A more circuit-oriented view inside the CV meter
125
Negative Capacitance Effect Observed for On-Wafer CV

Measurement*
Note
Problems with a R - CG Measurement. The DUT exhibits also a considerable series resistance.
This is the circuit diagram of the previous page.The impedance Zmeas that the LCR meter
measures is expressed by V1 / I2
This is the circuit diagram of the previous page to show circuit analysis. The impedance
Zmeas that the LCR meter measures is expressed by V1/I2.
Impedance Meter using IV Method On-Wafer*

Advanced IV method : Agilent 4294A + 42941A
RF IV method : Agilent E4994A
The advance I-V method is modified from the 4 terminal pair auto-balancing bridge to
perform grounded measurement. The measurement resources, i.e. signal source, volt
meter, and current meter, are positioned at the Gate side. In this case, one terminal of
the on-wafer devices is connected to the substrate of wafer, and therefore, it is also
connected to the wafer chuck. This grounded measurement config. is sometimes
necessary to reduce the effects of wafer chuck
126
Packaged Devices
Measurement
Test fixture ground and unused pinsconnected to CV meter Guard.
Test fixture ground and unused pinsconnected to CV meter Guard. For packaged devices,
we first perform an OPEN CV calibration including the test fixture.
Measuring packaged devices

Measure
Simulate
Apply either 2-pin or fully guarded measurements,but later, simulate exactly what you
have measured.
127
Then, for the measurement, we can either apply the 2-pin measurement (what yields the
total capacitance between the measurement pins), or a guarded measurement yielding
nothing but the desired capacitance between the pins D and S in the example below.
Whatever measurement principle is more adequate in your case, make sure, however,
that you simulate later exactly what you have measured before !!!
References
Agilent Technologies Impedance Measurement Handbook July 2006P/N 5950-3000 :

URL http://cp.literature.agilent.com/litweb/pdf/5950-3000.pdf
Agilent AN 369-5 Multi-frequency C-V Measurement of SemiconductorsP/N 5950-
2953 : URL http://cp.literature.agilent.com/litweb/pdf/5950-2953.pdf
Agilent Evaluation of MOS Capacitor Oxide C-V Characteristics Using the 4294AP/N
5988-5102EN : URL http://cp.literature.agilent.com/litweb/pdf/5988-5102EN.pdf
On-wafer Balanced Component Measurement using the ENA RF Network Analyzerwith
the Cascade Microtech Probing SystemP/N 5988-5886EN : URL
http://cp.literature.agilent.com/litweb/pdf/5988-5886EN.pdf
IV and CV Measurement Using the Agilent B1500A MFCMU and SCUU
P/N 5989-3608EN : URL http://cp.literature.agilent.com/litweb/pdf/5989-3608EN.pdf
Agilent 4070 Series Accurate Capacitance Characterization at the Wafer LevelP/N

5968-2499E : URL http://cp.literature.agilent.com/litweb/pdf/5968-2499E.pdf
128
CV Measurement And Calibration
Techniques
For CV (capacity versus voltage) measurements, two instrument settings are basically
possible:- a capacitance in series with a resistor- a capacitance in parallel with a
conductor.
This implies that when the device exhibits a more complex CV circuit than what is
anticipated with one of these settings, the capacitance (and the resistance/conductance)
will become a function of the measurement frequency. Therefore, for device modeling, a
frequency of 1MHz is typically applied. This frequency is sufficiently high enough to
achieve a good resolution with the capacitance meter (sufficient phase shift for the
resolution of the instrument), but also sufficiently low to avoid second order effects
(secondary capacitances, inductive influences etc.).
Regarding the measurement principle itself, there are two different methods available: ->
the current/voltage method e.g. HP 4280 -> the auto-balancing method e.g. Agilent
4284etc. Both can include an additional DC bias voltage for the required CV
measurements.
With the current/voltage method (fig.1), the unknown impedance Zx of the component
under test is calculated from the stimulating voltage, V1, and the measured current, I.
This current I is determined from the voltage drop, V2, along a known precision resistor,
R. In practice, R is replaced by a low-loss transformer in order to circumvent the problems
associated with a low-valued resistor. However, this limits the lower measurement
frequency. On the other hand, this method has advantages over the auto-balancing
method, in that it allows measurements of components connected to ground. More details
see the literature /Honda/, /Haruta/.
(1)Fig. 1: Capacitance measurement using the I-V method, e.g. HP 4280

In the auto-balancing method, shown in fig.2, the meter current again flows through
both the DUT and the reference resistor, R. Yet, the negative input of the OpAmp acts as a
virtual ground. The advantage of this method over the I-V is that the measurement range
can be changed by selecting a suitable value for R. In practice, the OpAmp is replaced by
a zero detector and a modulator, thus expanding the frequency range into the 100MHz
region.
(2)Fig.2: Capacitance measurement using the auto-balancing method, e.g. Agilent 4284
Important hint about CV measurements of biased semiconductor capacitances using an

auto-balancing bridge measurement
When adding applying a bias voltage to the auto-balancing method, however, special care
most be taken to ensure that the measurement OpAmp is not saturated by a low
resistance DUT. This can easily happen when measuring the CV curves of diodes into the
ON-state.
Referring to equation (2) of fig.2, the real (ohmic) part of the resistance Zx of the DUT,
i.e. the semiconductor resistance in parallel to the CV capacitance, has to be much bigger
than the reference resistor R. Otherwise, the output of the OpAmp can reach saturation
(?15V supply voltage). This happens e.g. when measuring semiconductor capacitances in
forward bias mode, where the ohmic part of Zx falls rapidly from quasi-infinite to a few
ohms !
A practical approach: when measuring a CV curve into the On-state of the diode, don't
apply a DC bias bigger than what corresponds to a capacitance ~3-5 times the 0V
capacitance CJ0.
129
CV Meter Settings:
Like with S-parameters, a CV measurement is a linear measurement, and therefore,

harmonics should be avoided. Therefore the applied CV signal level should not be too big,
but also not too small (reduce signal-noise ratio).
(Since there is usually no DC current flowing at CV measurements, the reason for
harmonics is the nonlinear capacitance itself !!).
You can verify the max. applicable signal level by increasing/reducing its value until you
get independent CV-curves.
Furthermore, a CV (LCRZ) meter interprets the measurement result with respect to a

capacitor in series with a resistor R, or as a capacitor in parallel with a conductor G. If this
pre-requisition is not met by the DUT, the capacitance and/or the resistor or conductor
become frequency dependent. In many cases, especially for silicon, this is true.
Therefore, check reducing the frequency from default 1MHz down to e.g. 100kHz and
check if the measured CV curve is affected or not. If it is not or no longer affected, the
mentioned simple underlying schematic is correct, since second-order effects are
switched-off due to the reduced measurement frequency. We now can perform a correct
CV measurement.
Hint:in IC-CAP, for the CV measurement, specify the measurement frequency in
Instrumnet Options, and declare the Setup Variable CV_FREQ for the simulation
frequency.
CONNECTING THE DEVICE

A so-called four-wire method is used for both CV measurement methods. Similarly to the
DC Kelvin measurement procedure, both the low and high pin of each port have to be tied
together at the DUT.
Hint: For proper 4-wire CV measurements, it is mandatory to have a proper guard
connection at the end of the 4 coax cables: connect all 4 coax shieldings together at the
DUT level, as depicted below.
Capacitance measurement with the auto-balancing bridge

In the auto-balancing method, the shieldings of these four coax cables are connected to
the virtual ground of the instrument's OpAmp. A special feature of this auto-balancing
method is that it is insensitive to stray capacitances towards Guard, the coax cables
shielding potential (virtual ground). As depicted in fig.3 below, this becomes clear since
the parasitic capacitance Cp1 is tied directly to the supplying bias and stimulus voltages,
and the parasitic capacitance Cp2 is located between virtual ground (minus input of the
OpAmp) and ground. This means, both parasitic capacitances are not included in the
measurement !
130
Fig.3: parasitic capacitances towards the coax cable shielding do not affect the Zx
measurement.
Extending the Guard shield to the DUT itself:if DUT Zx exhibits not only a capacitance
between its 2 connections, but also capacitances towards ground, these capacitances can
either be included with the measurement (fig.4a) or they can be excluded if this ground is
connected to the guard (shielding) of the CV meter coax cables (fig.4b).
Fig. 4: Shielding techniques for auto-balancing CV measurements:
In the same way, parasitic capacitances from the 'hot' CV pins to the other pins of the
DUT can be excluded frothe measurement result by connecting these extra pins to Guard,
like depicted in fig. 5 below.
Fig. 5: Measuring semiconductor capacitances
NOTE: How to connect a MOS transistor for Gate Oxide CV measurements see
further down in section CV On-Wafer Measurement Considerations For MOS Transistors
Calibrating the CV Meter
For modeling purposes, it is usually fully sufficient to perform an OPEN calibration (a
SHORT calibration would be required if inductors were measured). This means to measure
first the test fixture without the DUT (OPEN) and then, after this calibration was done, to
automatically (within the instrument) subtract this OPEN capacitance value from the
measurement. For measurements on the chip, an open dummy structure is contacted
during the OPEN calibration.
Shielding Of On-Wafer CV Measurements

For on-wafer measurements, connect the chuck as sketched below:
131
The following slide of Cascade Microtech emphasizes this guarding technique once again:
A study about 2-pin, 3-pin an 4-pin CV On-Wafer Measurements
This section is after the book:J.Berkner, Kompaktmodelle fr Bipolartransistoren, Expert-

Verlag Renningen (Germany),ISBN 3-8169-2085-3, Februar 2002
Notes:
This section refers to CV measurements applying the auto-balance measurement

method.- The results of these considerations can also be applied to any other device
like MOS or GaAs transistors.
When measuring a transistor on wafer, there are three contacting scenarios possible,
as depicted below: a 2-pin only method (leaving the chuck floating), a guarded
Chuck situation (as recommended by wafer prober manufacturers) as well as a full 4-
pin connection.
2-pin Measurements, case (a)

In this case, only the pads of the capacitance under test are contacted. For the
calibration of the capacitance meter (and thus the compensation of the cables plus
the contact pads capacitances), the pins are connected to an empty dummy structure
first. However, applying this 2-pin method to the DUT measurement, all parallel
capacitances are included in the measurement result as well, as shown below.
132
CV measurement of the BC junction capacitance with a 2-pin setup. left: measuring

the Dummy-Deviceright: measuring the DUT.
As depicted, the calibration including the dummy structure reflects a parasitic
capacitance of
while there is a bigger capacitance present during the 'calibrated' measurement of
the DUT, where also the other capacitances CCS, CBE as well as Cpad of the Emitter
and the out-calibrated Cpad of the Collector play a role !Since only Cpad/2 + Ccross
is calibrated out, the DUT measurement represents a too big capacitance.
CV Measurements with guarded chuck, case (b)

This is the recommended setup by wafer prober manufacturers: -> the chuck is
connected to the CV meter Guard:
In this case, during the OPEN calibration, the two pads capacitances are towards the
Guard potential of the CV meter, and, thus, cancelled out. The calibration basically
accounts for Ccross and second order capacitance effects. The Emitter pin is open and not
included in the calibration.
During the measurement, the Emitter pin plays a role: it gives a path to Guard for CBE ,
and, thus, cancels that capacitance out. CCS , with its connection to Substrate, is cancelled
out due to the measurement principle. The measurement result is the capacitance CBC, as
desired.
133
CV measurement of the BC junction capacitance with guarded chuck (3-pin setup)left:

measuring the Dummy-Deviceright: measuring the DUT.
4-pin Measurements, case (c)
In this case, all unconnected pads of the 2-pin measurement, including the chuck, are
connected with additional pins to the Guard, i.e. shielding of the CV meter coax cables.
With this forced guarding setup, all capacitances from the active pins towards guard are
excluded from the measurement result without any compromise. As can be seen in the
sketch below, the measurement setup is the cleanest method to measure nothing but the
desired CBC alone.
CV measurement of the BC junction capacitance with a 4-pin setup.left: measuring the

Dummy-Deviceright: measuring the DUT.
Important Hint:
You may now think, 'why should I calibrate when all parasitics are eliminated anyway'.Do
not forget that the CV meter calibration eliminates the offset capacitance between B and
C, which is not eliminated by the CV measurement principle !
CV On-Wafer Measurement Considerations For MOS Transistors
This section is based on Hiroshi Haruta, Agilent Technologies Japan, June 22, 2000
CV characteristic measurement is the standard method of analyzing MOS device
parameters including the thickness of gate insulation film, impurity concentration of the
substrate, flat band voltage, surface carrier density, threshold voltage, etc. It is also
performed to analyze the doping profile of a PN junction device.
How to set up the measurement:
134
An LCR Meter is connected to a wafer prober with test cables as illustrated.

What parameters are measured?
Voltage-capacitance characteristics are measured with a low-level signal (typically 20
mVrms) at 10 kHz, 100 kHz or 1 MHz where DC bias voltage is swept. Bias voltage is
typically varied from -5 V to 5 V in 10 mV minimum steps.
Key measurement points and issues:
Very low capacitance needs to be measured with high accuracy and high resolution (1 fF
resolution) at a low test signal level. Compensation must be properly performed at the
prober tips. The AC (RF) prober with a guarded wafer stage should be used to avoid noise
interference and to reduce the effects of stray capacitance around the probes. It is often
difficult to measure MOS devices below the 0.1 mm design level because gate tunneling
leakage current causes bias voltage error and test signal distortion.
NOTE: How to connect a MOS transistor for Gate Oxide CV measurements when
measuring a MOS Gate with a balanced bridge instrument, it is proposed to connect
the Gate to the low connectors, and the Substrate to the high connectors of the
CV meter.When measuring it using the I-V method, the Substrate is at low potential,
and the Gate is connected to the signal terminal of the CV meterFor more details, see the
publication: H.Suto et.al, 'Methodology for Accurate C-V Measurements of Gate Insulators
below 1.5nm EOT', Extended Abstracts of the 2002 Int'l Conf. on Solid State Devices and
Materials, Nagoya, pp. 748-749
----------------------------------------------------------------------------------------------------
----
In the example below, from the Agilent 4294A app.note, a special Setup for measuring the
capacitances Cgs, Cgd or Cds of a MOS FET is proposed, with the drain voltage varied.
Measurement setups for Cgs, Cgd and Cds measurements are illustrated. The DC bias
function of the instrument is used to apply the required drain voltage to the device. To
measure the intended capacitance without the influence of other capacitances, the
electrode of the device that is not connected to the measurement terminals must be
connected to the guard (outer shield conductor of the test cable).
Key measurement points and issues.

The impedance of the guard connection lead should be as low as possible to minimize
measurement error. The capacitor C2 used in Cgs measurement must have a sufficiently
low impedance to thoroughly de-couple the signal at the drain electrode.
CV Measurement of Packaged Devices
For packaged devices, we first perform an OPEN calibration including the test fixture.
135
Example of a text fixture (Interconnect Microwave: www.icmicrowave.com)

Then, for the measurement, we can either apply the 2-pin measurement (what yields the
total capacitance between the measurement pins), or a guarded measurement yielding
nothing but the desired capacitance between the pins D and S in the example below.
IMPORTANT NOTE:Apply either 2-pin or fully guarded measurements, but later, simulate
exactly what you have measured !!!
For device modeling, decide carefully what you want to measure!! You need to measure
what will be simulated later!! Or, the other way around, make sure to simulate what you
have measured before!!
Device Modeling of CV Curves and Corresponding Measurement Aspects
1. Comparing CV measurements directly with CV model equations, without calling a

simulator.
In this case, when you enter for example in IC-CAP the CV formula directly as a PEL
program (without using a simulator), you must ensure that your measurement represents
this situation as well. This is especially important for components with more than 2 nodes,
e.g. a transistor. For a bipolar transistor for example, you may have programmed with a
PEL function the CV formula
Therefore you need to make sure that the measured CBE does not include effects of the C
BC nor the CCE capacitance. This means, the open pins of the transistor (Collector and
Substrate) have to be connected to the Guard, i.e. the shielding of the CV meter coax
cables (virtual ground).
Of course, a proper CV meter calibration has to exclude the contact pad capacitances !
However, there is a caveat: don't forget that the DC behavior itself as well as the
additional parasitics (RB, RE, RC etc) can also influence the CV measurement and
simulation result !
The reason is that for a CV measurement, and also for the CV model equations as stand-
alone, it is assumed that - there is nothing in series with the capacitance- there is nothing
but a parallel resistance with the capacitanceBut with real measurements, - the
conductivity in the ON state of diodes and transistors acts like a SHORT to the CV curve-
and the transit time is overlying the CV measurement in the ON state.
2. Comparing CV measurements with CV simulations

This case is again a bit special for devices with more than 2 pins, e.g. transistors.In this
case, remember that in the model itself (e.g. a bipolar transistor), there is also a
capacitance between Base and Collector, and another one between Collector and Emitter !
I.e a CV simulation, with Input vBE (voltage) between nodes B and E, will return also the
Base-Collector capacitance in series with the Collector-Emitter capacitance. Like depicted
below, these 2 are in series, and together they are in parallel to CBE !!
136
Therefore, a standard CV simulation will always return the total capacitance between Base
and Emitter , including CBC and CCE!
NOTE:
Measurement-wise, make sure that this simulation, i.e. the total effective capacitance
between the two nodes, is respected with your measurement. Therefore, for the example
of CBE, it is not recommended to connect the open Collector to the virtual Ground of the
capacitance meter !
HINT:
capacitances to Ground (e.g. for our bipolar example from above: C to GND, B to GND, E
to GND and/or S to GND) do not contribute to CV simulations of the BE, BC, CE, and the
CS capacitance !! Therefore, once again, make sure to also have them excluded with the
measurements !!
HINT: if you want to check the simulated capacitance vs. the model equation itself,
switch-off the Base-Collector and the Collector-Emitter capacitance, i.e. set CJC=0,
CJS=0, or divide their actual values by 1E6 (to keep the previously extracted numbers
!!).And like in the case before, don't forget that the DC model itself and the additional
parasitics (RB, RE, RC etc) can also influence the CV simulation result! Therefore, for a
bipolar transistor, divide the actual parameter values of IS, RB, RE and RC by 1E6 each
for these CV simulations.
3. Comparing CV meter measurements with CV curves calculated from S-parameter
measurements
Here, the CV curves are calculated out of (properly de-embedded) S-parameter curves.
I.e. those S_deemb data are converted to Y-parameters, and we get
for a bipolar and the corresponding CGS, CGD and CDS for a MOS.
These matrix conversion formulae assume that there is no component in series with the
capacitances, and that there is nothing in parallel but a resistor.
Measurement-wise, these converted measurements should be compared with guarded or
4-pin CV measurements. See /D.MacSweeny/. Simulation-wise, they should be compared
either
with the pure CV model equations, - with a CV simulation where the other capacitors
and resistors are switched-off, - or with S-parameter simulations converted to CV
curves like the measurement data.
Simulating converted S-parameter simulations is best, since it takes everything into

account (the DC model and the outer parasitics), what corresponds exactly to the
measurement situation as well!
-> A Practical Approach for Modeling Transistor Capacitances:

It has become clear, that CV measurements look easy and simple, but can be quite
cumbersome depending on the actual device situation. Whatever method you prefer,
keep in mind that the model parameters obtained from CV measurements should be
considered as first order CV model parameters. The real world is non-linear RF, or at
least S-parameter fitting. Therefore, a practical and accurate solution is- to de-
embed the measured S-parameters, - convert them to Y-parameters, - show a plot of
MAG and PHASE (or REAL/IMAG) of all Yxx parameters- simulate S-parameters, and
convert them to these kinds of Y-parameters too,- and tune the capacitance
parameters (and the other model parameters also).
Hints on CV Meters:When configuring the CV meter, make sure that it is capable of

measuring the complete voltage range, and not merely some individual voltages:
e.g.: HP4284, opt.001
Probes using specific shielded CV probes, a measurement resolution for on-wafer CV

measurements down to a few Femto-Farad (fF) can be obtained with a CV meter, see
the following calibration verification result (after calibration, remeasured OPEN
dummy).
137
Publications:
Hiroshi Haruta, Agilent Technologies Impedance Measurement Handbook, 2nd
edition,
Agilent Technologies product number 5950-3000
Note: the older version was:M.Honda, The Impedance Measurement Handbook,
Yokogawa-Hewlett-Packard LTD,HP Literature Number 5950-3000
D. MacSweeney, Influence of Probing Conguration and Data Set Size for Bipolar
Junction Capacitance Determination, ICMTS Conference March 2002,
138
Network Analyzers
Contents
Measurement (iccapmhb)
Calibration (iccapmhb)
S-Parameters (iccapmhb)
De-embedding Techniques (iccapmhb)
Vector Network Analyzer - Basics for Modeling Engineers (iccapmhb)
139
Measurement
Contents
Linear Vector Network Analyzer Measurements (iccapmhb)

Displaying Phase Shift (iccapmhb)
140
Displaying Phase Shift
This section covers a special PEL routine which corrects the steps in the phase of S-
parameters. The following figure gives an example:
The PEL program given below shows how to implement this feature into IC-CAP.
!this routine adds 2*PI, if the phase sign changes and the phase step is
!more than 1.5*PI (to prevent from adding 2*PI at an ordinary phase
!cross-over!)
tmp=S.B
COMPLEX res.B.22[SIZE(tmp)]
!=============S11==========S11==========
!--------------treat meas.data.11 first
n = 0 ! counter for how many phase shifts have to be applied
i = 1 ! index
res.M.11[0] = ph(tmp.M.11[0])
WHILE i < SIZE(res)
phi1 = ph(tmp.M.11[i-1])
phi = ph(tmp.M.11[i])
IF ((phi*phi1) < 0 AND (ABS(phi-phi1) >= 1.5*PI)) THEN n = n + 2
res.M.11[i] = phi -n*PI
i =i + 1
END WHILE
!-------------------now treat simul.data.11
n = 0
i = 1
res.S.11[0] = ph(tmp.S.11[0])
WHILE i < SIZE(res)
phi1 = ph(tmp.S.11[i-1])
phi = ph(tmp.S.11[i])
IF ((phi*phi1) < 0 AND (ABS(phi-phi1) >= 1.5*PI)) THEN n = n + 2
res.S.11[i] = phi -n*PI
i =i + 1
END WHILE
!do the same with the partial arrays res.12, res.21, res.22 and finally return the whole array
RETURN res
141
Linear Vector Network Analyzer
Measurements
In this section, you will learn about:
Standalone VNA's
DC biased VNA's
Pad layout suggestions for transistor DC and S-parameter measurements
Frequency Limits for Connectors
Standalone VNA's
In order to characterize components above some hundred Megahertz, vector network
analyzers (VNA) are used. They measure the complex power transmission and reflection
coefficients at two locations, Port 1 and Port 2. The VNA basically consists of a frequency-
swept, high frequency generator (RF synthesizer), an S-Parameter Test Set to acquire the
injected and reflected power at the DUT at both ports, and a control and display unit
(Mainframe).
Figure: Vector Network Analyzer Structure
For the measurement, the DUT is connected to the two ports. The RF synthesizer signal is
fed into the S-Parameter Testset and applied alternately to port 1 and port 2. Measured is
the complex power signal of that specific frequency reflected back from the DUT to port 1,
the power signal transmitted from port 1 to port 2, and then, when the RF signal is
applied to port 2, the reverse behavior.
The block diagram of fig.2 shows the core of this meter combination, the S-parameter
testset. The RF Input source at the top, connected to the RF synthesizer, provides the
stimulus power. The PIN switch directs the signal to either a forward or a reverse S-
parameter measurement. Directional couplers then detect the injected and reflected
power of the DUT. The detected signals are then downcoverted into four IF signals for
further analysis in the VNA mainframe, where each input is digitized and signal processed
in order to give the S-parameters.
Figure: Block diagram of an S-Parameter Testset (4-sampler)
There are two main operating principles for VNAs: 3-sampler and 4-sampler VNAs. In case
of a 3-sampler, the reference signal is coupled-out before the Port1-Port2 switch.
Therefore, the switching errors are not covered by the calibration. For a 4-sampler VNA,
however, where the reference signals are detected after the switch, a better calibration
can be performed, including all error terms. Also, with the 4-sampler VNA, the user has
complete freedom with regard to his own calibration techniques and does not need to use
simplified procedures such as TLR* and LRM*. (Refer to Agilent Application Notes).
142
Figure: Principle of a 3-sampler and a 4-sampler VNA
Note
In both cases, the signal power at the reference port R should not be too small. Otherwise, the
downconverter might have phase locking problems. The error message reads in this case: "phase lock
lost". However, if the signal is too big, signal distortion may/will happen when measuring nonlinear
devices such as Transistors or diodes. This has to be absolutely avoided ! Otherwise, the VNA only
measures the fundamental frequency, ignores the harmonic frequencies, and the measured S-parameters
are completely wrong !!
They will not look distorted. They exhibit typical traces, however, the magnitude and the phase is wrong if
the measurement was overdriven, and thus, non-linear!!!
DC Biased VNA's
For S-Parameter measurements of active components, the operating point (DC bias) must
also be set and provided. Applying the DC bias Force-Sense technique, the measurement
setup then appears as shown in fig. 4. In this example, we apply external bias TEEs. Their
internal schematic is given in fig. 5. These bias TEEs allow Force-Sense techniques down
to some 10 nano-Ampere.
Using external bias TEEs reduces the 2? internal series resistance to typical remaining 1?,
and avoids the 1M? ESD protection resistance of the older VNA S-Parameter Testsets
(electro-static discharge).
Figure: VNA measurements with DC bias
Figure: The inner circuitry of the external bias TEE HP11612
143
However, also for these special bias TEEs, the total ohmic loss is not 0? Furthermore, as
can be seen from the schematic above, the ground unit GNDU of the DC analyzer has to
be connected always in order to have the DC ground tied together with the NWA ground !
About The Resolution of the External Bias Tees

The reason for the roughly 1nA DC current resolution limit of the external bias TEEs
HP11612 is that with the internal anti-oscillation circuits, there is a capacitor of roughly
5nF to ground for each DC port. This may confuse the internal firmware of the DC
analyzer. Because these instruments are optimized for use with semiconductors, they
assume a capacitive load of less than ~500pF. The charging/uncharging of such parasitic
capacitors is taken into account by the firmware's algorithm which decides, when to begin
with the DC current measurements. If there are bigger capacitive loads associated with
the DUT, the internal logic considers the current to be settled and starts the current
measurement. In fact, it still charges the parasitic capacitor and thus measures a current
which belongs rather to charging the capacitor rather than the DC current of the
semiconductor! Because the anti-oscillation capacitors are selected for rather worst-case
conditions, they may be replaced by smaller ones (~500pF). Of course, this depends on
the specific transistor. Also, it could be considered to add ferrite beads to the bias-TEE's
DC path.
Conclusion About BIAS Tee

GaAs modeling engineers usually simply apply standard bias TEEs without the force-sense
biasing. Silicon modeling engineers sometimes prefer the force-sense bias TEEs. As it was
mentioned above, in both case the DC bias loss is not 0?. For this reason, the DC losses
have to be taken into account anyway, with the Dut Test Circuit in the IC-CAP ModelFile.
As a consequence of all this, if the additional ~1.5? loss in the DC bias path can be
tolerated, you can also connect the triax cables of your DC analyzer via a triax-coax
adapter (triax middle shield (guard) left OPEN!), directly to the rear DC connectors of your
NWA.
Simply cover the ohmic loss by specifying an appropriate TestCircuit in the DUT level of
your IC-CAP ModelFile.
About Using the Internal Bias Tees of the VNA's S-Parameter

Testset
Note that there are a few things to consider: For adapting the DC bias to the rear panel of
the S-parameter testset ,make sure to apply a pair of TRIAX(f) - BNC(m) ADAPTERs with
the triax-middle-shielding unconnected.
The internal bias TEEs do not allow the employment of the Force-Sense technique,
however, this is no problem when specifying these DC losses in the IC-CAP Test Circuit
level. Typically, the S-Parameter Testset's internal bias TEEs have an overall resistance
between 2 .. 2.5? modern Agilent PNAs typically show 1? Thiswill lead to a lower test
voltage at the DUT than desired, if applying more than a few mA bias current. This effect
has to be taken into account by adding such a series resistor to each connection of the
transistor in the simulator circuit description, AC-wise shorted by a big capacitor (without
these big capacitors, the simulator would not use the correct bias setting for its S-
parameter simulations). Remember: only with these additional resistors in place, curve
fitting can be correct! Otherwise, we would compare simulated to measured S-parameters
whose bias conditions were different!)
Example for a enhanced SPICE circuit deck:
Note
For older HP VNAs, each port has an internal ESD-protection resistance of 1M? to ground. When
measuring both, the S-parameters and the DC bias current, this would reduce the min. current resolution
to a few micro-Ampere.
Measurement Hints
144
It may be smart to use 2 additional SMUs of the DC analyzers for the S-Parameter DC
bias.
In this case, no recabling is required when measuring S-parameters. All what is needed is
to tell the IC-CAP software to use the SMUs for DC measurements and to use the other
SMUs for the S-parameter biasing.
VNA Dynamic Range

Terminate e.g. Port2 with a 50 Ohm LOAD cal.standard and measure S21.
The VNA dynamic range is then the difference between max. applickable RF signal level
dBm and the measured noise level of MAG(S21)dBm .
You can improve the dynamic range by reducing the IF bandwith. Typical settings for
modern Agilent PNAs is 10Hz. This results in a typical signal-noise ration of ~100dB.
For older HP VNAs, you should increase the averaging instead.
Note
Measurement resolution improvement for PNAs is recommended via IF bandwidth reduction, and not
increased averaging.
Probing on the Wafer

The slides and text of this section are copied with friendly permission from the file
'RF-Microwave Training Presentation.pdf' of the Cascade Microtech Winter 2002 European
Probing Seminar CD-ROM
145
146
Use the ISS substrate to check the recommended 25um overtravel of your probing
contacts:
The following figure gives and idea, how transistor layouts should look like to cover both,
DC measurements and biased S-parameter measurements. For bipolar transistors, it is
valid correspondingly.
147
Figure: Layout for DC and S-parameter measurements of transistors
For S-parameter measurements, a good Grounding is essential, and Ground loops must be
avoided. Therefore, the Bulk and Substrate contacts are shorted to Ground reliably with
the above layout and accurate S-parameter measurement results can be obtained. If an
individual DC bias is required for the Bulk or Substrate contact, it is recommended to use
Ground-Signal probes, and to use a 3rd probe of this type to bias the 3rd contact. This
reduces possible device oscillation considerably.
If an individual DC bias is required for the Bulk or Substrate contact, it is recommended to
use Ground-Signal probes, and to use a 3rd probe of this type to bias the 3rd contact. This
reduces possible device oscillation considerably.
Figure: Layout for S-parameter Measurements and Individual DC bias for Bulk or Substrate, using
Ground-Signal Probes
Note
The distance between the center of the pads is called pitch. Typical pitches for on-wafer transistor
modeling is about or below 100um.
See also:
T.E.Kolding, A Four-Step Method for De-Embedding Gigahertz on-Wafer CMOS
Measurements, IEEE Transactions on Electronic Devices, vol.47, no.4, April 2000
Frequency Limits For Cables and Connectors
Connector Frequency Limit

3.5mm < ~ 30GHz
2.4mm < ~ 50GHz
1.85mm < ~ 65GHz
1mm <~110GHz
Web Info
www.agilent.com: search for Application Note 95-1, S Parameter Techniques for Faster,
More Accurate Network Designs.
Publications
1. Paul Schmitz, Vector Measurements of High Frequency Networks, Hewlett-Packard,

Publication HP5958-0387, April 1989
2. Agilent Technologies Application Notes
1287-1: Understanding the Fundamental Principles of Vector Network Analyzers,
Pub.No. 5965-7710E, 1997
1287-2: Exploring the Architectures of Network Analyzers, Pub.No 5965-7708E, 1997
1287-4: Network Analyzer Measurements: Filter and Amplifier Examples,
Pub.No5965-7710E, 1997
3. Hewlett-Packard 1997 Back to Basics Seminar, D.Ballo, Network Analyzer Basics,
Pub.No. 5965-7917E
148
4. Hewlett-Packard Application Note 8510-5A: Specifying Calibration Standards for the
HP8510 Network Analyzer, Publication HP5956-4352, February 1997
5. Cascade Microtech, Microwave Wafer Probe Calibration Constants, HP8510 Network
Analyzer Input, Instruction Manual, 1990
6. Cascade Microtech Application Note: On Wafer Vector Network Analyzer Calibration
and Measurements, 1997, Pub No. PYRPROBE-0597
7. Cascade Microtech Application Note: Layout Rules for GHz Probing, 1992, Pub.No
LAYOUT19
8. CascadeMicrotech Tecnical Brief: A Guide to Better Vector Network Analyzer
Calibrations for Probe-Tip Measurements
9. Cascade Microtech, Winter 2002 European Probing Seminar CD-ROM
149
Calibration
Contents
Calibration Procedures (iccapmhb)

Defining Custom Calkits (iccapmhb)
Introduction to NWA Calibration (iccapmhb)
NWA Calibration With Gating in the Time Domain (iccapmhb)
VNA Cal Kit Manager (iccapmhb)
150
Calibration Procedures
In calibration procedures, you will learn about:
NWA Calibration in General (iccapmhb)

Entering CalKit Data into the Network Analyzer (iccapmhb)
Entering Calkit Data for On-Wafer Probes (iccapmhb)
Entering Calkit Data for Packaged Devices (iccapmhb)
Entering Calkit Data for Connectorized Devices (iccapmhb)
Entering SOLT Calkit Data - HP Network Analyzer specific (iccapmhb)
Entering SOLT Calkit Data - Agilent PNA Network Analyzer Specific (iccapmhb)
Entering TRL Calkit Data - Agilent Network Analyzer specific (iccapmhb)
Checking Maximum Applicable RF Signal Level (iccapmhb)
NWA Instrument Options in IC-CAP (iccapmhb)
Performing the NWA Calibration from IC-CAP (iccapmhb)
NWA Calibration Verification (iccapmhb)
DC Bias Losses During S-Parameter Measurements (iccapmhb)
NWA Calibration Procedure for At-Temperature Measurements (iccapmhb)
151
Checking the max. Applicable RF Signal
Level
It is very important to not overdrive the test device with too much RF signal. As
mentioned, it must be absolutely ensured that particularly the transistor's output signal is
not clipping. When this happens, the transistor output signal is no longer a sine function,
i.e. the transistor behaves nonlinear. This is a contradiction to a linear NWA ! In this case,
the NWA will only measure the fundamental frequency, and ignore the occurring
harmonics !
On the other hand, a too small RF signal will result in noisy S-Parameter measurements.
In this context, keep in mind that:

-> a typical RF signal level for modeling transistors is about -40dBm
i.e. the NWA port1 is set to -30dBm, plus additional ~10dB attenuation of cables and
connectors.
-> -40dBm corresponds to a RMS voltage of ~2.2mV, what corresponds to ~3mV peak or
~6mV peak-peak !
-> small signal condition for a transistor is usually defined as VT / 10 = 2.7 mV
-> Last not least, when harmonics occur, i.e. the test device is overdriven,
the DC operating point is affected and thus, the DC bias current is affected.
We can use this effect of a shifted DC bias to identify the max. applicable RF signal for
NWA measurements with a smart and simple measurement method :
When measuring a DC output characteristics and calculating Rout out of it, the resulting
curve is very sensitive. Therefore, we can use this plot to identify possible effects of a too
big a AC power applied to the transistor. This means, we measure the DC output
characteristics, and let the NWA operate in continuous mode, i.e. unsynchronized to the
DC measurement. When this happens, i.e. when the operating point is shifted (ib and ic
for bipolar, id for MOS), harmonics do occur! We then reduce the RF power a bit and know
the max. allowed RF power for the NWA S-parameter measurements!
The following plot reflects such a test. The disturbed curves happen when there is too
much RF power applied to the transistor.
And here another example:
Here some more examples of too big RF signals, causing the transistor to behave
nonlinear:
Too big RF signal affects the iB of a bipolar current, when comparing the S-parameter iB
with the previously measured DC iB:
152
The same, now referring to iC:
Referring to the Gummel plot:
Last not least, a too big RF signal can disturb also the fT curve:
153
Note
By the way, self heating affects the slope 'after' the maximum fT.
Important Note on the Agilent N5250A 110GHz PNA

Network Analyzer
In this network analyzer, the power in the 67-110 GHz range is not controlled by the PNA.
It is provided directly by the 67-110 GHz Olesson test head. The power generated by the
source in the head is constant and it is around 0dB.
Therefore, the only way to decrease this power is by using the attenuators.
In other words, the power settings in the PNA only controls the power up to 67 GHz.
Therefore the trick of setting -30 dB and using the power slope to compensate for the
losses between the PNA port and the head port, can only be used in the first band. For a
sweep from 10 MHz ... 110 GHz, across the 67 GHz, it is recommended to set the source
power in IC-CAP to 0dBm, and to use the attenuatiors on Port1 and Port2 for signal levels
which are low enough for your transistor.
Considering the DC Bias Losses During S-Parameter
Measurements
After the network analyzer has been properly calibrated, and this calibration has been
verified, we finally have to take the DC bias losses into account, when measuring and then
modeling the S-parameters of e.g. diodes or transistors.
These DC losses are NOT calibrated out by the VNA calibration !

Therefore, we perform 2 measurements:
we place the GSG probes on the SHORT standard of the ISS calibration substrate and
measure the ohmic loss.
we perform the same measurement on the SHORT dummy on the wafer.
Hint: The following notes refer to the NPN_MEASURE_MASTER.mdl file of the IC-CAP
demo_features directory:
When placing the GSG probes on the SHORT standard of the ISS substrate, and when
stimulating a current into both probes, we can measure the voltage drop of the contacts.
Such a measurement result is depicted below:
Note
The contact resistance to the gold on the ISS is usually different (much lower) compared to the contact
resistance to the aluminum pads on the wafer, see the next measurement.
For the measurement results, we consider these resistors:
154
Assuming an ideal contact on the gold SHORT of the ISS substrate, we measure in this
setup basically Rbiastee.
Then, we repeat these measurement conditions, but now for the SHORT on the wafer. This
will show up as an additional resistor due to the worse GSG-aluminum contact compared
to the previous GSG-gold contact.
Note
You can use the SHORT dummy on the wafer, but you can also place the GSG probes an any aluminum
area on the wafer. It only needs to make sure we have a good aluminum short between the GSG probe
pins.
For the interpretation of the aluminum SHORT measurements, we consider the following
schematic:
Assuming 3x the same contact resistance for the GSG probe, we measure in this setup
basically
Rtotal = Rbiastee + 1.5 * Rcontact.
Rbiastee is known from the previous ISS contact measurement. So we get for Rcontact:
Rcontact1 = (Rtotal1 - Rbiastee1) / 1.5

Rcontact2 = (Rtotal2 - Rbiastee2) / 1.5
How to take these contact and bias-TEE resistors into account for
your S-Parameter modeling?
For your S-parameter measurements, the DC operating points are affected by these DC
losses.
This loss is not taken into account by your VNA calibration. Therefore, the easiest way to
perform a correct S-parameter modeling is to to simply take these DC losses into account
during your simulations.
This means:
In your S-parameter DUTs, add a Test Circuit which reflects these DC loss resistors, and
enter there
Rtotalport1 = Rbiastee1 + 1.5* Rcontact1
Rtotalport2 = Rbiastee2 + 1.5* Rcontact2
Make sure to add a big parallel capacitor to Rtotalport1 and Rtotalport2, in order to not
affect the S-parameter simulations by this ohmic loss resistor (which affects the DC bias
point, but NOT the S-parameters!!!).
155
For an example, see the Test Circuit in DUT NWA_Spar in
demo_features\3_MEAS_ORGANIZE_n_VERIFY_DATA\0_MASTER_FILES\NPN_MEAS_MASTER_demodata_PELdep.mdl
Here is the corresponding ADS syntax:
define emul_bias_TEE (P1 P2)

;The following network emulates the losses in the bias-TEEs
;and the wafer contact losses.
;Since the losses affect only the DC path, they are shorted
;for the AC signals by the capacitors C_short_x
;The values of R_serx are measured in Setup 'check_DC_contact_Spar'
R:R_ser1 P1 P11 R=1m
C:CAC1 P1 P11 C=.1
R:R_par1 P11 0 R=1MEG
R:R_ser2 P2 P22 R=1m
C:CAC2 P2 P22 C=.1
R:R_par2 P22 0 R=1MEG
;call the subcircuit description of the transistor,
;as defined in Tab 'Circuit' of the model file
VBIC_EXTRACT_MASTER:X1 P22 P11 0 0
end emul_bias_TEE
And this is the SPICE syntax:
.SUBCKT emul_bias_TEE 1=P1 2=P2

R_ser1 1 11 1.5
C_short1 1 11 .1
R_par1 1 0 1MEG
R_ser2 2 22 1.5
C_short2 2 22 .1
R_par2 2 0 1MEG
*call the subcircuit description of the transistor,
*as defined in Tab 'Circuit' of the model file
X1 22 11 0 0 VBIC_EXTRACT_MASTER
.ENDS
Entering Calkit Data for Connectorized Devices

When using and Agilent network analyzer with connector calkits, e.g. the 3.5mm calkit,
the calkit data are already available in the instrument and need not to be entered
manually. For the 26.5GHz 3.5mm calkit, the data are for an 'insertable device', i.e. one
cable with a male connector, the other with a female connector (Tdelay_THRU = 0s) , are
shown in the table below:
HP85052B 3.5mm calkit (26.5GHz) for an 'insertable device':
Note
For non-insertable devices, see further below.
HP85033D 3.5mm calkit (6GHz) for HP8753 NWAs and again for an 'insertable device':
Note
For non-insertable devices, see further below.
For the 2.4mm 85056A calkit (50GHz) and the 8510C network analyzer, the data are
(neglecting the frequency-dependent C's and L's), again for an 'insertable device':
Insertable and Noninsertable Devices
Note
The THRU cal standard specification with 0 ps is assuming a 'male-female' connection during the THRU
measurements. This assumes, on the other hand, a male and a female connector at your DUT.
If your DUT exhibits connectors with the same sex, you need to use a 3.5mm THRU
connector for the THRU calibration. Such a THRU connector may not be included in the
calkit box. In this case, assume a delay of
156
TD=measured_ThruLength / C0 with C0 = 3E+8 m/s
for the delay of such a conventional THRU connector. In this case, however, you have to
modify the calkit and enter the THRU delay time into your NWA calkit data!
A typical value for HP/Agilent 3.5mm THRU adapter (12 mm length 'w/o connectors')
is TD = 95ps .
Below a list of modeled 3.5mm CalKit THRUs.
If you instead have the two cable connectors with the same sex, and need to connect later
to a male-female DUT, screw a THRU connector to one of your cables, and leave it
connected also during the calibration.
HINT:
Even if you measure on-wafer, the separate connectorized calkit box might be interesting
to buy as well, since it allows you to check the network analyzer performance separately.
IMPORTANT HINT:
Under www.vnahelp.com, you can find a wealth of information regarding NWA calibration
for connectorized devices.
Entering Calkit Data for On-Wafer Probes
CASCADE MICROTECH
The following is a table of calkit standards definitions for Cascade 100um GSG probes.
For the other Cascade ACP (air coplanar probe) G-S-G probes the calkit standards values
are:
CalKit constants for Cascade G-S-G probes:
Pitch C-Open L-Short L-Termination

100 -9.3fF 2.4pH -3.5pH
125 -9.5fF 3.6pH -2.6pH
150 -9.7fF 4.8pH -1.7pH
Independent of the pitch, the THRU delay time is always 1ps.
This is the table for the assignment of the cal kit classes:
Table 2:standard class assignment, for all kinds of probes
157
S11A 1 OPEN
S11B 2 SHORT
S11C 3 LOAD
S22A 1 OPEN
S22B 2 SHORT
S22C 3 LOAD
FORWARD TRANSMISSION 4 THRU
REVERSE TRANSMISSION 4 THRU
FORWARD MATCH 4 THRU
REVERSE MATCH 4 THRU
FORWARD ISOLATION 3 ISOLATION
REVERSE ISOLATION 3 ISOLATION
Note on the 500 entry for Z0 of the Cascade LOAD:
or:What to do if I the NWA does not allow to enter a calkit value:
The Cascade LOAD represents a precisely trimmed 50 DC characteristics, plus an

inevitable inductance. In order to reflect this inductive performance at higher frequencies,
it would be necessary to enter the value of that inductance into the NWA calkit. However,
most NWAs do not allow to enter an inductance. However, they allow to enter the
characteristic impedance Z0 plus the delay time of a lossless delay line. Now, it is
important to know that the S11 curve of a delay line of exactly Z0delayline = 50 is
completely in the center of the Smith chart. Delay lines with a characteristic impedances
Z0delayline above 50 look like an inductor in series with a resistor of 50, and delay line
impedances below Z0delayline = 50 look like a capacitance in parallel with a resistor of
50. See the sketches below:
Therefore, since the Cascade LOAD behaves like a precisely trimmed resistor of 50 in
series with an inductor, a delay line with Z0delayline = 500 (what is the max. entry
allowed for Agilent's NWAs) are entered plus a delay time. This delay time value is
obtained from the lumped inductance by the following conversion formula:
T_delay = L / Z0delayline = L / 500

On the other hand, the Picoprobe LOAD behaves like a resistor of 50 in parallel with a
capacitor. Therefore, we have to enter a delay line impedance of Z0delayline = 5 (what
corresponds to the min. entry allowed for Agilent's NWAs) and a corresponding delay
time! See the specification sheet of the Picoprobe GSG probes.
Note
If The Network Analyzer does not allow L0 FOR THE SHORT, but only a delay, apply the above formula as
well. Use Z0delayline = 500 for both, Cascade and Picoprobe (since we want to convert an inductance to
a delay line, see the notes on a delay line with Z0delayline > 50). For example, to replace the L_short =
2.4pH , a 4.8fs delay and Z0delayline = 500 are entered into the NWA calkit table instead.
IF THE NETWORK ANALYZER DOES NOT ALLOW TO ENTER A NEGATIVE DELAY TIME: In
this case, the negative series inductance has to be converted into a parallel
capacitance:For a load resistance R0 in series with an impedance LSer < 0, there is
....(1)
Setting (1)=(2) gives
what finally gives a modified conversion formula
Example:
R0 = 50, Z0 = 5, Lser = -3.5pH, and so we enter as calkit data, following the hints on
delay lines with Z0 < 50 (see above):
158
and
Acknowledgements: thanks to Mr. Meiser of Infineon/Munich for this hint.
Note on the negative capacitance of Cascade probes for an OPEN

The goal of a calibration is to ensure later a correct measurement of the DUT without any
parasitics caused by the measurement cables etc. If the DUT was a 50 coplanar line, it is
expected that the power from port 1 is fed from the probe tip completely into this DUT.
Now, an OPEN Cascade probe is also a 50 line, yet cut off at the end of the probe tips.
When measuring this lift-off probe, the electro-magnetic field at the probe end is distorted
and no longer the homogeneous one of a 50 line. In other words, it will of course not
correspond to a precisely cut-off homogeneous field. It will rather extend a bit over the
OPEN end into the air. On the other hand, if the probe is later connected to that 50
coplanar line, it is obvious that there should be no discontinuity of the electro-magnetic
field any more, see the following figure. Therefore, the distortion, shown with the OPEN
probe tip, must correspond to a nonideality which is not an ideal infinite Ohm
representation.
Note
Another explanation for a negative capacitance with probes in the air is that the wave propagation in air is
faster than on the wafer.
For more details, see the publications of Cascade Microtech mentioned at the end of this
chapter.
PICOPROBE
CalKit constants for Picoprobe G-S-G probes:
Pitch C-Open L-Short LOAD Delay .

75 6.5fF 3.9pH 19.7fs and Z0=5
100 6.5fF 5.0pH 15.5fs and Z0=5 or 10fs/500
125 6.5fF 7.1pH 11.5fs and Z0=5
150 6.5fF 8.8pH 8.0fs and Z0=5 or 17.6fs/500
Independent of the pitch, the THRU delay time is always 1.13ps, because the Picoprobe
cal substrate THRU is 200um long, probe contact distance from each end is 25um each, so
the effective THRU is 150um. Corresponding to the dielectric constant of the Picoprobe cal
substrate, the corresponding delay time is then 1.13ps.
Note
For an explanation of the Z0=5 of the LOAD, please see above.
CalKit constants for Picoprobe G-S probes (to be used with the CS-8
Cal.Substrate)
Pitch C-Open L-Short LOAD Delay

100 4.3fF 9.2pH or 18fs/500 0
Independent of the pitch, the THRU delay time is always 1.13ps
Note
The next slides are from Cascade Microtech, so we have decided not to upload them.
159
Entering Calkit Data for Packaged Devices
Intercontinental Microwave (packaged devices) Test Fixture
Standard Class Assignment
S11A OPEN
S11B SHORT
S11C LOAD
S22A OPEN
S22B SHORT
S22C LOAD
FORWARD TRANSMISSION THRU
REVERSE TRANSMISSION THRU
FORWARD MATCH THRU
REVERSE MATCH THRU
FORWARD ISOLATION ISOLATION
REVERSE ISOLATION ISOLATION
FREQUENCY RESPONSE
TRL THRU 1
TRL REFLECT 4
TRL LINE 2 3
Entering CalKit Data into the Network Analyzer

General Info
About Calibration Kits (a copy of the Agilent PNA Help pages)
A calibration kit is a set of physical devices called standards. Each standard has a precisely
known or predictable magnitude and phase response as a function of frequency.
In order for the analyzer to use the standards of a calibration kit, the response of each
standard must be mathematically defined and then organized into standard classes that
correspond to the error models used by the analyzer.
To be able to use a particular calibration kit, the known characteristics from each standard
in the kit must be entered into analyzer memory or recalled from a default list of
calibration kits stored in the analyzer.
Calibration Standards
Calibration standards provide the reference for error-corrected measurements in the
network analyzer. Each standard has a precisely known definition that includes electrical
delay, impedance, and loss. The analyzer stores these definitions and uses them to
calculate error correction terms.
During measurement calibration, the analyzer measures standards and mathematically
compares the results with "ideal models" of those standards. The differences are
separated into error terms that are later removed from device measurements during error
correction.
Standard Type
A standard type is one of four basic types that define the form or structure of the model to
be used with that standard. The four standard types are shown below:
160
Standard Terminal Impedance
SHORT zero ohms
OPEN infinite ohms
LOAD system impedance, Z0
THRU/LINE no terminal impedance
ARBITRARY user-defined
Standard Definitions
Standard definitions describe the electrical characteristics of the standards and the
frequencies they will be used. Refer to the "standards definition table" included with the
calibration kit. Standard definitions include:
Minimum Frequency Specifies the minimum frequency the standard is used for
calibration.
Maximum Frequency Specifies the maximum frequency the standard is used for
calibration.
Z0 Specifies the characteristic impedance of the standard (not the system
characteristic impedance or the terminal impedance of the standard).
Delay Specifies a uniform length of transmission line between the standard being
defined and the actual calibration plane.
Type Specifies type of standard (SHORT, OPEN, THRU/LINE, LOAD, ARBITRARY).
Loss Specifies energy loss, due to skin effect, along a one-way length of coaxial
cable.
Loss Model Equation
The value of loss is entered as ohms/second at 1 GHz.

To compute the loss of the standard, measure the delay in seconds and the loss in dB
at 1 GHz. Then use the following formula:
Capacitance Model Equation
C0, C1, C2, C3. Specifies the fringing capacitance for the open standard.
C = (C0) + (C1 x F) + (C2 x F) + (C3 x F)

(F is the measurement frequency).
The terms in the equation are defined when specifying the open as follows:
C0 term is the constant term of the third-order polynomial and is expressed in
Farads.
C1 term is expressed in F/Hz (Farads/Hz).
C2 term is expressed in F/Hz.
C3 term is expressed in F/Hz.
Inductance model equation:
L0, L1, L2, L3. Specifies the residual inductance for the short standard.
L = (L0) + (L1 x F) + (L2 x F) + (L3 x F)

(F is the measurement frequency).
The terms in the equation are defined when specifying the short as follows:
L0 term is the constant term of the third-order polynomial and is expressed in
Henries.
L1 term is expressed in H/Hz (Henries/Hz)
L2 term is expressed in H/Hz.
161
L3 term is expressed in H/Hz.
Class Assignments
Once a standard is characterized, it must be assigned to a standard "class". A standard
class is a group of standards that are organized according to the calibration of the PNA
error model.
The number of classes needed for a particular calibration type is equal to the number of
error terms being corrected.
A class often consists of a single standard, but may be composed of multiple standards,
such as loads or delay lines. Refer to the calibration kit "class assignment" table.
Example: A response calibration requires only one class, and the standards for that class
may include an OPEN, or SHORT, or THRU. A 1-port calibration requires three classes. A
full 2-port requires 10 classes, not including two for isolation.
The number of standards assigned to a given class may vary from one to seven for
unguided calibrations. Guided calibrations allow as many standards as needed.
The different classes used in the PNA:
S11A, S11B, S11C (S22A, S22B, S22C and so forth)

These are the three classes for port 1-reflection calibrations (three classes also for S22
and S33). They are used in the one-port calibrations and the full two-port calibration.
They are required in removing the directivity, source match, and reflection tracking errors.
Typically, these classes might consist of an open, a short and a load standard for each
port.
Transmission and Match (forward and reverse)
These classes are used to perform a full two-port calibration. The transmission class
relates primarily to the transmission tracking, while the match class refers to load match.
For both of these classes, the typical standard is a thru or delay.
Isolation
The isolation classes are used to perform a full two-port and the TRL two-port calibrations.
The isolation classes apply to the forward and reverse crosstalk terms in the PNA error
model.
TRL thru
These are used to perform a TRL two-port calibration. The TRL thru class should contain a
thru standard or a short line. If it contains a non-zero length thru standard, then the
calibration type is called LRL or LRM.
TRL reflect
This class is used to perform a TRL two-port calibration. The TRL reflect class should
contain a standard with a high reflection coefficient, typically an open or preferably short.
The actual reflection coefficient need not be known, but its phase angle should be
specified approximately correctly ( 90 deg).
The exact same reflection standard must be used on both ports in the TRL calibration
process.
TRL line (or match)
These are used to perform a TRL two-port calibration. The TRL line or match class should
contain line standards, load standards, or both. If a line standard is used, its phase shift
must differ from that of the TRL thru standard by 20 to 160. This limits the useable
frequency range to about 8 to 1. Two or more line standards of different lengths may be
specified to get broader frequency coverage. It is also common to include a load standard
for covering low frequencies, where the line's length would be impractically long. When a
load is used, the calibration type is called TRM or LRM.
Note
For more information, read application note 8510-5A, "Specifying Calibration Standards for the Agilent
8510 Network Analyzer". Although the application note is written for the Agilent 8510 series of network
analyzers, it applies to the PNA as well. The part number of the publication is 5956-4352.
Entering SOLT Calkit Data (Agilent PNA-

NetworkAnalyzer-specific)
Procedure for the older PNA Firmware:
1. Use the mouse, go to the 'Calibration' menu, click 'Advanced Modify Cal Kit'.
2. Select 'Edit Version 1 Kits'.
3. In the new window, select the desired calibration kit ID number, e.g. 7.
Note
To restore the selected calibration kit to its predefined values (ID numbers 1-6), click Restore
Default Kit.
4. In the 'Cal Kit Name' box, enter your calibration kit name for the new calkit.
Assign Standards to Ports
1. Click the 'Port 1 Label' box and enter your port 1 name, e.g. probe1.
162
2. In the 'Port 2 Label' box, enter your port 2 name, e.g. enter probe2.
3. In the section 'Choose Standard Nos. for Cal Classes, enter a name for the Cal class
and make the selection for the Standard number:
Cal.Class Cal.Class Label Standard No.
S11A 'OPEN' =2
S11B 'SHORT' =1
S11C 'LOAD' =3
S21T 'THRU' =4
S22A 'OPEN' =2
S22B 'SHORT' =1
S22C 'LOAD' =3
S21T 'THRU' =4
Modify Standard Definition

4. In the Standard No. box in the 'Modify Standard Definition' portion of the dialog box,
click the arrow button to select the desired standard number.
5. For the standard number selected in the previous step (during Assign Standards to
Ports), enter a label (e.g. Standard #1 is 'SHORT', #2 is 'OPEN', #3 is 'LOAD', #4 is
'THRU'.
6. Then, enter all the definitions of the standard (for example, Minimum Frequency,
Maximum Frequency, Z0, Delay, Loss, etc). Refer to the calkit data of your probes,
as given ins the tables further below.
7. Repeat steps 1 to 3 for each standard number.
8. Click OK.
You are back in the starting menu item.
Click OK.
Procedure for the new PNA Firmware (PNA Type 'B'):

On the PNA, select Calibration/AdvancedModifyCalKit:
163
164
165
After this, we are back in IC-CAP.
We have to consider two cases:
A- In Instrument Options, 'Use PNA Calibration Settings'= No

B - 'Use PNA Calibration Settings'= Yes (This is recommended).
CASE (A): In Instrument Options, 'Use PNA Calibration Settings'=

No
To perform the PNA calibration from IC-CAP, load the file

demo_features\3_MEAS_ORGANIZE_n_VERIFY_DATA\0_MASTER_FILES\2_NWA_CAL_VERIFICATION\CAL_VERIFY_MASTER_PELdep.mdl
Begin with Setup 'check_max_RF':
Connect the transistor

Check the DC output characteristics sweep conditions
On the PNA, set the port attenuations to the highest dB values and the PNA signal
level to the lowest possible value.
Set the PNA to continuous measurement mode
166
In IC-CAP, hit measure and check the plots.
You should get an ordinary output characteristics. No distortions.
Convert these measured data to reference data (in the Outputs of the Setup, change
from Type 'B' to 'M', hit enter, and change from 'M' to 'B'. This means that there are
now pseudo-simulation data available which represent the output characteristics with
no RF distortion (small signal RF excitation condition).
Then, on the PNA, manually increase the power level (if necessary, adjust the
attenuator settings), so that you get as starting RF conditions e.g. -30dBm at port1
and -10dBm at port2.
For transistors, you should select power-decoupled ports, i.e. you specify for each
port an individual power level.
Make sure the PNA is still in continuous measurement mode (remember, the PNA is
NOT controlled by IC-CAP at this stage !!)
Re-perform the DC output measurement with IC-CAP.
If the power levels are still ok, the new (red) curve is identical to the previous
(yellow) reference curve. If there is a distortion visible, namely for low currents, the
applied power levels are already too big for the transistor. Repeat this step with
reduced PNA power levels.
If there is no distortion, you may try to increase the power levels.
After that, proceed in IC-CAP with Setup 'OPEN'.
Check the frequency range
In 'Instrument Options', specify
Port Power Coupled: No
Port Atten Auto: Yes
Cal Type: H
and enter the power levels for each port as evaluated before.
For 'Cal/State File Name' enter e.g. myCal_1.cst (this will be the calset filename
stored after the manual PNA calibration)
Hit 'Calibrate' in IC-CAP.
IC-CAP displays a window stating that you should now manually calibrate the PNA.
DO NOT CLOSE THIS IC-CAP WINDOW UNTIL YOU ARE FINISHED WITH THE
CALIBRATION AND HAVE SAVED THE RESULT IN THE PNA myCal_1.cst FILE
On the PNA, select the 'Calibration Wizzard' menu, and work through until you see the
above depicted calibration window.
connect the cal. standard on the ISS substrate and click on the corresponding green
button in the PNA cal window
Finally, save the calibration manually as calset myCal_1.cst in the default directory
of the PNA:
On the PNA, select File/SaveAs, and select the file type *.cst),
Save the .cst file on the PNA-computer under
'C:\Program Files\Agilent\Network Analyzer\Documents '.
This is the directory IC-CAP will ask the PNA to look for the correction data
when executing 'Measure' later.
Final Note: make sure that for all subsequent IC-CAP Setups applying this PNA calibration,
the Input 'freq' must be identical as well as the 'Instrument Options' !!!
I.e.: copy-paste the 'Instrument Options' to all other PNA Setups in your IC-CAP ModelFile
!
167
Note
On the Agilent PNA instrument mainframe, the default directory for saving and reading calibration (.cal)
and state/cal (.cal) files is:
C:\Program Files\Agilent\Network Analyzer\Documents
Calibration files can be saved and read in a different directory by setting the System Variable
PNA_CAL_PATH_FILE to another directory.
Example: PNA_CAL_PATH_FILE=C:\my_dir
CASE (B): (RECOMMENDED) In Instrument Options, 'Use PNA

Calibration Settings'= Yes
In this case, you perform the PNA calibration without any initialization from IC-CAP !
I.e. applying WinCal, SuessCal or the PNA Cal Menu
After that, on the PNA save the calset manually to the CalFile Name specified later in
IC-CAP.
On the PNA, select File/SaveAs, and select the file type *.cst).
Save the .cst file to the default directory (IC-CAP will search the file there !).
After that, you only need to set the frequency points in the IC-CAP Setup accordingly to
what you had used before on the PNA. And as mentioned just before the
InstrumentOptions entry 'Cal/State File Name' must match the .cst file name used before
when we had saved th PNA calibration.
Note
Make sure that for all subsequent IC-CAP Setups applying this PNA calibration, the Input 'freq' must be
identical as well as the 'Instrument Options', that is, copy-paste the 'Instrument Options' to all other PNA
Setups in your IC-CAP ModelFile
Entering SOLT Calkit Data HP NetworkAnalyzer

specific
STEP BY STEP PROCEDURE FOR HP8510 and 87xx:
Note
using Picoprobes, you may simply use the provided floppy disc and follow the instructions given in the
Picoprobe Calibration Substrate CS-5 Manual.
We start with a calkit list as below.
168
The calkit entering procedure below refers to the calkit table above. If in your own calkit
table the OPEN and SHORT entries are flipped, i.e. Standard Nr.1 = SHORT, Standard Nr.2
= OPEN, either exchange these lines in your calkit table and proceed as described below,
or keep that sequence in your calkit table and change the corresponding lines in the
procedure below.
Changing calibration constants (modifying a calkit) for the HP 8510

(very similar for Agilent 87xx family):
see also HP8510-5A product note: Specifying Calibration Standards for the HP8510
Network Analyzer
Enter the calkit definitions
On the 8510 network analyzer, press hardkey CAL, the select softkey MORE, MODIFY1 (or
2),
Note: on the 8753E, press hardkey CAL, then select softkey CAL KIT, MODIFY,
In a first step, we define the calibration standards.
Press DEFINE STANDARD
Press 1 and then X1, to enter the 1st line of the table below, i.e. the OPEN calibration
standard data
VERY IMPORTANT NOTE: the softkey OPEN has to be underlined now!
If not, repeat the above step: i.e. hit the OPEN softkey, and enter 1 X1,until the
underlined softkey and the entered table row match !!
NOTE: if your table's entries have a different order for the OPEN, SHORT, LOAD and
THRU,
e.g. SHORT-OPEN-LOAD-THRU, apply this method correspondingly!
Select OPEN
Hit, CO and enter the new C value. Do the same for C1, C2 and C3
Hit SPECIFY OFFSET, and enter the new value.
Do the same with softkeys OFFSET DELAY, OFFSET LOSS, OFFSET Z0,
MINIMUM FREQUENCY, MAXIMUM FREQUENCY
After that, finish this submenu by selecting STD OFFSET DONE
Select LABEL STD (OPEN is displayed in the upper left display corner), and enter the
standard
name as given again in the table below, then select TITLE DONE to finish this sub-menu
Select STD DONE (DEFINED)
Standard no.1 has now been defined. To define the remaining standards, refer to the
calkit table
above and repeat the steps above from DEFINE STANDARD to STD DONE
-> for the SHORT (press 2 and X1 and select SHORT etc),
-> for the LOAD (press 3 and X1 and select LOAD etc.)
-> and for the THRU (press 4 and X1 and select THRU etc.)
The next, i.e. the 2nd step for entering the calkit data, is the entering of the class
assignments (in this step we associate the location of the correction factors in the NWA
corr.vectors with the table entries from above)
Proceed as follows:
Select SPECIFY CLASS,

hit S11A, enter 1, x1 (to inform the NWA to use standard no.1 for the S11A class of
calibration)
hit S11B, enter 2, x1
hit S11C, enter 3, x1
hit S22A, enter 1, x1
hit S22B, enter 2, x1
hit S22C, enter 3, x1
hit MORE,
hit FWD.TRANS, enter 4, x1
hit REV.TRANS, enter 4, x1
hit FWD.MATCH, enter 4, x1
hit REV.MATCH, enter 4, x1
hit RESPONSE, enter 4, x1
hit FWD.ISOL'N, enter 3, x1
hit REV.ISOL'N, enter 3, x1
Select SPECIFY CLASS DONE
The 3rd step of our calkit entering procedure is to label the class
Select LABEL CLASS,
hit S11A, enter the title 'OPEN', press LABEL DONE
hit S11B, enter 'SHORT' (use the wheel to select the characters!!!)
hit S11C, enter 'LOAD'
hit S22A, enter 'OPEN'
hit S22B, enter 'SHORT'
hit S22C, enter 'LOAD'
hit MORE
hit FWD.TRANS, enter 'THRU'
hit REV.TRANS, enter 'THRU'
169
hit FWD.MATCH, enter 'THRU'
hit REV.MATCH, enter 'THRU'
hit FWD.ISOL'N, enter 'ISOLATION'
hit REV.ISOL'N, enter 'ISOLATION'
Select LABEL CLASS DONE
Finally, we reach step 4 of our procedure: the labelling of the calkit:

Select LABEL KIT,
hit ERASE TITLE, and enter the new title,
press TITLE DONE,
press KIT DONE (MODIFIED)
The calkit is now ready to be stored: Hit 'Save'.

Entering TRL Calkit Data Agilent NetworkAnalyzer
specific
Select the corresponding Standard Type:
THRU
REFLECT: use a SHORT (recommended, alternatively use an OPEN)
LINES (in the Menu depicted below, select THRU for entering the different TRL lines,
and then change the name to LINE in the up-popping next menu.)
Note
It is absolutely mandatory to select in the menu below the right standard type, since this selection will
specify internally in the cal.software what type of standard it is.
Fill in the calkit data.
Note
For the LINES, change the Label to LINEx.
170
After having entered the TRL standards (as mentioned, it is strongly recommended to use
a SHORT rather than an OPEN), select 'Class Assignments' as TRL. Then hit 'Edit'.
Click through the 'Calibration Kit Klass' readio boxes and associate the Calkit Standards
with these classes.
Then, for the Calibration Reference Z0, select SystemZ0.
For the 'Testport Reference Plane', select THRU standard (this will mike the cal reference
plane to the middle of the THRU).
171
Finally, save the Calkit on the PNA, typically in the default directory.
NWA Calibration in General
VNA Standard Calibration Techniques and Verification
A considerable challenge in S-parameter VNA measurements is to define exactly where
the measurement system ends and the DUT begins. This location is called 'reference
plane'. This means, all error contributions, inside the VNA and also from the cables up to
this reference plane, have to be calibrated out.
The calibration of a NWA is performed by rather complex procedures. Such are Short-
Open-Load-Through (SOLT), Through-Reflection-Load (TRL) or Load-Refection-Match
(LRM) and the associated error correction model.
As can be seen in the following figure, there are 6 error contribution terms in forward
direction:
Directivity Cross-talk of the power coupler

Crosstalk Cross-talk inside the S-parameter test set, overlying the DUT
Source Mismatch Multiple reflections due to non-Z0 input and output
impedance of the cables and connectors
Load Mismatch The same for the opposite port
Reflection Tracking A/R Frequency dependence of signal path R->A
Transmission Tracking A/R Same for signal path R->B
For the reverse calibration, another 6 error terms add up to a total of 12 terms. These 12
terms are often referred as the 12-term error correction.
For the different calibration procedures, specific, accurately known standards have to be
measured. For connectorized DUTs, with e.g. 3.5mm connectors, usually 4 calibration
standards like OPEN, LOAD, SHORT and THRU are applied, with exactly known standard
inaccuracies.
It is interesting to note that each corrected S-Parameter is a function of all other
(uncorrected) measured S-Parameters, see the following figure:
172
For the case of on-wafer measurements, SOLT calibration, and using Ground-Signal-
Ground (GSG) coplanar probes, fig. 3 depicts the corresponding test structures, which are
usually available on a RF-high-performance ceramic substrate (ISS substrate).
About Calibration Standards

While in the case of the CV meter, the OPEN calibration corrects for an ideal offset
capacitor, a NWA calibration uses cal standards (OPEN, SHORT, LOAD, THRU etc.) from a
cal kit.
These cal standards do not represent ideal standards. They represent the real, existing
standard, including its nonidealities! It means that a SHORT is not an ideal SHORT, but
instead represents rather an inductance. The same applies to the THRU, which has a non-
ideal delay time. The OPEN corresponds rather to a capacitor than to an ideal OPEN.
Therefore, these nonidealities of the G-S-G probes have to be entered into the NWA
before calibration. This is called 'modifying the cal kit'.
While this procedure refers to the nonidealities of the calibration standards, the
subsequent calibration is related to the selected frequency range, the RF power, the
averaging of the NWA etc. After it has been performed, the correction terms are stored in
the cal set of the NWA. In other words, the 12-term error vectors are 'filled up'.
Afterwards, when the measurement is performed, the raw measured data arrays will be
corrected using a correction technique related to the selected calibration method, and
referring to the specified cal set. Finally, this corrected measurement result is transferred
into IC-CAP and displayed there (as well as on the NWA monitor).
Therefore, after the calibration, a re-measurement of the OPEN will not represent an ideal
173
open, but instead exactly those parasitic components as described in the documentation
of the OPEN. In the same way, a THROUGH shows up after calibration with its real delay
time, and a SHORT represents its inductive behavior!
This re-measuring of the standards together with a simulation of their known values can
be used to verify the VNA calibration, and this can be done easily, using IC-CAP.
We simply re-measure the cal.standards, e.g. the OPEN, the SHORT, the THRU and the
LOAD. As said above, this measurement will correspond to the nonidealities of the
selected cal.standard. In case of Cascade probes, the OPEN, for example, behaves like a
negative capacitance of roughly -9fF . Now, after this measurement has been made, we
can define a test circuit for that setup in IC-CAP, entering the netlist of the calibration
standards. Using SPICE3, a simulator which also permits negative capacitances, we can
simulate the expected behavior of the OPEN probes. If the calibration was executed
correctly, there is an excellent match between measured and simulated curves.
In a next step, we measure the SHORT, define in IC-CAP the SHORT nonidealities in a
SPICE circuit, and simulate. Again, an excellent match between measured SHORT data
and simulations has to be achieved. We then continue with the THRU and LOAD
measurements and simulations. Only if all 4 standards exhibits an excellent fit, we can
assume a correct calibration of the NWA.
Note
NWA Calibration Procedure For At-Temperature

Measurements
Example using WinCal of Cascade Microtech
1. Place the wafer on the thermal chuck and the ISS substrate on the aux chuck
2. Pre-align the wafer in theta and use the WinCal software to align the ISS
3. Set desired temperature
4. Once at temperature, put the probes just on the wafer pad and allow to stabilise at
temperature for 15 mins.
5. Move the probes over to the ISS and place on the alignment mark
6. Perform Auto calibration (probe alignment and calibration should be done as fast as
possible before probe temperature changes)
7. Move probes back to device and place probe back on device pads
8. Leave for 5 minutes to re-stabilize
9. Use the WinCal verification tool to check calibration stability
10. Measure the device
NWA Calibration Verification
In order to verify the calibration, it is highly recommended to re-measure the calibration
standards and to model them, using the calkit data of the GSG probe or the connectorized
standards manufacturer.
For this verification, and for an on-wafer SOLT calibration for example, we re-measure the
cal.standards, e.g. the OPEN, the SHORT, the THRU and the LOAD. We know that this
measurement will correspond to the nonidealities of the selected cal.standard. In case of
Cascade probes, the OPEN, for example, behaves like a negative capacitance of roughly -
9fF . Now, after this measurement has been made, we can define a test circuit for that
Setup in IC-CAP, entering the netlist of this calibration standards. Using SPICE3, a
simulator which also permits negative capacitances, we can simulate the expected
behavior of the OPEN probes. If the calibration was executed correctly, there is an
excellent match between measured and simulated curves.
Note
The following figures give an example for Cascade G-S-G probes, 100um pitch.
Verifying the calibration by modeling all the standards of the calibration kit
174
Only if the fitting between simulated and measured data is in the few-percent range, for
all 4 re-measured calibration standards, the NWA calibration can be considered as good. If
only one fit is bad, re-perform a new NWA calibration.
HINT:
use the IC-CAP CAL_VERIFY_MASTER_demodata.mdl File in the demo_features directory.
The scheme of the IC-CAP CAL_VERIFY_MASTER.mdl file
including checking of max. RF power, and the calibration verification of the standards.
Additionally, DC Setups are added for each DUT to allow verification of good DC contacts
during calibration.
NWA Instrument Options in IC-CAP
Instrument Options for the HP/Agilent 87xx and 8510 Network
Analyzers
Usually, for transistors, these settings might be correct:
Note
How to find the max. applicable power is described in Chapter 'CHECKING THE MAX. APPLICABLE RF
SIGNAL LEVEL' of the ModelingHandbook.
Source Power -10dBm

Port 1 Attenuation -20dB
Port 2 Attenuation 0dB
Avg Factor 32
Cal Set No THE CALSET NUMBER USED FOR THE SPECIFIC MEASUREMENT SETUP
For passive devices, typical values are:
Source Power -10dBm

Instrument options for the Agilent PNA Network Analyzers
Setting the Port Power Levels is the main difference between the Instrument Options of
the PNA series compared to the older Agilent/HP network analyzers (87xx or 8510):
Set the entry 'Port Power Coupled' to 'No'. This gives an individual, fully decoupled and
175
independent 'Port Src Power' and 'Port Src Atten' for each port ! I.e. select an individual
combination of 'Port Src Power' and 'Port Src Atten' for each port which corresponds to the
max. applicable power rate of your device.
Note
How to find the max. applicable power is described in Chapter 'CHECKING THE MAX. APPLICABLE RF
SIGNAL LEVEL'
of the ModelingHandbook.
Example of a typical setting for transistors:
Port 1: Power -20dBm, Attenuation 10dB,

Port 2: Power -10dBm, Attenuation 0dB
Note
For each port of the PNA, you can achieve any specified, desired power level, however, not all
combinations of 'Port Source Power' and Port Source Attenuation' and 'Power Slope' are possible (error
message 'Source unleveled')
Therefore, it is recommended to first try manually the possible 'Port Src Power' and 'Port
Src Atten' settings directly at the PNA, and to enter these values afterwards into the
Instrument Options in IC-CAP, (before clicking on 'Calibrate' in the IC-CAP Setup).
Instead of Cal Set numbers, the PNAs offer to save the Cal Sets with filenames in the
memory of the PNA processor. Therefore, save the Cal Set in the default directory of the
PNA and give it a filename. Then, enter that filename into the IC-CAP Instrument Options
Field 'Cal File Name'.
Available IC-CAP Instrument Options for the PNA series (IC-CAP 2006B Update3)
In details (from the IC-CAP Help File):

YELLOW: must-enter-fields (if Use PNA Calibration Settings=yes).
HINT: leave non-yellow fields default
176
Entry Field Note Default Typical
Use User Yes = use user sweep, i.e. IC-CAP triggers every frequency point. No -
Sweep No = use instrument's internal sweep.
Hold Time Time, in seconds, the instrument waits before beginning a sweep to 0 -
allow for dc settling.
Delay Time Time the instrument waits before setting each frequency in user 0 -
sweep mode.
Sweep Time Time the instrument takes for each sweep. 0 = Auto 0 -
Sweep S = Stepped mode. A = Analog (ramp) mode S -
Type[SA
Port Power Yes = Coupled mode. No = Non-Coupled mode. When Ports are Yes No
Coupled coupled, Port Src Power is used for both Port 1 and 2. Port 2 Src
Power is ignored. Attenuators are also coupled so that Port Src Atten
is used for both ports and Port 2 Src Atten is ignored.
Port Src Define the source Power for Port 1 and 2 when ports are coupled or - -20
Power the source power for Port 1 when ports are uncoupled. The power (Transistor)
range depends on the attenuator settings and the PNA model and -10
options. (pass.dev.)
Port 2 Src Define the source power for Port 2 when ports are uncoupled. This - -10
Power option field is ignored when ports are coupled. The power range (Transistor
depends on the attenuator settings and the PNA model and options. and
pass.dev.)
Port Atten Yes = Auto mode. No = Non-Auto mode. When attenuators are in No -
Auto auto-mode, the PNA will set the most efficient values for the
attenuators to obtain the requested output power at the port. In auto-
mode, the full power range is directly available at the output port. In
auto-mode, the instrument options Port Src Atten and Port 2 Src
Atten are ignored.
Port Src Possible Values: 0, 10, 20, 30, 40, 50, 60, 70 dB. The available range 0 5 - 10
Atten depends on the PNA model. For example, the E8364A attenuator (Transistor)
range is 0-60 dB. This option is ignored when attenuators are in auto- 0 (pass.dev.)
mode.
Port 2 Src Possible Values: 0, 10, 20, 30, 40, 50, 60, 70 dB . The available range 0 0
Atten depends on the PNA model. For example, the E8364A attenuator
range is 0-60 dB. This option is ignored when attenuators are in auto-
mode.
Power Slope Can be any value between -2 and +2 dB/GHz 0 -
Dwell Time * Sets the dwell time, in seconds, between each sweep point. Only 0 -
available in Stepped sweep type. (Auto - PNA will minimize dwell
time)
IF Bandwidth Possible Values: 1, 2, 3, 5, 7, 10, 15, 20, 30, 50, 70, 100, 150, 200, 1000 200
300, 500, 700, 1k, 1.5k, 2k, 3k, 5k, 7k, 10k, 15k, 20k, 30k, 35k, 40k
Hz
Note: If a non valid value is specified, the PNA will not round it to the
nearest available value, rather it will use the lowest valid value close
to specified value.
Avg Factor Number of averages per measurement. 1-1024 1 1
Cal TypeHN H = Hardware calibration. N H
N = No calibration
Cal File Name of PNA cal file to be used. If no file extension is specified, .cst myCal.cst -
Name (cal file & instrument state) is assumed.
Use PNA When set to 'yes', IC-CAP will not set the PNA InstrumentOptions, and No -
Calibration will accept all user-specified (or WinCal-specified) settings. All
Settings InstrumentOptions entries are ignored by IC-CAP, except 'Cal File
Name' and 'Cal Type'. As a user, you only need to set the Start-Stop-
Step in your 'freq' Input accordingly to the PNA frequency setting.
Delay for For long-running measurements use this option to avoid 0 -
timeouts measurement timeouts.
Init Command field to set the instrument to a mode not supported by the - -
Command option table. Command is sent at the end of instrument initialization
for each measurement. Normal C escape characters such as \n (new
line) are available.
NOTE on DWELL time:
Dwell time is the time spent at each frequency point before sampling starts. For most
applications, you should set dwell time to auto mode. In auto mode, the PNA increases the
dwell time as the sweep time increases to comply the total sweep time. If long delays are
present in the circuit and additional settling time is needed, set the dwell time to an
appropriate value.
Dwell time is not active in analog mode--only in step mode. If the sweep time in analog
mode is increased significantly (because of a setting), the PNA can internally switch to
step mode and set an optimum value for the dwell time.
NOTE on ATTENUATORS IN AUTO MODE:

When port attenuators are set to auto mode, the PNA automatically chooses the
attenuator value that provides the requested power level at the output port. Since
accurate S-parameter calibration requires that the attenuator settings do not change
during measurements or calibration, auto mode is not recommended.
NOTE on SWEEP TYPE:

Step sweep mode is more accurate than analog (ramp) mode, but analog mode is usually
faster than step sweep mode. In step sweep mode, RF phase locking is performed at each
frequency, which ensures that the frequency value is very accurate. This results in a
longer transition time from one frequency point to the next and a longer total sweep time.
In analog mode, the RF frequency is swept across the frequency range and its frequency
accuracy depends on the linearity of the VCO (Voltage Controlled Oscillator).
177
NOTE on Cal TypeHN :
-> NEVER use 'N' (no calibration)
NOTE on Use PNA Calibration Settings:

If you prefer to set your PNA manually (or use e.g. WinCal for all PNA settings), then set
Use PNA Calibration Settings=Yes
In this case, IC-CAP will accept all settings of the PNA and will no longer be 'the boss' of
the PNA instrument (what is the default setting of IC-CAP). You only have to enter the
frequency settings and the .cst file correctly.
NOTE: when set to 'Yes', all irrelevant InstrumentOptions fields should commonly be
greyed-out (inactive). This is currently not possible for the IC-CAP source code, but this
will be changed in ICCAP2008.
IMPORTANT NOTE ABOUT AVERAGING AND PNAs:
It is often preferable to apply some signal-to-noise algorithm to measured data. In the

case of the 8510 network analyzer, the general practice is to set averaging to ~16 or ~32.
(the added benefit of setting it to higher values is generally negligible).
In the newer PNAs, however, the recommended use paradigm to achieve the same benefit
as averaging in the 8510, is to define IF Bandwidth, rather than averaging. The value
setting is up to the user but generally values of 200 Hz or less are recommended for
modeling applications.
In other words:
reducing IF Bandwidth in PNA is ~ increasing averages in the older 8510...87xx NWAs.
Background: the PNA apply digital signal processing, while the older 8510 and 87xx
applied analog signal processing.
Another reason that the 'best practices' recommendation to set IF Bandwidth vs.
averaging to reduce the noise, is measurement efficiency. If one sets 'Number of
Averages' to 'n', this forces the PNA to take 'n' sweeps (vs. by setting IF Bandwidth, the
time for data acquisition at each point may be increased, but there is only 1 sweep).
IMPORTANT NOTE: if Cascade/Wincal is being used to automate the PNA calibration,

averaging is not saved with the PNA calibration file, as Cascade recommends the use of IF
Bandwidth settings.
So even if one had set averaging > 1 in the IC-CAP Setup Instrument Options, this
averaging would not be used or saved in the PNA cal file.
Performing the NWA Calibration from IC-CAP
NWA CALIBRATION UNDER IC-CAP
Provided that the Cal Kit Data which is provided with every GSG probes calibration
substrate have been entered correctly into the network analyzer (see the other chapter
(.pdf file) located in the same directory), and we have determined the maximum
applicable RF signal level, the calibration of the NWA under IC-CAP is pretty simple.
The steps are:

After the IC-CAP Setup has been defined with a frequency sweep under 'Inputs' and an S-
parameter output, the next step is to define the NWA options under 'Instrument Options'.
Enter particularly the port power levels, the averaging and the Cal Set number (for the
PNAs the Cal Set Filename).
Make sure to select an appropriate RF power level, in order to not overdrive the DUT.
THE NWA IS A LINEAR MEASUREMENT SYSTEM,

THEREFORE HARMONICS (OVERDRIVEN DUTs) ARE NOT ALLOWED !
see the chapter on evaluating the correct RF signal level
in the other chapter (.pdf file) located in the same directory.
H3. Using Agilent PNAs
1. In the IC-CAP Setup's InstrumentOptions, when you set 'Use PNA Calibration
Settings' = Y, then you only need to enter the field for CalType = H, and
Cal/StateFileName.cst only . All other InstrumentOptions fields are ignored by IC-
CAP.
2. Then, click 'Calibrate' in the pulldown menu of the actual IC-CAP setup.
IC-CAP transfers all instrument settings to the NWA.
3. Then, a box appears with a note:
4. Press OK.
178
Warning
Do not click ok now.
5. Perform the calibration manually on the NWA first or use WinCal or any other cal.
software.
6. Save your calibration in the Calset specified earlier in the IC-CAP instrument options
(see below the infos about older HP NWAs and the Agilent PNA series).
7. After all that, finally click OK in IC-CAP.
Series 8510x and 87xx Network Analyzers of HP / Agilent
1. On the NWA, click 'local', 'Cal', and select the appropriate calkit.
2. Click 'Calibrate Menu', and then click 'Full 2-port'.
3. Perform that full TwoPort calibration manually.
4. Finally, save the calset in the same calset number which was entered into IC-CAP
before (under instrument options).
AGILENT PNA NETWORK ANALYZER SERIES
Note
For each port of the PNA, you can achieve any specified, desired power level, however,
not all combinations of 'Port Source Power' and Port Source Attenuation' and 'Power Slope'
are possible (error message 'Source unleveled'){info)
Therefore, it is recommended to first try manually the possible 'Port Src Power' and 'Port
Src Atten' settings directly at the PNA, and to enter these values afterwards into the
Instrument Options in IC-CAP, (before clicking on 'Calibrate' in the IC-CAP Setup).
After you have entered the Instrument Options in IC-CAP, click on Calibrate, wait until you
get the IC-CAP message window as described above, (do NOT click 'ok' on that IC-CAP
window), and go to the PNA. Hit the 'local' button.
With the mouse, select Calibrate/Calibration Wizard,
and you get
Defining Custom Calkits

When measuring packaged devices, the network analyzer calibration is typically performed
using 3.5mm connector-based calibration standards.
As an example, the table below gives the HP85033D 3.5mm calkit data for the HP8753
NWAs,for an 'insertable device' (Delay THRU = 0 ps !!) with a male 3.5mm connector at
one port and a female 3.5mm connector at the other port.
This means, the calibration plane is ending at the 3.5mm connectors.

It has to be extended to the very location of the packaged device, within the test fixture.
179
This is usually done by modeling the test fixture consisting of the other 3.5 mm connector
and the strip line(s) down to the DUT (device under test). A TDR would be applied in order
to evaluate the equivalent circuit along the connections. After that, a NWA would be
applied to find out the crosstalk between the connections, and to add the corresponding
enhancements to the equivalent circuit. After inserting the DUT, modeling continues, and
all model components required from now on for device modeling are part of the 'DUT
model'.
Another possibility would be to de-embed the modeled test fixture, to obtain the DUT S-
parameters out of the measurements of test fixture and DUT.
On the other hand, if custom calkits, i.e. OPEN, SHORT, LOAD and THRU, were available
for the custom test fixture, i.e. and if these custom calkits were accurately known
(modeled), we could also use them instead of the 3.5mm connector-based calkits and
obtain a calibration including the test fixture. After NWA calibration, the calibration plane
would then end where the DUT is connected to, on the fixture !
In other words, we need to specify a table (like the one above) for the custom calkit. This
chapter sketches the realization of this idea.
Before we begin, here a copy of parts of Agilent Application Note AN 1287-9:

In-Fixture Measurements Using Vector Network Analyzers
Characterizing calibration standards for SOLT calibration:

Most network analyzers already contain standard calibration kit definition files that
describe the characteristics of a variety of calibration standards. These calibration kit
definitions usually cover the major types of coaxial connectors used for component and
circuit measurements, for example Type-N, 7 mm, 3.5 mm and 2.4 mm connector-based
calkits.
Most high-performance network analyzers allow the user to modify the definitions of the
calibration standards. This capability is especially important for fixture-based
measurements, because the in-fixture calibration standards will not have the same
attributes as the coaxial standards. Custom calibration standards, such as those used with
fixtures, require the user to characterize the standards and enter the definitions into the
network analyzer. The calibration kit definition must match the actual standards for
accurate measurements. Definitions of the in-fixture calibration standards can be stored in
the analyzer as a custom user-defined calibration kit.
Characterizing a SHORT
The electrical definition of an ideal short is unity reflection with 180 degrees of phase shift.
All of the incident energy is reflected back to the source, perfectly out of phase with the
reference. A simple short circuit from a single conductor to ground makes a good short
standard. For example, the short can be a few vias (plated through holes) to ground at
the end of a micro-strip transmission line. If coplanar transmission lines are used, the
short should go to both ground planes.
To reduce the inductance of the short, avoid excessive length. A good RF ground should
be near the signal trace. If the short is not exactly at the contact plane of the DUT, an
offset length can be entered (in terms of electrical delay) as part of the user-defined
calibration kit.
Characterizing an OPEN
The open standard is typically realized as an unterminated transmission line. Electrical
definition of an ideal open has "unity reflection with no phase shift." The actual model for
the open, however, does have some phase shift due to fringing capacitance.
How to determine open capacitance:
The fringing capacitance can be measured as follows:
1. Perform a one-port calibration at the end of the test cable. Use a connector type that
is compatible with the fixture. For example, use APC 3.5-mm standards for a fixture
using SMA connectors.
2. Connect the fixture and measure the load standard. This data should be stored in
memory and the display changed to "data minus memory". This step subtracts out
the reflection of the fixture connector (assuming good consistency between
connectors), so that we can characterize just the open. An alternative is to use time-
domain gating to remove the effect of the connector.
3. Measure the short standard. Set the port extension to get a flat 180 degrees phase
response. To fine-tune the value of port extension, set the phase-off set value for the
trace to 180 degrees and expand degrees-per-division scale. Mismatch and directivity
reflections may cause a slight ripple, so use your best judgment for determining the
flattest trace, or use marker statistics (set the mean value to zero).
4. Set the network analyzer display format to Smith chart, the marker function to Smith
chart format G+jB (admittance) and then measure the open standard. Markers now
read G+jB instead of the R+jX of an impedance Smith chart. Admittance must be
used because the fringing capacitance is modeled as a shunt element, not a series
element. The fringing capacitance (typically 0.03 to 0.25 pF)can be directly read at
the frequency of interest using a trace marker. At RF, a single capacitance value (C0)
is generally adequate for the calibration kit definition of the open. In some cases, a
single capacitance number may not be adequate, as capacitance can vary with
frequency. This is typically true for the measurements that extend well into the
180
microwave frequency range. Capacitance may vary with frequency, especially at
frequencies above 3 GHz.
When measuring the fringing capacitance, a problem can arise if the short standard is
electrically longer than the open standard. The measured impedance of the open
circuit then appears to be a negative capacitor, indicated by a trace that rotates
backwards (counter-clockwise)on the Smith chart. This problem is a result of using
an electrically longer short standard as the 180 degrees phase reference. The
electrically shorter open will then appear to have positive phase. The remedy for this
is to decrease the port extension until the phase is monotonically negative. The
model for the open will then have a normal (positive)capacitance value. The value of
the negative offset delay that needs to be included in the open standard definition is
simply the amount by which port extension was reduced (for instance, the difference
in the port extension values between the short and the open).In effect, we have now
set the reference plane at the short. Alternatively, the offset delay of the open can be
set to zero, and a small positive offset delay can be added to the model of the short
standard. This will set an effective reference plane at the open.
Characterizing a LOAD
An ideal load reflects none of the incident signal, thereby providing a perfect termination
over a broad frequency range. We can only approximate an ideal load with a real
termination because some reflection always occurs at some frequency, especially with
non-coaxial actual standards.
At RF, we can build a good load using standard surface-mount resistors. Usually, it is
better to use two 100-ohm resistors in parallel instead of a single 50-ohm resistor,
because the parasitic inductance is cut in half. For example, 0805-size SMT resistors have
about 1.2 nH series inductance and 0.2 pF parallel capacitance. Two parallel 100-ohm
0805 resistors have nearly a 20-dB better match than a single 50-ohm resistor. Also note
that special RF-performance SMT resistors are available !
Characterizing a THRU
The thru standard is usually a simple transmission line between two coaxial connectors on
the fixture. A good thru should have minimal mismatch at the connector launches and
maintain a constant impedance over its length (which is generally the case for PCB thrus).
The impedance of the thru should match the impedance of the transmission lines used
with the other standards (all of which should be 50 ohms).
If the PC board is wider for the transmission line where the DUT will be soldered and since
we want the two halves of line to be equal in electrical length to the thru line, the PCB
must be widened by the length of the DUT. With a properly designed PC board fixture, the
short (or open) defines a calibration plane to be in the center of the fixture. This means
the thru will have a length of zero (which is usually not the case for fixtures used in
manufacturing applications, where a set of calibration standards is inserted into a single
fixture). Since the length is zero, we do not have to worry about characterizing the loss of
the thru or its phase shift.
After this prologue of the Application Note, we want to enhance the above mentioned
ideas and apply device modeling techniques for obtaining an accurate model of our
custom calibration standards.
First of all, this is the test fixture layout:
181
In this scenario, the DUT will be measured in a shunt scenario, i.e. towards Ground.
Note
Another test fixture layout would be two non-connected striplines, each of the length of the later OPEN,
SHORT and LOAD cal. standard, leading to both ends of the DUT. The THRU would be exactly twice the
length of the individual OPEN, SHORT and LOAD striplines.
In this case, the DUT would be measured between the NWA ports, and not towards
Ground.
The next figures depict the layout of the corresponding calkit standards:
182
Regarding the calkit standards from above, the modeling has to be performed in such a
way that we can enter the model into the calkit definition table of the NWA, taking the
limitations of the table entries into account.
This means, we need to specify a shifted calibration plane from the connector ends to the
middle of the THRU standard. Therefore, the THRU calkit data will be delay = 0 ps.
For the modeling, we will have to model the two connectors at each side, plus a lossy
delay line. From this model, the left connectors and half of the THRU delay line will be
used later for de-embedding the OPEN, SHORT and LOAD standards. The remaining
performance of these de-embedded standards will be modeled as a frequency-dependent
capacitance for the OPEN, plus possibly a fine-tuning delay (taking into account a possible
different electrical length of the OPEN compared to half the length of the THRU
standard),an inductor for the SHORT standard, plus possibly a fine-tuning delay, a delay
and an impedance Z0 in series with the 50 Ohm resistor for the LOAD standard (taking
into account the frequency-dependence of the 50 Ohm resistor).
STEP-BY-STEP MODELING PROCEDURE FOR THE CUSTOM CALKITS:
Note
Pls. contact [email protected] for a copy of the IC-CAP file shown below.
1. a conventional NWA calibration is performed based on a standard 3.5mm connector

calkit
2. this calibration is verified by re-measuring all 4 cal standards, and by comparing the
4 measurement results by the corresponding 4 SPICE simulations. See the
corresponding calibration verification Model file of IC-CAP, in the demo_features
directory.
After that, the custom calkits are connected and measured individually. For each calkit
standard, a model will been developed. Since we will have to de-embed the connector and
half of the THRU for the OPEN, SHORT and LOAD; what is done by inverse ABC matrix
(chain matrix) multiplications, we will use only ABC matrices for the standard modeling,
and not a simulator like SPICE.
183
CUSTOM THRU CALKIT: Measurement and SPICE Model

As mentioned, we want the new reference plane in the middle of the THRU standard, so
the THRU is modeled first.
The plots above show the measured data and the fit of the model. Particularly important is
the fitting of the Sxy plots, which corresponds to the length of the thru microstrip line and
its loss.
The connectors show basically up in the Sxx plot, as a deviation of the measurement from
the center of the Smith chart, i.e. Z0 = 50 Ohm.
The modeling transforms in IC-CAP look like this:
Modeling:
we commence with the model parameters of the strip line (plus the possibly existing
delta-L of the THRU if the test fixture offers 2 microstrip lines instead of a single one. In
this case, the THRU stripline is Ldelta longer than 2*Lopen.)
The next detail of IC-CAP depicts the PEL program 'S_simul' which simulates the THRU
standard, i.e. the chain of Aconnleft*Aopen*Adeltathru*Aopen*Aconnright
Transform 'Aopen' holds the A matrix data of the microstrip line of the same length as the
OPEN standard, Adeltathru the A matrix data of the possibly existing Ldelta strip line (DUT
between the ports), Aconnleft and Aconnright model the 3.5mm connector with each a
series inductor and a shunt capacitor plus a typ. 38ps delay line of 50 Ohm. Watch out
that Aconnleft and Aconnright contain flipped schematics, since A matrices always model
from 'left to right' !
Parameter fine-tuning is performed for the Sxy plot with optimizer

'o_Sxy_alpha_beta_Z0', which either tunes or optimizes the delay and the attenuation of
the THRU microstrip line. Keep an eye on the starting trace of the Sxx curve: if it ascends
from Z0=50 Ohm, then Z0 of the microstrip line is >50 Ohm, if it descends, then it is <50
Ohm !
Then, the Sxx plot is fine-tuned or optimized with 'o_Sxx_Z0conn_L_C'.

184
Note
the resonance in the Sxx stems from the connectors.
Note
all modeling parameters for this step are Model Variables: stripline: Lopen, alpha, beta, Z0, connector: L,
C, Z0conn, TDconn
(because they will be reused in the modeling of the OPEN, SHORT and LOAD)
but the Ldelta microstrip parameters alphathru, betathru and Z0thru are local Setup Variables.
To finish with the THRU, we enter its calkit data into the table:
CUSTOM OPEN CALKIT: Measurement and SPICE Model

Now that we know the model of the microstrip line on the calibration substrate, as well as
the connector model, we can measure the OPEN standard, and de-embed the left
connector and the microstrip line from the measurement. After that, we will obtain the
measurement of the open ended line.
In the IC-CAP file, this is done in Setup specify_open/OPEN.
The following IC-CAP screen dump shows the modeling programs:
The simulation program 'S_simul' consists of a multiplication of A matrices of the left

connector, the microstrip of the line (with alpha, beta and Z0 as modeled before in the
THRU standard) and a conversion into S-parameters.
For the modeling, we need to de-embed the measurement from the 3.5mm connector and
the microstrip line.
This is done in Program 'S_wo_line', see the next slide:
These de-embedded data are converted into a corresponding OPEN capacitor to Ground,
185
with the following PEL Program:
The next plot shows this OPEN capacitance, and its frequency-dependency.
Both, the converted measurement data as well as the fitted curve with
With these capacitance values, we check the fitting in the S-parameters, and obtain:
A little fine-tuning of the OPEN delay fits the S11 data very well, as already shown in the
first figure of this chapter.
So, the obtained parameters for the OPEN cal. standard are:
After all, we enter its calkit data into the table:
Custom SHORT calkit: Measurement and SPICE Model

In the same way as with the OPEN standard, we proceed with the SHORT.
186
The measured data are again de-embedded from the 3.5mm connector and the microstrip
line, in the same way as above for the OPEN.
Then, the shorted-end data are converted into a shunt inductance, what is depicted
below:
The inductance shows basically no frequency drift, and so the Cal. standard parameter L0
can be derived directly. A little fine-tuning of this inductor together with an additional
delay line is performed on the data of the total, non-deembedded data in order to take the
electrical difference of the structure with the geometrical structure into account.
So, we obtain these calkit data:
which we enter again into the standards table:
Custom LOAD calkit: Measurement and SPICE Model

In the same way as with the OPEN and SHORT standards, we proceed with the LOAD.
187
The data of the total measurement, as shown above, are de-embedded from the 3.5mm
connector and the microstrip line. The modeling goal is now a Z0 = 50 Ohm shunt
resistor, in series with again a delay line. Since the ohmic 50 Ohm resistor will exhibit a
frequency dependence, this delay line can help modeling this effect.
Also, this resistor should be a real high-frequency resistor, and as mentioned in the
application note from above, it is preferable to have 2* 50 Ohm resistors in parallel to
reduce their inductance.
The data after de-embedding look like this:
In this case, due some non-idealities of the standard, we can only make a compromise.
THis is shown as LOAD_Standard data in the same plot above.
These data are:
We enter its calkit data into the table and obtain the final calkit table result:
NEXT AND FINAL STEP: VERIFICATION OF THE CALKIT

A new NWA calibration is performed with a modified calkit, and the data from the table
above.
After that, all cal standards are re-measured and compared to a simulation of SPICE
circuits, which consist of the components of the calkit table from above. The calibration
was successful only if
ALL 4 RE-MEASUREMENTS MATCH THE SIMULATIONS EXTREMELY WELL !!
Introduction to NWA Calibration

VNA Standard Calibration Techniques And Verification
A considerable challenge in S-parameter VNA measurements is to define exactly where
188
the measurement system ends and the DUT begins. This location is called 'reference
plane'. This means, all error contributions, inside the VNA and also from the cables up to
this reference plane, have to be calibrated out.
The calibration of a NWA is performed by rather complex procedures. Such are Short-
Open-Load-Through (SOLT), Through-Reflection-Load (TRL) or Load-Refection-Match
(LRM) and the associated error correction model.
Figure 1: The 12 error contributions for S-parameter measurements with a VNA
As can be seen in fig.1, there are 6 error contribution terms in forward direction:
Directivity: cross-talk of the power coupler

Crosstalk: cross-talk inside the S-parameter test set, overlying the DUT
Source Mismatch: multiple reflections due to non-Z0 input and output impedance of
the cables and connectors
Load Mismatch: the same for the opposite port
Reflection Tracking A/R: frequency dependence of signal path R->A
Transmission Tracking A/R: same for signal path R->B
For the reverse calibration, another 6 error terms add up to a total of 12 terms. These 12
terms are often referred as the 12-term error correction.
For the different calibration procedures, specific, accurately known standards have to be
measured. For connectorized DUTs, with e.g. 3.5mm connectors, usually 4 calibration
standards like OPEN, LOAD, SHORT and THRU are applied, with exactly known standard
inaccuracies.
It is interesting to note that each corrected S-Parameter is a function of all other
(uncorrected) measured S-Parameters, see fig.2 below.
Figure 2: The 12-term error correction formula
For the case of on-wafer measurements, SOLT calibration, and using Ground-Signal-
Ground (GSG) coplanar probes, fig. 3 depicts the corresponding test structures, which are
usually available on a RF-high-performance ceramic substrate (ISS substrate).
Figure 3: SOLT calibration structures for on-wafer measurements
189
Some notes about these calibration standards:

while in the case of the CV meter, the OPEN calibration corrects for an ideal offset
capacitor, a NWA calibration uses cal standards (OPEN, SHORT, LOAD, THRU etc.) from a
cal kit.
These cal standards do not represent ideal standards. They represent the real, existing
standard, including its non-idealizations! It means that a SHORT is not an ideal SHORT,
but instead represents rather an inductance. The same applies to the THRU, which has a
non-ideal delay time. The OPEN corresponds rather to a capacitor than to an ideal OPEN.
Therefore, these non-idealizations of the G-S-G probes have to be entered into the NWA
before calibration. This is called 'modifying the cal kit'.
While this procedure refers to the non-idealizations of the calibration standards, the
subsequent calibration is related to the selected frequency range, the RF power, the
averaging of the NWA etc. After it has been performed, the correction terms are stored in
the cal set of the NWA. In other words, the 12-term error vectors are 'filled up'.
Afterward when the measurement is performed, the raw measured data arrays will be
corrected using a correction technique related to the selected calibration method, and
referring to the specified cal set. Finally, this corrected measurement result is transferred
into IC-CAP and displayed there (as well as on the NWA monitor).
Therefore, after the calibration, a re-measurement of the OPEN will not represent an ideal
open, but instead exactly those parasitic components as described in the documentation
of the OPEN. In the same way, a THROUGH shows up after calibration with its real delay
time, and a SHORT represents its inductive behavior!
This re-measuring of the standards together with a simulation of their known values can
be used to verify the VNA calibration, and this can be done easily, using IC-CAP.
We simply re-measure the cal.standards, e.g. the OPEN, the SHORT, the THRU and the
LOAD. As said above, this measurement will correspond to the non-idealizations of the
selected cal.standard. In case of Cascade probes, the OPEN, for example, behaves like a
negative capacitance of roughly -9fF . Now, after this measurement has been made, we
can define a test circuit for that setup in IC-CAP, entering the netlist of the calibration
standards. Using SPICE3, a simulator which also permits negative capacitance, we can
simulate the expected behavior of the OPEN probes. If the calibration was executed
correctly, there is an excellent match between measured and simulated curves.
Note
NWA Calibration With Gating in the Time Domain

This chapter refers to a de-embedding technique for strip-line test fixtures as described by
/Gronau/. Based on this publication, see also /Bhm/ and /Stassen/.
It is assumed for this type of de-embedding that the vector network analyzer (VNA or
NWA) has been calibrated already up to the ends of its connection cables. These cables
are then connected to a test fixture. The idea is that the DUT is measured on a ceramic
substrate (or similar) and is connected by strip lines on this test fixture. If we can correct
the VNA calibration for the connector and the strip lines at both sides of the DUT, we
would be able to shift our calibration plane to every location on the strip lines. And
therefore we can de-embed the effects of the test fixture until the physical limits of the
DUT. This is a very important feature and allows to measure devices like dummy
packages, passive components like resistors, capacitors, spiral inductors and vias alone
190
without being affected by parasitic effects from the test fixture.
The measurement setup for the DUT and the test fixture is given in following figure.
The measurement environment for the DUT
The basic idea of this special calibration technique is to describe the effects of the
connectors A and B as well as those of the lines A and B with A-matrices each and to de-
embed the DUT later with A-matrix manipulations.
Therefore, the A-matrix of the total VNA measurement is:
ATOTAL = Aconn_a * Aline_a * ADUT * Aline_b * Aconn_b (1)
Once the matrices Aconn_x and Aline_x are know, we can solve for ADUT :
ADUT = Aline_a-1 * Aconn_a-1 * ATOTAL * Aconn_b-1 * Aline_b-1 (2)
The problem is now to determine the A-matrices of the strip lines and the connectors. This
is done with measurements on an auxiliary calibration substrate that holds a THROUGH
line and two OPEN strip lines with different length.
Calibration Substrate
The VNA measurements are Fourier transformed to the time domain in the VNA, a gating
function is applied to select either the connector reflection or the end reflection, and finally
this gated time domain reflectogram is transformed back to the frequency domain. The
resulting S-parameters are then transferred back to IC-CAP. These auxiliary measurement
results are stored in the setups of the model file and the de-embedding calculations will
refer to these data later. That means we will perform a "software calibration" on the later
measurement data. In other words, we will obtain partly calibrated data (until the ends of
the VNA cables) from the device under test (DUT), and then perform matrix manipulations
on this data using the previously measured calibration data. After these manipulations
within IC-CAP, we will obtain the S-parameters of the DUT alone and convert them to A-
parameter to match equ.(2).
The following figure gives an overview of this concept of user-definable reference planes
and the signal flow chart.
Test fixture with user-definable reference planes and the signal flow chart for the fixture.
191
In order to be able to set a correct gating, i.e. a high time domain resolution, it is
necessary to sweep the VNA frequency as high as possible. In the file
'cal_w_td_gating.mdl' , the upper frequency is 26.5GHz. Therefore, an HP8510 or similar
with time domain option is recommended. Also, macro "run_cal" (that performs all
required measurements of the setups Sxxx) is using HP-IB commands for the HP 8510.
The following figure gives the user interface of the IC-CAP file 'cal_w_td_gating.mdl' :
The IC-CAP model file includes:
setups 'cal / Sxxx' are used to keep the gated S-parameters

setup 'cal / Calculate' contains all the matrix manipulations
setup 'cal / m_thru_line' holds the un-gated measurement of L1
192
setup 'meas_DUT / meas_DUT' contains the measurement of the DUT and the test
fixture, its transforms are linked to setup 'cal / Calculate' in order to perform the de-
embedding.
The individual steps of the procedure are:

The S-matrices of the connectors A and B are called matrix A and matrix B and the strip
lines A and B are described with the propagation coefficient , i.e. exp(- * length).
It is further assumed that

A12 = A21 = A (3)
and B12 = B21 = B (4)
To begin with, the VNA has to be calibrated to the end of its connection cables, and
the calibration has to be stored in a VNA cal set.
Then we connect the calibration substrate THRU line L1 and have the following
situation:
Determination of S11 of connectors A and B with time-domain gating.
Note
Using IC-CAP, the VNA is not re-setted totally before every measurement. Therefore, we can perform a
'classical' S-parameter measurement from IC-CAP, set the VNA to local afterwards, transform the S-
parameters in the VNA to time domain, and set the gating. If we then re-perform a measurement from IC-
CAP, the S-parameters transmitted to IC-CAP will reflect the gated S-parameters.
Next,
The transmission line of length L1 is measured. L1 has to be large enough, such that
the individual reflections of connector A and B do not overlap in the time domain.
i.e.: we set the VNA to local, perform a measurement, display S11, transform it to
the time domain and set the gate on the left reflection (connector A). Then, we re-
perform a measurement in IC-CAP. After that, we hold in
SETUP S11A the gated S-parameter measurement S11 of connector A.
S11A = A11 (5)
This is depicted in the following figure:
The gating of S11 for the separation of S11A, the S11 parameter of connector A
193
After that, we have to
Set the VNA to local, and position the gate on the right reflection (connector B). We
re-perform a measurement in IC-CAP and hold in
SETUP S11B the gated S-parameter measurement S11 of connector B,

what is mathematically:
S11B = A21*exp(- *L1)*B22*exp(- *L1)*A12
= A^2*B22*exp(-2* *L1) (6)
with respect to the signal flow chart.

See the following figure for details.
The gating of S11 for the separation of S11B, the S11 parameter of connector B
Note
index 1 refers always to port 1, index 2 to port 2 of the VNA.
Now we concentrate on S22:
We set the VNA to local, measure and display S22, transform it to the time domain
and set a gate on the left reflection (connector B, the first reflection to be seen in
time domain for S22).
Then, we re-perform a measurement in IC-CAP. After that, we hold in
SETUP S22B the gated S-parameter measurement S22 of connector B.
S22B = B11 (7)
194
The following two figures give the details.
The measurement of S22A and S22B with time domain gating.
The gating of S22 for the separation of S22B, the S22 parameter of connector B.
Then,
Measuring again L1 with the VNA set to local, we now position the gate on the right
reflection (connector A). We re-perform a measurement in IC-CAP and hold in
SETUP S22A the gated S-parameter measurement S22 of connector A,

what is mathematically:
S22A = B^2*A22*exp(-2* *L1) (8)
The following figure shows this procedure:
The gating of S22 for the separation of S22A, the S22 parameter of connector A.
Lastly,
we position the gate on the strip line in the time-domain equivalent of S21 and yield
the measurement result of
SETUP S21L1,
S21L1 = S12L1 = A*B*exp(- *L1) (9)
This is shown in following figure:
The gating of S21 and the 'gated S-parameter' S21L1.
195
After having measured the THRU line L1, we do not know yet all parameters that are
required for the de-embedding procedure. We still need to know the value of parameter
. And that is why we have to perform some more measurements on lines L2 and L3, which
have each an open end.
Port1 of the VNA is therefore connected with L2 and port2 with L3.
Measurement setup for the first measurement of lines L2 and L3
We
measure S11 and set a gate on the OPEN reflection of line L2 at port1, and transform
back into the frequency domain. After having performed a measurement in IC-CAP,
SETUP S11L2 contains
S11L2 = A^2* exp(-2 *L2) (10)
This is shown in the following figure:
Signal flow chart for line L2 and the gating in the time domain to separate the OPEN of S11L2
we do the same gating on the OPEN reflection of line L3 at port2, and transform back
into the frequency domain. The following IC-CAP measurement yields
SETUP S22L3, which contains
S22L3 = B^2* exp(-2 *L3) (11)
This is depicted in following figure:
Signal flow chart for line L3 and the gating in the time domain to separate the OPEN of S22L3
196
Next, we have to flip the connections of the calibration substrate. Now, port1 of the VNA is
connected with L3 and port2 with L2.
Measurement setup for the second (flipped) measurement of lines L2 and L3
We re-measure with the VNA set to local, then
set a gate on the OPEN reflection of line L3 at port1, and transform back into the
frequency domain. Thus, after having measured with IC-CAP,
S11L3 = A^2* exp(-2 *L3) (12)
See the following figure:
The gating in the time domain to separate the OPEN of S11L3
again, we do
the same gating on the OPEN reflection of line L2 at port2, and transform back into
the frequency domain. Thus,
S22L2=B^2* exp(-2 *L2) (13)
given in the following figure:
The gating in the time domain to separate the OPEN of S22L2
197
After all, we are ready to calculate all unknown S-parameter values:
First, the propagation coefficient of the lines on the substrate is with the assumption of
S11L3 / S11L2 = S22L3 / S22L2 (see literature /Gronau 1992/):
where the programs 'phs' contain phase correction algorithms for a continuous phase shift
without 2PI steps.
The calculated propagation coefficient
The reflection coefficient is:
198
where 'phs' contains again the phase correction algorithms for a continuous phase shift
without 2PI steps.
The calculated reflection coefficient
while,
and
and
and
199
what looks in IC-CAP like:
Now we are done and can calculate the S matrices of the connectors and the strip lines:
These S-matrices are converted to A-matrices using the IC-CAP TwoPort function.
Finally, these A matrices are manipulated following equation (2), what looks like:
tmp1 = (TwoPort(S_lineP1.m,"S","A"))^-1
tmp2 = (TwoPort(S_conn1.m,"S","A"))^-1
tmp3 = TwoPort(S.m,"S","A")
tmp4 = (TwoPort(S_conn2.m,"S","A"))^-1
tmp5 = (TwoPort(S_lineP2.m,"S","A"))^-1
tmp = tmp1 * tmp2 * tmp3 * tmp4 * tmp5
RETURN TwoPort(tmp.m,"A","S")
with
A_total = A_connA * A_LA * A_DUT * A_LB * AconnB (see 1)
or solved for A_DUT:

-1 -1 -1 -1
A_DUT = A_LA * A_connA * A_total * A_connB * A_LB (see 2)
It should be noted, that a nice feature of this method is to being able to set the calibration
plane to every location on the strip line on the ceramic substrate. This means that once
the calibration is done, the performance of the DUT alone can be calculated by simply
setting the strip line length variables to the length of interest. This feature is especially
important for devices like SMT devices, vias etc.
Calibration Verification
In order to check the quality of the calibration, we reconnect LINE_1 of the calibration
substrate, set the lengths L_P1 and L_P2 (length of the strip lines of port1 and port2) to
0.5*L1 and 0.5*L2, perform a measurement of strip line L1 and apply the equations (20)
to the measurement results. In this case, the calibrated or better de-embedded result
should correspond to an ideal THRU: Sxx=0 and Sxy=1 (all power transmitted). Fig.20
shows the result of such a verification measurement for frequencies from 45MHz to
26.5GHz. It can be seen that the procedure works very well.
Verification of the calibration quality
200
Note
It is mandatory for this calibration method to exactly determine the calibration substrate strip line lengths
up to 1/10 mm.
Publications
G.Gronau, Scattering Parameter Measurement of Microstrip Devices, Microwave Journal,
Nov.1992.
G.Gronau, I.Wolff,: A Simple Broad-Band Device De-embedding Method Using an

Automatic Network Analyzer with Time-Domain Option, IEEE Trans.on Microwave Theory
and Tech., Vol. 37, No.3, March 1989.
M.Bhm, Diploma Thesis Technical University Ilmenau/Germany: Metechnische und

rechnerische Untersuchungen von Verbindungsstrukturen fr elektronische Schaltungen in
einem Frequenzbereich von 45MHZ bis 25GHz, 30.4.1993.
Rudolf Stassen, 'Einsatz eines Mikrowellen-SPitzenmeplatzes zur Charakterisierung von

Transistore direkt auf dem Wafer im Frequenzbereich von 0,045 bis 26,5GHz',
Diplomarbeit am Institut fr Schicht- und Ionentechnik am Forschungszentrum Jlich
GmbH,
available from Technische Informationsbibliothek, Hannover.
VNA Cal Kit Manager 2.0

Version 2.0 (983 kBytes) for Windows 95 / 98 / NT4 / 2000
from the internet site www.vnahelp.com
This PC-based software allows the easy entering of calkit data into a NWA, without hitting
manually the many buttons especially on the elder NWAs.
201
Installation Hints:
The file setup.exe should be stored in any convenient spot on your hard drive. Then
execute the file by double-clicking it in Windows Explorer - this will start the installation
program. You can then follow the instructions on the screen.
VNA Cal Kit Manager is absolutely free, provided you abide by the license agreement.
202
S-Parameters
Contents
Definition of S-Parameters (iccapmhb)

Understanding S-Parameter Plots (iccapmhb)
Interpreting S-Parameter Plots (iccapmhb)
S-Parameter Basics for Modeling Engineers (iccapmhb)
S-Parameters Measurements for Modeling (iccapmhb)
Calculating S-parameters from Voltages (iccapmhb)
Small Signal Versus Large Signal S-Parameters (iccapmhb)
Converting N-Port S-Parameters to M-Port (iccapmhb)
Differential S-Parameters (iccapmhb)
Amplifier Characteristics from S-Parameters (iccapmhb)
Q-Factor Calculation (iccapmhb)
S-Parameter Impedance Matching (iccapmhb)
S-Parameter Utilities (iccapmhb)
Note
Additional Information: Agilent Technologies Application Note AN-95
S-Parameter Techniqueshttp://contact.tm.agilent.com/Agilent/tmo/an-95-1/index.html
203
Amplifier Characteristics from S-
Parameters
We can calculate more information about the 2-port network from the S-parameters. Here
some examples. For more details see the publication list below, esp. Hewlett-Packard
application note AN95-1.
Input Reflection Coefficient with arbitrary load

impedance Z L
with the reflection coefficient:
Note
For Z L = Z 0, it is S' 11 = S11
Output Reflection Coefficient with arbitrary source

impedance Z S
Voltage Gain with arbitrary Z S and Z L
Stability:
Rollett's stability factor, calculated from the S-parameters, is given by
where D is the determinant, i.e.
Devices with a K-factor K > 1 and a non-negative determinant at a given frequency are
said to be unconditionally stable at that frequency, meaning they will not oscillate due to
load mismatch. In this case, ANY impedance connected to the ports of the two-port will
NOT cause oscillation.
In general, transistors have K < 1 for certain frequencies, particularly lower frequencies,
and are therefore only stable for certain impedance values.
Note
K can also be calculated from Y parameters.
Power Gain from S-Parameters

With the abbreviations:
where S means conjugate complex S-Parameters, we can calculate the following gain
characteristics:
204
Power Gain
of the load impedance Z L
Available Power Gain
with the reflection coefficient,
of the source impedance Z S
Transducer Power Gain
in the case of unilateral transducer power gain, i.e. S12=0, this simplifies to:
Maximum Available (stable) power amplification Gain

This maximum gain is obtained when the transistor or amplifier is inserted between
matching load/source resistance. I.e. for conjugate complex impedance matching at both
sides of the transistor, the amplifier gain is different, usually higher than for a Z0 = 50?
environment..
The maximum available gain is:
with K from (4) and only valid if K>1
MAG calculates the gain of the entire network (DUT plus matching networks). There is still
a Z0 termination outside the matching networks.
Note
It is a typical characteristic of RF transistors that they are conditionally unstable for lower frequencies
(K<1). Calculating MAG for those frequencies makes therefore no sense. In other words, only for
frequencies with K>1, an impedance matching for max. gain MAG can be designed.
It is further helpful to draw |S21|2 and MAG together in the same plot versus log(freq).
Usually, MAG is bigger than |S21|2, what means that the power amplification can be
improved with impedance matching, compared to the conventional measurement
condition Z0=50. See the following figure. Note that for a wide frequency range, the
transistor is not unconditionally stable (K<1), and therefore MAG is not defined.
205
Comparing power amplification |S21|2 to the maximum available gain.
Note
For modeling, it is interesting to compare MAG_measured and MAG_simulated, since basically all S-
parameters contribute to the MAG formula. A good fitting for MAG is an indication for a good model fit.
In case that S 12 =0, the MAG formula from (10) simplifies and we get for the
Maximum Unilateral Transducer Power Gain
When the transistor/amplifier being measured by a conventional vector network analyzer,

i.e. in a Z 0 environment, the Current amplification for AC-wise shorted output port is
This gives the Power Amplification in dB

for this Z 0 environment condition:
Mason's power gain

This is another often cited power gain definition:
with
as in equation (4) above, and
Definition of the Transit Frequency

The transit frequency ft is defined as the frequency at which |h21| hits '1'. This frequency
ft determines the max. switching frequency and is therefore most interesting for digital
applications (output AC-wise shorted).
The definition of the transition frequency assumes, that the transistor behaves like a
single pole lowpass with a -20dB decrease per frequency decade. Looking at real
measurement data shows, that this assumption is true in most cases, even for very high
frequencies.
As we assume a 1-pole low-pass for h21, the gain-bandwidth product is a constant.

Therefore it is sufficient to measure a h21 at a fixed frequency higher than the -3dB
frequency. In other words, this fixed frequency should be from a -20dB/decade range of
h21 . This measurement frequency can be found when transforming the measured S-
parameters to H-parameters (using the TwoPort function in IC_CAP). From the dB-plot of
ABS(h21(f)) versus log(frequency) we determine a frequency where the slope fits a -
20dB/decade roll-off.
206
This frequency is now used as a fixed frequency f-20dB. After the S-parameters are
converted into H-parameters, we get for the constant gain-bandwidth product of this one-
pole low-pass filter:
Using Y parameters, ft can also be calculated as:
Especially for MOS Transistors:
(M.T.Yang et.al, On The High-Frequency Characteristics and Model of Bulk

Effect, IEICE Trans. Electron., Vol.E88-C, No.5, May 2005)
A short excerpt from the above literature:

At low Vgs in the weak inversion region, the increased electron (hoe) velocity also reduces
the total transit time of the carriers. these lead to a significant increase of gm, resulting in
a higher fT. For PMOS, the fT increase is less due to less mobility. With increasing Vgs, in
strong inversion, fT reaches a maximum due to velocity saturation. Beyond that, fT
decreases again, due to mobility reduction. Applying more vB shifts the VTH, and thus,
reduces the fT and shifts its maximum to higher vGS .
In practice, the simplest way to see fT is to multiply MAG(H21) by the stimulating
frequency array. This gives a plot as shown below:
This visualizes easily where the prerequisite of the formula presented above is valid:
where the ft-curve is flat vs. frequency. This is marked by the box in the plot above.
The different ft-values, represented by the blue arrows, represent the different DC bias
conditions.
Yet, this is not the max. frequency for the transistor in analog operation and without the
assumption of an AC-wise short at the output of the twoport.
Definition of the Maximum Oscillation Frequency

fmax:
This max. frequency is defined by fmax , which refers to the maximum available gain
MAG. As long as MAG> 1, the transistor can still generate oscillations. The max. possible
oscillation frequency is therefore
or, referring to Manson's power gain:
Although neither U nor MAG are representing a 1-pole low-pass filter, and, therefore, a -
20dB/decade interpretation does not make real sense for MAG. But many modeling
engineers do so, and calculate:
where f equals the measurement frequency at the falling slope (where the slope is
assumed to drop with -20dB/decade).
In J.Berkner, Kompaktmodelle fr Bipolartransistoren, expert Verlag, ISBN 3-8169-2085-
3, Fig. 6-45, a nice sketch about the problems which may arise with this method is given.
See the next figure:
207
As can be seen, neither MAG does not follow a -20dB/decade slope for all frequencies, nor
does U.
However, theoretically, both curves should hit the 'real fmax' at 0dB,
as described by Gupta, M.S.: "Power Gain in Feedback Amplifiers, a Classic Revisited",
IEEE Trans.on Microwave Theory and Techniques, Vol.40, No.5, May 1992. So, applying a
-20dB/decade slope interpretation will result in a fmax which is frequency dependent, see
the sketched -20dB/decade extrapolation. In many cases, the max(fmax) is then
published in data sheets of foundry processes.
Note
This frequency-dependency of fmax has also been presented during the European IC-CAP workshop 2002
in Berlin by P.Brenner, Infineon, Munich, "Test Structures, Methodology And Specific Problems In
Measuring Power Gains And fMAX Characteristics Of High Speed Transistors With fMAX > 50GHz".
For H21 of a 1-pole low-pass filter (a transistor for example), the MAG(H21) decreases by -20dB per
decade in a dB vs. log(freq) plot. In this case, an extrapolation of the -20dB/decade slope by means of:
!spar_amp_characteristics_equation32.gif!like described further above makes sense.
There are approximation formulas published which allow fmax to be calculated

independent of the max. available gain MAG.
For a bipolar transistor, B.Ardouin proposes:
B.Ardouin, Contribution la modlisation et la caractrisation en hautes frquences des

transistors bipolaires heterojunction Si/SiGe, PhD Thesis at the University Bordeaux,
cole doctorale de sciences physiques et de l'ingenieur, Dec.10, 2001
For MOS transistors, G.Knoblinger proposes:
G.Knoblinger, Modellierung des Hochfrequenzverhaltens von MOS-Transistoren, PhD

Thesis at the Universitt der Bundeswehr, Munich, Fakultt Elektrotechnik, Aug.10, 2001
and M.T.Yang et.al:
M.T.Yang et.al, On The High-Frequency Characteristics and Model of Bulk Effect, IEICE
Trans. Electron., Vol.E88-C, No.5, May 2005
Note
Like with the ft Plot, the fmax is calculated from the traces inside the box (where the prerequisites are
valid), and is a function of the applied DC biases.
Notes on Estimating fmax vs fT by Extrapolation, and

Comparison of fT vs fmax Measurements
Source: Labnotes on calculating fmax and ft, with friendly permission from Nathan
Perkins, Agilent Technologies, 2004
208
As expected, |H21| rolls off at one decade of magnitude per decade of frequency.
Therefore, fT (H21 magnitude=1) is always calculable as fT=mag(H21)*freq for any
frequency.
However, |U| rolls off by two decades of magnitude per decade of frequency. This is
because U is a power gain, not a current gain. Therefore, the simple approximation of
fmax=mag(U)*freq would badly overestimate the fmax value !
Linearizing this, we get log(U) = -2*log(freq) from U ~ freq-2. Anywhere along a line of a
specific DC bias, the product U*freq2 is a constant. If any value of U and freq are
measured, and the freq fmax at which U goes to unity is desired, we can calculate:
This leads to a solution/description for plotting fT vs fmax on dB scale graphs in IC-CAP.
Plotting raw values will give H21 a rolloff of -20dB/decade (expected value), but U will roll
off at -40dB/decade. This is because we are inconsistent in our definition of dB.
209
This is because an IC-CAP plot always does dB scaling using 20*log(val) and not
10*log(val); e.g. always assumes a measurement of voltage gain rather than a
measurement of power gain. Re-plotting with U as a power gain term (e.g., entering UdB
as 10*log10(U)), and H21 as a current gain (e.g., 20*log10(MAG(H21)), we see that both
now have a roll-off of -20dB/decade of frequency. This is more consistent with many
literature publications.
Considerations About the Validity of Unilateral Gain

Calculations
All twoport models are bilateral, so both the forward and the reverse signal flow must be
considered. If the signal flow in the reverse direction is much smaller than the flow in the
forward direction, it is possible to make the simplification that the reverse flow is zero.
The Unilateral Figure of Merit U is a quick calculation that can be used to determine where
this simplification can be made without significantly affecting the accuracy of the complete
gain formula.
This gives the error limits on Unilateral Gain Calculations:
Example:
Suppose that a transistor is to be used in a simple amplifier, and optimized for power gain
at a certain frequency by means of lossless input and output matching networks. Since the
reverse gain S12 for this transistor is quite small, -50 dB smaller than forward gain S21 ,
there is a possibility that it can be neglected. If this is so, the design problem will be much
simpler, because setting S12 equal to zero will make the design equations much less
complicated. In determining how much error will be introduced by assuming S12 = 0, the
first step is to calculate the unilateral figure of merit U, using the formula given above.
U is then plotted against log(freq). As an example, if the maximum value of U is 0.03, so

the maximum error in this case turns out to be about + 0.25 dB at that frequency. This is
small enough to justify the assumption that S12 = 0.
Note
A small i.e. neglectible reverse gain, or feedback factor, S12 , is an important and desirable property for a
transistor to have, for reasons other than it simplifies amplifier design. A small feedback factor means that
the input characteristics of the completed amplifier will be independent of the load, and the output will be
independent of the source impedance. In most amplifiers, isolation of source and load is an important
consideration.
210
Calculating S-parameters from Voltages
As it was mentioned before, S-parameters can be interpreted in terms of voltage at the
DUT in a Z0 environment.
The following sketch gives an explanation about how to calculate them for a given twoport
(DUT), imbedded in an external circuit, which itself represents the characteristic
impedance.
Z1 = Z2 = Z0
e.g.Z0 = 50
Z1 = Z2 = Z0
e.g.Z0 = 50 Ohm
211
Converting N-Port S-Parameters to M-
Port
Converting 3-port parameters to 2-port and back
In order to transform e.g. Common-Emitter S-Parameters to Common-Base, we will now

consider 3-Port S-Parameters. Provided a certain port is connected to ground, we are then
able to evaluate the conversion formulas.
Note
For details, see the publication /Stassen/ mentioned at the end of this chapter.
3-PORT -> 2-PORT
Provided that the device under test is connected with 50 to all its 3 ports, the 3-pole
network S-parameters are defined as below:
If port 3 is connected to ground, we get for the corresponding reflection coefficient 3
-----------------------------------(2)
or
what changes (1c) to
------------------------------------(3)
substituting (3) in (1a) and (1b), we get the 2-port network parameters from the 3-port
ones as:
Converting from 2-PORT -> 3-PORT

Using the matrix in (4), we can also calculate the 3-port S-parameters out of 2-port
parameters.
As described in publication /Stassen/, see below, the sum of each column and row in a 3-
port S-parameter matrix equals '1', i.e.
and
Referring to the above equations (1) and (2), we can calculate the 3-port S-parameters
out of the 2-port ones with Index T (Twoport Measurement) :
Once these 3-port S-parameters are known, they can

be used to calculate the common-base or common-
collector 2-port S-parameters of a bipolar transistor
212
etc.
Publications:
Rudolf Stassen, 'Einsatz eines Mikrowellen-SPitzenmeplatzes zur Charakterisierung von
GmbH, available from Technische Informationsbibliothek, Hannover.
Converting from 2-PORT -> 1-PORT and its

Application for the Q factor of RF-passive Components
This is required if a component has been measured in 2-Port mode, and some of its
components characteristics might be affected by the characteristic impedance of the
opposite VNA port.
We refer to equation. (3) from above and obtain:
213
Definition of S-Parameters
Scatter Parameters, also called S-parameters, belong to the group to twoport parameters
used in twoport theory. Like the Y or Z parameter, they describe the performance of a
twoport completely. Different to Y and Z, however, they relate to the traveling waves that
are scattered or reflected when a network is inserted into a transmission line of a certain
characteristic impedance ZL. Therefore, S-parameters can be compared to reflection and
through-pass of a pair of spectacles.
S-parameters are important in microwave design because they are easier to measure and
to work with at high frequencies than other kinds of twoport parameters. They are
conceptually simple, analytically convenient and capable of providing detailed insight into
a measurement and modeling problem. However, it must kept in mind that -like all other
twoport parameters, S-parameters are linear by default. I.e. they represent the linear
behavior of the twoport.
This means that S-parameters do relate traveling waves (power) to a twoport's reflection
and transmission behavior. Since the twoport is embedded in a characteristic impedance
of Z0, these 'waves' can be interpreted in terms of normalized voltage or current
amplitudes. This is explained below.
Looking at the S-parameter coefficients individually, we have:
S11 and S21 are determined by measuring the magnitude and phase of the incident,
reflected and transmitted signals when the output is terminated in a perfect Zo load. This
condition guarantees that a2 is zero. S11 is equivalent to the input complex reflection
coefficient or impedance of the DUT, and S21 is the forward complex transmission
coefficient.
Likewise, by placing the source at port 2 and terminating port 1 in a perfect load (making
a1 zero), S22 and S12 measurements can be made. S22 is equivalent to the output
complex reflection coefficient or output impedance of the DUT, and S12 is the reverse
complex transmission coefficient.
The accuracy of S-parameter measurements depends greatly on how good a termination
we apply to the port not being stimulated. Anything other than a perfect load will result in
a1 or a2 not being zero (which violates the definition for S-parameters). When the DUT is
connected to the test ports of a network analyzer and we don't account for imperfect test
port match, we have not done a very good job satisfying the condition of a perfect
termination. For this reason, two-port error correction, which corrects for source and load
match, is very important for accurate S-parameter measurements.
Let's now discuss some characteristic S-parameter values.
S11 and s22
214
Value Interpretation
-1 All voltage amplitudes towards the twoport are inverted and reflected (0)
0 impedance matching. no reflections at all (5).
1 voltage amplitudes are reflected (infinite ).
The magnitude of S11 and S22 is always less than 1. Otherwise, it would represent a
negative ohmic value. On the other hand, the magnitude of S21 (transfer characteristics)
respectively S12 (reverse) can exceed the value of 1 in the case of active amplification.
Also, S21 and S12 can be positive an negative. If they are negative, there is a phase shift.
Example: S21 of a transistor starts usually at about S21 = -2 -10. This means signal
amplification within the Z0 environment and phase shift.
S21 and S12
Magnitude Interpretation
0 no signal transmission at all.
0...+1 input signal is damped in Z 0 environment.
+1 unity gain signal transmission in the Z 0 environment.
>+1
The numbering convention for S-parameters is that the first number following the S is the
port at which energy emerges, and the second number is the port at which energy enters.
So S21 is a measure of power emerging from Port 2 as a result of applying an RF stimulus
to Port 1.
In order to better understand the Sxx parameters, let's consider an ideal transmission
line, connected to a port of the VNA, and terminated with either a LOAD, an OPEN or a
SHORT, or an ideal RESISTOR. (This chapter is from HP App.Note 1287-1).
a transmission line connected to a VNA
At very low frequencies, with wavelengths much larger than the line, the transmission line
can be thought of a simple wire. This is adequate for conducting DC or very low frequency
power. The resistance of the wire is relatively low and has little effect on low-frequency
signals. The voltage and current are the same no matter where a measurement is made
on the wire. At higher frequencies, wavelengths are comparable to or smaller than the
length of the transmission line (or a conductor in a high-frequency circuit), and power
transmission can be thought of in terms of traveling waves. When the transmission line is
terminated in its characteristic impedance, maximum power is transferred to the load.
When the termination is not equal to the characteristic impedance, that part of the signal
that is not absorbed by the load is reflected back to the source.
If a transmission line is terminated in its characteristic impedance Z0 (LOAD condition), no

reflected signal occurs since all of the transmitted power is absorbed by the load. Looking
at the envelope of the RF signal versus distance along the transmission line shows no
standing waves because without reflections, energy flows in only one direction.
When the transmission line is terminated in a SHORT circuit (which can sustain no voltage
and therefore dissipates zero power), a reflected wave is launched back along the line
toward the source. The reflected voltage wave must be equal in magnitude to the incident
voltage wave and be 180 degrees out of phase with it at the plane of the load. The
reflected and incident waves are equal in magnitude but traveling in the opposite
directions.
If the transmission line is terminated in an OPEN-circuit condition (which can sustain no

current), the reflected current wave will be 180 degrees out of phase with the incident
current wave, while the reflected voltage wave will be in phase with the incident voltage
wave at the plane of the load. This guarantees that the current at the open will be zero.
The reflected and incident current waves are equal in magnitude, but traveling in the
opposite directions. For both the short and open cases, a standing wave pattern is set up
on the transmission line. The voltage valleys will be zero and the voltage peaks will be
twice the incident voltage level.
If the transmission line is terminated with say an ideal 25 RESISTOR, resulting in a

condition between full absorption and full reflection, part of the incident power is absorbed
and part is reflected. The amplitude of the reflected voltage wave will be one-third that of
the incident wave, and the two waves will be 180 degrees out of phase at the plane of the
load. The valleys of the standing-wave pattern will no longer be zero, and the peaks will
be less than those of the short and open cases. The ratio of the peaks to valleys will be
2:1.
Voltage standing wave ratio
The traditional way of determining RF impedance was to measure VSWR using an RF

probe/detector, a length of slotted transmission line, and a VSWR meter. As the probe
215
was moved along the transmission line, the relative position and values of the peaks and
valleys were noted on the meter. From these measurements, impedance could be derived.
The procedure was repeated at different frequencies. Modern network analyzers measure
the incident and reflected waves directly during a frequency sweep, and impedance results
can be displayed in any number of formats (including VSWR).
The most general term for ratioed reflection is the complex reflection coefficient,
(gamma). The magnitude portion of is called (rho). The reflection coefficient is the
ratio of the reflected signal voltage level to the incident signal voltage level. For example,
a transmission line terminated in its characteristic impedance Z0, will have all energy
transferred to the load so Vrefl = 0 and = 0. When the impedance of the load, ZL is not
equal to the characteristic impedance, energy is reflected and is greater than zero.
When the load impedance is equal to a short or open circuit, all energy is reflected and
= 1. As a result, the range of possible values for is 0 to 1.
Reflection Parameters
Return loss is a way to express the reflection coefficient in logarithmic terms (dB). Return
loss is the number of decibels that the reflected signal is below the incident signal. Return
loss is always expressed as a positive number and varies between infinity for a load at the
characteristic impedance and 0 dB for an open or short circuit. Another common term
used to express reflection is voltage standing wave ratio (VSWR), which is defined as the
maximum value of the RF envelope over the minimum value of the RF envelope. It is
related to as
(1 + )/(1 - ).
VSWR ranges from 1 (no reflection) to infinity (full reflection).
After these thoughts on the Sxx parameters, lets finish by considering the properties of
the Sxy parameters. The transmission coefficient is defined as the transmitted voltage
divided by the incident voltage. If the absolute value of the transmitted voltage is greater
than the absolute value of the incident voltage, a DUT or system is said to have gain. If
the absolute value of the transmitted voltage is less than the absolute value of the
incident voltage, the DUT or system is said to have attenuation or insertion loss. The
phase portion of the transmission coefficient is called insertion phase.
Transmission parameters
Direct examination of insertion phase usually does not provide useful information. This is
because the insertion phase has a large (negative) slope with respect to frequency due to
the electrical length of the DUT. The slope is proportional to the length of the DUT. Since it
is only deviation from linear phase that causes distortion in communications systems, it is
desirable to remove the linear portion of the phase response to analyze the remaining
nonlinear portion. This can be done by using the electrical delay feature of a network
analyzer to mathematically cancel the average electrical length of the DUT. The result is a
high-resolution display of phase distortion or deviation from linear phase.
Applying electrical delay
Note
Adding electrical delay should not be used for modeling.
Another useful measure of phase distortion is group delay. This parameter is a measure of
the transit time of a signal through a DUT versus frequency. Group delay can be
calculated by differentiating the DUT's phase response versus frequency. It reduces the
linear portion of the phase response to a constant value, and transforms the deviations
from linear phase into deviations from constant group delay, (which causes phase
distortion in communications systems). The average delay represents the average signal
216
transit time through a DUT.
Understanding group delay
217
Differential S-Parameters
When measuring S-parameters using a VNA, the DUT is characterized by reflection and
thru measurements at both ports. These signals are referring to signal ground.
In some applications, however, the device may be used in differential mode, e.g. in
differential amplifier stages etc. In such a case, the characteristics of the device may be
different from those measured in the conventional S-parameter measurement
environment. This is especially important for passive RF components and their Q factor,
i.e. the 'quality factor' of their RF performance. Therefore, it is desirable to also model
both, the performance and the Q factor, in the appropriate stimulus condition.
Since the S-parameter are linear twoport parameters, they can be converted into Z-
parameters, and the Z-parameters can be converted in special Z-parameters for common
mode excitation and differential mode excitation.
Let's begin with the conventional Z matrix definition:
For differential mode condition, we can define:
and
And for common mode,
and
Referring to the Z matrix,
we can substitute by the common and differential voltages and currents and obtain
with
This conversion scheme can be applied to get differential mode and common mode S-
parameters:
Convert S-parameters to Z
Apply common/differential conversion
Convert Z-parameters back to S
Calculating differential and Common-mode S-

parameters directly from S-parameters
Publication: Ch.Inui, M.Fujishima, Characterization of T-Shaped Terminal IMpedances of

Differential Short Stubs in Advanced CMOS Technology, 2008 IEEE ICMTS Conference,
Edinburgh, UK
Related to the Q-Factor of Spiral Inductors Operated

in Differential Mode
In non-differential mode, the Q factor is calculated from the Z-parameters, which have
been converted to 1-port Z-parameters as
This refers to a measurement condition where the 1st port is tested, and the 2nd is
grounded.
Under differential mode operating conditions, this is not the case. Referring to the sections
above, we now have:
218
what gives
Wrap-up recipe: Under differential excitation of a symmetric inductor, differential

impedance, inductance and quality factor are defined as
Zdiff = Z11 + Z22 - Z12 - Z21
Ldiff = 1/(2*PI*f) * IMAG(Zdiff)
Qdiff = IMAG(Zdiff) / REAL(Zdiff)
Differential 1-Port S-Parameters

The differential one-port S-parameter can be calculated from the conventionally measured
S-parameters by
From that, we obtain the differential Z-parameter (with the diff. characteristic impedance
Z0):
Acknowledgments:
Special thanks to Laurent Gambus of Philips in Can for bringing up this topic of
differential/common 2-port parameters and also for the basic matrix conversion method.
Publications on Differential/Common Mode Z-

parameters
M.Danesh, J.R.Long, R.A.Hadaway, D.L.Harame, "A Q-Factor Enhancement Technique For
MMIC Inductors", 1998 IEEE MTT-S Digest, p.183-186
M. Danesh, et al., "Differential driven symmetric microstrip inductors", IEEE TMTT, vol. 50,
no. 1, Jan., 2002.
Differential 1-Port S-Parameters
after M.Danesh, J.R.Long, "Differentially Driven Symmetric Microstrip Inducors",

IEEE Trans. Microwave Theory and Techniques, vol.50, no.1, pp.332-341, Jan 2002
219
Interpreting S-Parameter Plots
INTRODUCTION TO UNDERSTANDING S-PARAMETER
PLOTS
IC_CAP file: 1_Sxy_plots_LCR.mdl
This sub-chapter covers expected s-parameter measurements for typical LCR circuits.
Additional delay lines would shift the phase of the following plots, but not the magnitude.
If the delay line is matching the impedance of the NWA ports, the phase shift is linear.
Otherwise, it is non-linear. For more details about such phase shifts see the next sub-
chapter.
S11
S21
S21 is always 0, because both ports are decoupled. No power can be transmitted between
them.
220
S11:
S21
S11
S21
221
S11
S21
S11
222
S21
S11
223
S21
224
S11
S21
NOTE: the Sxy curve starts at "10:30 o'clock". This is a hint for the common resistor to
ground!
225
Since the second RC has a higher 3dB frequency than the first, the curve thends for high
frequencies to an overlay of both RC branches
Note
the Sxy curve starts still at "10:30 o'clock".
Since the second RC has a higher 3dB frequency than the first, the curve thends for high
frequencies to an overlay of both RC branches
note
The Sxy curve starts upwards, at "12 o'clock". This is a hint for the cross-coupling capacitor.
226
S11
S21
227
Note
the Sxy curve starts at "8:30 o'clock". This is a hint for the inductor to ground
S11
S21:
228
Phase Shift of Inductors and Capacitors
Attenuation of Resistors Related to S-Parameters

S21 VALUES FOR RESISTORS THAT ARE PARALLEL OR IN SERIES WITH THE NWA PORTS
The interpretation of S11 or S22 curves in terms of complex inductances is simple, since
we use a Smith chart. Yet, the interpretation of the traces of S12 or S21 seem to be more
complex. The following two tables may help in better understanding such Sxy plots. They
give the Sxy values for typical series and parallel ohmic resistors. These values will be
measured when R_PAR and R_SER are imbedded between 50 Ohm delay lines (Port1 and
Port2, respectively):
Note: The formula for R_SER is:
How to use these tables

Assuming we have a value of REAL(S21) equal to 0,95 (i.e. 95% power transmission from
229
Port1 to Port2). As Port2 has 50 Ohms, the series resistor between Port1 and Port2 has 5
Ohms.
If we assume a parallel resistor and again have a value of REAL(S21) equal to 0,95, the
parallel resistor to Port2 (50 Ohm) is now 500 Ohm.
Application
An application of the Sxy -> Rseries context is for example for diode modeling.
Let's assume a diode between Port1 and Port2. Its resistance varies a lot with the DC bias,
see the plot below:
Diode DC characteristics and corresponding diode resistance.
This means. when the diode current changes from pA to mA, the diode resistance drops
from G? to ~10?.
Related to S-parameters, for lowest bias, the S21 of a diode is 0, while for highest bias,
we expect S21 -> 1. (The end point on the real axis of S21 is determined by the ohmic
series resistance of the diode!).
Therefore, the question arises which DC biases to select for diode modeling.
This question can be simply answered by inspecting the R_diode plot and its
corresponding S21-value from the table above:
For S21 starting points between S21~0,1 and S21~0,9, R_diode should be between
roughly 1k? and 10?. For the diode above, suitable DC bias points would be between 2.5V
and 3.5V.
The next plot shows the measurement result based on these DC bias conditions:
S21 characteristics of a diode, biased after the considerations from above
While this S21 plot shows the 'inner diode', i.e. without any series inductance (package or
strip lines on the wafer for finger structured devices), the next S21 plots shows the same
device, but now including a series inductance.
HINT: watch the difference between fig.2 and fig.3:

Fig.2: the diode capacitance shorts the dynamic diode resistance for high frequencies.
The end point of all traces is equal to the RS, the ohmic series diode resistance,
usually close to '1'.
Fig.3: the shortening effect of the diode capacitance is overlaid by the extra phase shift
of the inductance. Therefore it can happen, that the starting angle for inrementing DC bias
changes from +90 abruptly to -90 !
230
S21 characteristics including a series inductance(e.g. package)
About the Interpretation of S-Parameters of an

Asymmetrical Tee Structure
This tutorial is intended to help with the understanding of a typical small signal modeling
problem and the interpretation of its Smith charts. It can be applied to better understand
the trace of S11 and S22 curves of small signal transistor models after de-embedding. It
explains especially the case where 'elliptic' curves are obtained instead of simple 'half
circles'.
Schematic of an asymmetrical TEE schematic
The total resistance as seen at port1 is:
For extreme frequencies, Rtot behaves mainly like a capacitor Ceff with frequency
dependent value.
For low frequencies, it behaves like Ceff_low = C1 ( 1 + C12/C1)
and for high frequencies like Ceff_high = C1
Referring to low frequencies and therefore Ceff_low, we can distiguish between two cases:
C12/C1 >> 1
and C12/C1 << 1
In the first case, Ceff_low ~ C12, and the influence of C1 can be neglected. The expected
S11 curve will start at 'infinite' resistance, and lead towards 50 Ohm for higher
frequencies (Port2 resistor R2). Yet for even higher ones, C1 will become conductive and
will conduct the power flow from Port1 directly to ground. Therefore, R2 will become
'invisible'. Between both ranges, we expect a transitional trace of S11. This situation is
depicted in the Smith chart of shown below.
In the case that C1 >> C12, we have Ceff ~ C1 for all frequencies. I.e. the signal flow
from Port1 directly into the 'big' C1 is dominating always over the other flow through the
'small' C2 into R2. This is shown in fig.2b.
Fig.3 shows the trace of fig.2a converted to the complex resistance plane. The trace
comes for low frequencies from 50? - j ? , and changes its direction to tend to 0?. As can
231
be seen, this kind of plot is ideally suited to extract the two ohmic resistance values of the
underlaying schematic.
Fig.3: S11 from fig.2a, converted to the complex resistance plane. The knee happens at
the same frequency as in fig.2a.
Application for transistors:

Fig.2a is often seen when measuring S22 curves of transistors. From the considerations
above, the interpretation might now be easier when understanding the importance of the
relative parameter values of CBE and CBC (for a bipolar transistor). This is covered in the
last plots of this section.
Another example is the modeling of the output resistance of MOS transistors in the
operating point. Here, the measured S-parameters are converted to Y, and Y22-1 is
displayed like in fig.3. This allows the easy separation in typically a high-value ohmic
output resistance and one or two series RC combination. More details see in the section
"Operating Point Modeling of MOS Transistors".
S11
232
S21
Interpreting S-Parameter Plots of a Transistor
Starting Points of Transistor S-Parameter Plots

example MOS transistor
IC_CAP file: calc_spar_start_from_DC.mdl
Of particular interest is the starting point of transistor S-parameter traces. These starting
ponts are completely determined by the DC performance. This means that, if the starting
points do not match for the transistor modeling, there is no chance to match the
measured S-parameter traces with the model at all.
The S-parameters of the DUT can be calculated out of these schematics:

Forward Simulation following the formula:
Reverse Simulation with the formula:
233
Proveded, Rin, gm and Rout are obtained and known from DC measurements, we can
calculate the voltages for Z0=50:
FORWARD:
REVERSE:
These results introduces into the S-parameter equations from above gives finally with
V01f = 1 and v02r = 1:
Note about a possible mismatch between measured and simulated S21 transistor data at
lowest frequencies
Although you have achieved an excellent fitting of the DC transfer curve, i.e. the output
plot and even the Rout plot match well, you might run into some problems when fitting
the S21 curves of a transistor. Since with S-parameter modeling, we can only affect the
trend of the simulated curves towards higher frequencies by changing the transit time
parameters, the capacitances or some external RF parasitics, we cannot change the
starting points at lowest frequencies. These points are completely determined by the DC
fit.
This means we might run into a modeling problem like depicted in the following figure:
mismatch between measured and simulated data of S21 at low frequencies
It has been observed that this effect can be due to the ohmic losses in the bias TEEs used
for the S-parameter measurements. For example, if we bias the transistor directly from
the DC instrument (HP414x, HP415x) into the bias connectors of the S-parameter test set
of the VNA, we have to account for a total Rbias of about 2.5 ? at each port. Using the
external bias TEEs, the remaining resistance is considerably lower, in the range of 0.5 .
What does this mean for a MOS transistor: no voltage drop at the Gate, since Ig=0,
However, a voltage drop for Vd. And, indeed, the biggest S21 mismatch can be observed
for DC bias conditions where the output characteristics curves (Id vs. Vd) are most non-
linear, i.e. for low Vd.
For a bipolar, the Rbias affects again mainly Ic, but also a bit Ib. On the other hand, the
output characteristics is usually more linear than that of a MOS.
This means, we define a certain DC bias value for Vd in the S-parameter setup, but this
Vd is not appplied to the transistor Drain. In order to compensate for this effect, we firstly
measure the Rbias at each port by connecting the probes to the SHORT calibration
standard, and by performing a DC current sweep (-10mA .. 10mA) for the SMU,
measuring the voltage. Using the derivative function of IC-CAP, we can easily calculate the
Rbias. This is done for both ports. After that, we add a Rbias1 and Rbias2 into the circuit
deck of our transistor. Rbias1 in series with the Gate, and Rbias2 in series with the Drain.
Both resistors are AC-wise shorted by e.g. 1F capacitors. (If we would not short these
Rbias, we would influence considerably the RF trace of our S-parameters. By shorting
them with these big capacitors, only the DC operating point of the transistor is affected,
and not its S-parameter traces).
234
Example for a enhanced SPICE circuit deck:
Applying this enhanced schematic, we obtain the curves of fig. 2 instead of those in the
following figure:
Interpreting S-Parameters of Strip Line

Strip Line Tutorial
A strip line exhibits a TEM field, i.e. a transverse electromagnetic mode field, where the
dielectric field is perpendicular to the magnetic field, and both are perpendicular to the
direction of wave propagation.
From this sketch, we can conclude the following basic properties of a microstrip line,
depending on the geometry factor W/L:
The characteristic impedance, Z0, ranges from about 20 to about 100. The limit of
100 exists for a very simple reason: the width is much less than the hight, and such a
structure cannot be manufactured (under-etching etc).
This sketch allows to make some fundamental considerations:
As a matter of fact, a small microstrip line exhibits less capacitance than a wide one.
Inspecting the plot, this concludes that a lower capacitance in a microstrip line comes
along with a lower impedance Z0.
This is obvious, when recalling that
Referring to crosstalk between lines, we can learn from the sketch above that a low
impedance microstrip line is capacitive. I.e. the energy is rather between the metal
conductor and the groud. I.e. two low impedance striplines side-by-side, will exhibit less
cross-talsk than two high impedance striplines. By the way, this is a key design rule for
235
packages and connectors.
As another important outcome, a shielding across a microstrip line will lead to the fact
that the impedance of the resulting strip line wil be lower, because more of the electro-
magnetic field will now be present in the enlarged electric field consisting of the previous
field in the dielectic layer plus the additional space between the active metal layer and the
top cover. Therefore, a cover across a microstrip line reduces the resulting impedance,
and, thus, reduces cross-talk between adjacent striplines.
To further reduce cross-talk of adjacent lines, i.e. to reduce the impedance of each line
(inclrease the electric field, i.e. make the lines more capacitive), reduce the height of the
dielectric material.
However, on the other hand, a cover 'kills' the performance of filters designed from strip
lines based on electric field coupling!
OPEN ENDED STRIP LINE VERSUS A THRU STRIP LINE
Note
Composing striplines from ABC matrices.
When defining a strip line of an ABC matrix (chain matrix), please note that in this case
the line is from Port1 to Port2. If you need to define an open ended strip line with an ABC
matrix, make sure to multiply the above ABC strip line matrix with another ABC matrix of
a thru with a very big series resistor value !
See the two PEL programs below:
THRU STRIPLINE
! A PEL program for a ABC matrix THRU microstip line

COMPLEX Athruline.M.22[SIZE(freq)]
i=0
WHILE i < SIZE(freq)
tmp=j*2*PI*T1.TD*freq[i]
Athruline.11[i]=cosh(tmp)
Athruline.12[i]=T1.Z0*sinh(tmp)
Athruline.21[i]=T1.Z0^-1*sinh(tmp)
Athruline.22[i]=cosh(tmp)
i = i + 1
END WHILE
RETURN Athruline
________________________________________________________________
OPEN STRIPLINE
! A PEL program for a ABC matrix OPEN microstip line

COMPLEX Aline.M.22[SIZE(freq)]
i=0
tmp=j*2*PI*T1.TD*freq[i]
Aline.11[i]=cosh(tmp)
Aline.12[i]=T1.Z0*sinh(tmp)
Aline.21[i]=T1.Z0^-1*sinh(tmp)
Aline.22[i]=cosh(tmp)
i = i + 1
END WHILE
! ABC matrix of a 1G resistor
! NOTE: the A matrix of an OPEN is NOT defined!
COMPLEX Aopen.M.22[SIZE(freq)]
i=0
Aopen.11[i]=1
Aopen.12[i]=1G
Aopen.21[i]=0
Aopen.22[i]=1
i = i + 1
END WHILE
RETURN Aline*Aopen
Note
Phase shift of an OPEN and a THRU strip line.
When simulating an OPEN ended strip line, the delay of the strip line shows up in Sxx.
Note that due to forward and reverse traveling of the waves along the open ended strip
line, the strip line delay shows up multiplied by 2. This becomes evident when comparing
such a result with a THRU strip line.
THRU strip line between Port1 and Port2
236
OPEN ended strip line at Port1
Note
Both strip lines have the same length L.
Conclusions
For the same phase shift, a THRU strip line needs a length 2*L compared to an OPEN
ended strip line of length L !
S-PARAMETERS OF SERIAL LOSSLESS DELAY LINES

WITH DIFFERENT IMPEDANCES
Delay lines show up in S-parameter plots with big phase shift. If they have no loss, like
the SPICE models they keep the magnitude. This is depicted below.
Yet, if there is a mismatch in the impedances, the phase is affected. The following plots
give an idea. Both lines have a delay of 30ps. The frequency is swept from 45MHz to
20GHz.
Case 1: impedance steps to lower value
237
Note
a step to lower impedance: -> more phase shift at low frequencies!
Case 2: impedance steps to higher value
238
Note
NOTE: a step to higher impedance:-> less phase shift at low frequencies!
As a final note, we can conclude that for lossless lines, there is no phase ripple, provided
there are no multiple reflections (connector mismatch, impedance mismatch). If we
encounter phase ripple, we have to take multiple reflections into consideration.
239
Q-Factor Calculation
The quality factor of a spiral inductor is given by:
The best way to get rid of the 50 Ohm terminations of 2-port S-parameters is to calculate
1-port S-parameters:
From that, we apply the basic S11 <> R conversion, mentioned in the chapter on Smith
charts,
and obtain for the input impedance at Port 1
From this, we can obtain the requested 1-Port characteristic like the Q factor of a spiral
inductor:
Note
The same result is obtained when converting the S-parameters to Y-parameters, and then calculating.
240
Small Signal Versus Large Signal S-
Parameters
So far, we introduced the S-parameters and compared them to the other twoport
parameters like Y or Z. This means, S-parameters are small signal parameters by
definition. For a transistor as an example, the S-parameters do not reflect non-linear
amplification phenomena like compression etc.
In general, twoport parameters of non-linear components like transistors or diodes vary as

a function of input power. Referring to S-parameters, the parameters |S1l|2 and |S21|2
are defined as a function of power incident at port 1 with no power incident at port 2;
whereas |S12|2 and |S22|2 are defined as a function of power incident at port 2.
Therefore, in RF simulators like ADS, all S-parameters of nonlinear electrical elements are
represented by a S-parameter matrix versus input power at port 1 for forward and port 2
for reverse operation.
This data set has been accepted as a convenient means of characterizing nonlinear
devices by their large-signal S-parameters and have been successfully used for designing
power amplifiers, oscillators, etc.
However, keep in mind that when the S-parameters become RF-power dependent,
harmonics occur. In order to characterize these harmonic frequencies, a spectrum
analyzer or a non-linear network analyzer should be applied!
Figure: Small signal S-parameters are a function of frequency and, for transistors, also of bias, while large signal S-
parameters also cover the signal power.
Note: even for the three-dimensional character of large signal S-parameters, they are by
definition linear. This means, a single frequency power |al|2 injected into port1 will lead to
reflected and transmitted power with exactly the same frequency. No harmonic
frequencies!
S-Parameter Basics for Modeling Engineers
This is the cover of the February 1967 issue of the Hewlett-Packard Journal, yet its
content remains important today. S-parameters are an essential part of high-frequency
design, though much else has changed during the past 30 years. During that time, HP,
now Agilent Technologies, has continuously forged ahead to create today's leading test
and measurement environment.
Describing Twoport Networks
To characterize the performance of such a network, any of several twoport parameter sets
can be used, each of which has certain advantages. Each parameter set is related to a set
of four variables associated with the twoport model. Two of these variables represent the
241
excitation of the network (independent variables), and the remaining two represent the
response of the network to the excitation (dependent variables).
Example: y parameters
S-Parameter Analogy
"Scattering parameters", which are commonly referred to as S-Parameters, relate to the
traveling waves that are scattered or reflected when a network is inserted into a
transmission line.
S-parameters are important in microwave design because they are easier to measure and
to work with at high frequencies than other kinds of parameters. They are conceptually
simple, analytically convenient, and capable of providing a goog insight into a
measurement or design problem.
Note
S-parameters are a mean to characterize an electronic device, in a well-defined environment (i.e.
terminated with the characteristic impedance Z0). To better understand this, we consider an example
from optics, a lens. If your lens is in a well-defined environment, e.g. air, your optician can characterize its
performance from measuring the reflection and throughput characteristics. If your lens would be put on
water, and the other side remained in the air, such a measurement would be much more difficult. So,
what is true for your spectacles, is true also for S-parameter. They allow to characterize a circuit in a well-
defined environment, the characteristic impedance Z0. This means that S-parameters relate traveling
waves (power) to a twoport's reflection and transmission behavior.
From Light Waves to S-Parameter 'waves'

at high frequencies, voltages and currents are difficult to measure
BUT: electrical high-frequency signals are represented by el.power. THEREFORE:
when light power works well for optics, why not using electrical power for RF
characterization?
Note
By injecting electrical power to the device, we will get the AC currents and AC voltages at the
terminals of the device by simple calculations, as will be seen later.
Considering a 2-port: stimulated at port1, then at port 2
Note
During S-parameter measurements, the power sources provide the RF signal to the twoport. The voltages
v1, i1, v2 and i2 are therefore a consequence of that stimulating power source. On the other hand, the
above plot explains how they can be calculated out of the measured or simulated S-parameters.
242
The S-Parameters relate the b(out of the device) PowerWaves to
the a (into the device) PowerWaves
The conditions a1=0 and a2=0 are fulfilled with a impedance of Z0 (typ. 50 Ohm) in
series with the stimulating voltage source at the input, and with an output termination of
again Z0.
S-Parameters in Power or Voltage Domain
243
Note
S-Parameters are unit-less (they are relations of 'waves').
These 'waves' can be interpreted as 'power waves', or 'voltage waves'. And this makes the comparison
between measurement results shown on the instrument's screen and results displayed e.g. in ADS Data
Display a bit confusing.
As shown above, instruments display the results as power waves -> you can make a simple test: measure
an OPEN standard at the NWA's port1, and then connect a resistor with 150 Ohm. MAG(S11) drops from 1
to . On the NWA display, when selecting 'LOG MAG', the trace drops by 3dB.
Simulators (like ADS), and S-parameter data files (e.g. Touchstone), apply the voltge wave interpretation-
> the experiment above, performed using an ADS S-parameter simulation, drops by 6 dB !!
Therefore, in IC-CAP Plots, we do not apply the IC_CAP built-in LOG10 scale, which is based on
20*LOG10(MAG()), but specify by ourselves 10*LOG10(MAG()).
This way we get the same Plot result on the NWA display and in IC-CAP (what is important when e.g.
verifying the calibration etc).
-----------------------------------------------
dB Voltage Waves: 20*LOG10(MAG())
dB Power Waves: 10*LOG10(MAG())
-----------------------------------------------
Interpreting Some Important S-Parameter Values
Reflection S-Parameters S11 and S22
Applying the Smith Chart to S11 and S22
What makes S xx-parameters especially interesting for modeling, is that S11 and S22 can
be interpreted as complex input or output resistance of the twoport (including the
termination at the opposite side of the twoport with Z0 !!). That's why they are usually
plotted in a Smith chart.
The Smith chart is a transformation of the complex impedance plane Z into the complex
244
reflection coefficient, i.e. Sxx, following:
!spar_figequation5.gif!with the system's reference impedance Z0 = 50 .
This means that the right half of the complex impedance plane Z is transformed into a
circle in the Sxx-domain with radius '1' .
In order to get more familiar with interpreting the Smith chart, let's consider a square
with the corners (0/0), (50/0), (50/j50) and (0/j50) in the complex impedance
plane 'Z' and its equivalent in the Smith chart with Z0=50. Watch the angel-preserving
property of this transform (rectangles stay rectangles close to their origins). Also watch
how the positive and negative imaginary axis of the Z plane is transformed into the Smith
chart, and where (50/j50) is located in the Smith chart. Also verify that the center of the
Smith chart represents Z0, i.e. for Z0 = 50, the center of the Smith chart is (50/j0).
Some more:
Sxx interpretation
-1 all voltage amplitudes towards the twoport are inverted and reflected (0 )
0 impedance matching, no reflections at all (50 )
+1 voltage amplitudes are reflected (infinite )
Note
The magnitudes of S11 and S22 are always 1. Do not forget that included in Sxx is the termination at
the opposite side of the two-port, usually Z 0
Interpreting the Smith chart of S11 and S22
Note
for TwoPorts,Sxx represents the input impedance at Sxx including Z0 of the opposite port. The slide above
gives Smith chart plots for different electrical components connected to a single port of a vector network
analyzer (VNA). For an ohmic resistor, all measurement results are located on the x-axis, R<Z0 is to the
left of the center, R>Z0 to the right. From the scaling of the Smith chart (x = 0 ..1.. infinite), we can
calculate the value of the resistor as x*Z0.For inductors and capacitors, however, the measurement result
is frequency dependent and follows the sketched traces. Note that the curves always turn to the right with
increasing frequency.
We consider the component between one port and ground. For a component between the ports, the 50
Ohm of the opposite port always show up together with the component's impedance Z
Transmission S-Parameters S21 and S12
Interpreting the Polar Plots of S12 and S21 (Sij)
245
S-Parameter Interpretation at a glance
Sxx Input/Output Resistance of the Twoport up to and Including Port 2/Port 1

Sxy Power Transmission of the Twoport in the Z0=50 System
Note
Note: when interested in the twoport's input/output resistance alone, without the char.impedance of the
opposite VNA port, use the Y parameters.
Typical S-Parameter Traces
S11:
freq S11 Resist. Notes

0 +1 infin. C is open
infinite 0 50 C is short
but we see 50 of
Port2
S21
freq S21 Notes
0 0 no signal is transmitted
infinite 1+ C is a short, all signal is transmitted
A Specific Sxx Parameter Trace (very often seen with transistors)
For the schematic above, we can distinguish between two cases: C12 >> C2and C12 <<
C2
246
In the first case, we have Ceff~ C12, and the influence of C2 can be neglected. The
expected S22 curve will start at 'infinite' resistance, and lead towards 50 Ohm for higher
frequencies (Port1 impedance Z0). Yet for higher frequencies, C2 will become more and
more conductive and will make the power flowing from Port2 more and more directly to
ground. Therefore, Z0 of port1 will become 'invisible'. The S22 curve will change its high
frequency end value, and will now tend towards 0Ohm for infinite frequency. Between
both ranges, we expect a transitional trace of S22.
Not to the case C2 >> C12: in this case, the 'easier' path for the signal, for all
frequencies, will be from Port2 directly to Ground. C12 and Z0 of port1 will not be visible
in the Smith chart.
Note
This behavior is very often found with the S22 of transistors.
Interpreting Transistor S-Parameter Traces
Interpreting Transistor S-Parameter Traces
S-Parameter Interpretation Basics

For frequency, 0 -> infinite,
Sxx Parameters of RC or RL circuits are half-circles
Sxx Parameters of CL circuits are full circles (resonance)
Sxy Parameters for active components have a magnitude >1 (power amplification in a 50
W system)
All S-parameter curves turn clock-wise!

Sxx and Sxy Parameters starting points at lowest frequency are identical to the DC
measurement.
Detailed Interpretation of S-parameters, assuming an underlying

schematic
247
PI schematic components vs. freq (at a fixed DC bias)
Conclusion
S-parameters are specific twoport parameters,
for linear circuits

related to power instead of voltage/current
As a very important last remark, S-parameters are members of the LINEAR Twoport
Theory.
In other words, a prerequisite of applying them to nonlinear devices like diodes and
transistors is to make sure that these nonlinear devices have been measured in small
signal excitation. I.e. the applied RF signal level has to be small enough to avoid any
non-linearities. This means, only the stimulating frequency is allowed to be present in
the measured S-parameters, and no harmonics of this stimulating, fundamental
frequency.
For Experts
Coming back to the transmitted and reflected PowerWaves
248
Note
During S-parameter measurements, the power sources provide the RF signal to the twoport. The voltages
v1, i1, v2 and i2 are therefore a consequence of that stimulating power source. On the other hand, the
above plot explains how they could be calculated out of the measured or simulated S-parameters.
Getting vx and ix from S-parameters
The normalization of v01=1 and v02=1 allows to simplify the above equations. It
corresponds to an AC stimulus at Port1 or Port2 of 1V AC signal. This is OK because S-
Parameters are linear, and represent the tangent to the DC curve (with 50 Ohm
termination load) in the operating point.
and now that the Vx and Ix are known from S-parameters. we can calculate some
transistor characteristics which are based on relations like:
Note
Due to the normalization to v01=1 and v02=2, we cannot get the absolute voltages v1, v2 and the
currents i1, i2 from the S-Parameters, but we can easily get characteristics like and gm which are the
quotient of ix and vx.
and inversely, we can also calculate S-Parameters out of a voltage simulation
Again, because S-Parameters are linear, we can scale the applied voltage to any value,
and get the corresponding (linearly) scaled other voltage and currents. Therefore, setting
V01=1V and V02=1V is pretty practical, because it just simplifies the above equation.
249
S-Parameter Impedance Matching (for
obtaining max. power gain)
This is an excerpt from HP App.Note AN 95-1
In order to amplify the maximum gain, we will now discuss to synthesize matching
networks that will transform the source and load impedances Z 0 to the impedances
corresponding to reflection coefficients of S11* and S22* , respectively, i.e. the conjugate
complex S11 and S22. This can be done for a single frequency or a frequency band.
For simplicity, let's consider a matching for a single frequency. In this case, the matching
networks need not be complicated. Simple series capacitor, shunt inductor networks will
not only do the job, but will also provide a handy means of biasing the transistor, via the
inductor, and of isolating the DC bias from the load and the source.
Values of L and C to be used in the matching networks for the small band amplifier at e.g.
300MHz are determined using the Smith Chart below.
Smith Chart for 300- MHz amplifier design example.
Note
See AN95-1 for an interactive movie on these manipulations.
First, points corresponding to S11 ,S11* ,S22 ,and S22* at 300 MHz are plotted in the
Smith chart. Each point represents the tip of a vector leading away from the center of the
chart, its length equal to the magnitude of the reflection coefficient being plotted, and its
angle equal to the phase of the coefficient. Next, a combination of constant- resistance
and constant-conductance circles is found, leading from the center of the chart,
representing Z 0 , to S11* and S22* . The circles on the Smith Chart are constant-
resistance circles; increasing series capacitive reactance moves an impedance point
counter-clockwise along these circles.
In this case, the circle to be used for finding series C is the one passing through the center
of the chart, as shown by the solid line in the figure above. Increasing shunt inductive
susceptance moves impedance points clockwise along constant-conductance circles. These
circles are like the constant-resistance circles, but they are on another Smith Chart, one
that is just the reverse of the chart shown! (This is a Smith chart for 1/Z !) The constant-
conductance circles for shunt L all pass through the leftmost point of the chart rather than
the rightmost point. The circles to be used are those passing through S11* and S22* , as
shown by the dashed lines in the Smith chart above.
Once these circles have been located, the normalized values of L and C needed for the
matching networks are calculated from readings taken from the reactance and
susceptance scales of the Smith charts.
Each element's reactance or susceptance is the difference between the scale readings at
the two end points of a circular arc. Which arc corresponds to which element is indicated
in the Smith chart above.
The final network and the element values, normalized and un-normalized, are shown
below:
250
A 300-MHz amplifier with matching networks for maximum power gain.
The following small tutorial gives some additional info about impedance matching:
We commence with the already discussed transform of the complex impedance plane into
the Smith chart plane, see the figure below.
We remember that:
On the impedance plane, with respect to increasing frequency, all curves turn
upwards (for a given, constant L,C,R). Referring to the Smith chart, they therefore
always turn clockwise
For a locus within the lower capacitance half-plane, only a series impedance can
bring the locus to the upper, inductive impedance half-plane
Adding a series capacitance to an impedance within the lower capacitance half-plane,
can bring the locus only to the x-axis at the most, even with C=infinite.
When adding a series inductor:
a zero inductor in series with a resistor does not affect the locus
a medium inductor in series with this resistor results in a bigger total impedance
a L=infinite gives an open circuit.
When adding a series capacitor

an infinite capacitor in series with the resistor does not affect the locus
a medium capacitor in series results in a bigger total impedance
a C=0 gives an open circuit
As a next step, we define an auxiliary overlay chart to reflect Smith chart locus shifts of
251
parallel components. After the locus shift has been determined, the resulting impedance is
then read from the original impedance Smith chart.
Again, it should be noted that when working with the auxiliary overlay chart, read the
resulting impedance from the original impedance Smith chart.
When adding a parallel inductor
a L=infinite inductor in parallel with a resistor does not affect the locus
a medium inductor in parallel with this resistor results in a smaller total impedance
a L=0 shorts the resistor
When adding a parallel capacitor

a C=0 capacitor parallel with a resistor does not affect the locus
a medium capacitor in parallel with this resistor results in a smaller total impedance
a C=infinite shorts the resistor
After these pre-remarks, we are now ready to consider the impedance matching
procedure:
252
The resulting matching network for the output of the transistor is therefore:
The following is an excerpt of the Agilent web page

http://www.tm.agilent.com/data/static/eng/tmo/Notes/interactive/an-95-
1/classes/imatch.html
(as of Jan.2001):
Smith Chart Circuit Elements Paths

You can think of impedance matching using the Smith Chart as driving a car to a specific
destination in Smith Town - a city were none of the streets are straight! By adjusting
circuit components in appropriate order, we can constrain the circuit reflectance to paths
along constant resistance or constant conductance circles. Just like road signs can direct a
car along the circular streets of Smith Town, so can we reach the matching impedance
condition in a straightforward and deterministic way. The graphs below demonstrate how
the various shunt and series L and C components change the circuit reflectance on the
Smith Chart. Assuming the given component is the last component in the matching
network, the circuit reflectance will move as indicated along constant resistance or
constant conductance circles.
253
Forbidden Regions of the Smith Chart

For a given load reflectance, only certain L-C matching networks will be capable of
transforming the source impedance to the load impedance. In fact, for any load
reflectance, exactly two of the four possible L-C matching networks in the Transistor
Amplifier-II model above will be able to do the matching job. But which two?
The charts below can be used to determine which matching networks will work in a given
load situation. If the load reflectance lies within the forbidden region of the Smith Chart
for the indicated matching network, then that network cannot perform the required
matching operation. You cannot drive your car into the forbidden neighborhoods of Smith
Town! They are unpaved!
Use these charts to determine which matching network should be used. First, visually
locate the position of the load reflectance from your inner model on each of the four color
Smith Charts. Then, eliminate the two networks whose forbidden regions overlap the
reflectance point, and use one of the remaining two networks to perform the impedance
match.
254
S-Parameters Measurements for
Modeling
10 Rules for Successful S-parameter Handling

Note
This paper is a collection of experiences with making reliable S-parameter measurements. The underlying
IC-CAP model files are located at:
demo_features\3_MEAS_ORGANIZE_n_VERIFY_DATA\0_MASTER_FILES and further
20_NWA_CAL_VERIFICATIONselect CAL_VERIFY_PNA_2port_MASTERFILE_demodata.mdl or
CAL_VERIFY_genl_NWA_2port_MASTERFILE_demodata.mdl for the older HP/Agilent NWAs.
and 31_DEEMB_VERIFICATION\
CHECK_OPEN_SHORT_DEEMB_with_a_THRU_demodata__MASTERFILE.mdl
DC Bias Voltage Drop in the VNA Bias TEEs
The NWA bias TEEs, internal or external, represent ohmic losses (Rbias) for the DC bias,
and -together with the DC loss of Rcontact-this loss is not calibrated out by the NWA(!)i.e,
the bias voltages at the device will be less than provided by the SMUs. This will shift the
starting point of the S21 curves towards 0.
During the S-parameter measurements, the transistor is DC-biased by bias TEEs. Typical
values for the ohmic losses in the DC path of NWA-built-in bias-TEEs, are ~1Ohm
(PNA/ENA), ~2Ohm (older HP/Agilent NWAs). External bias-TEEs may have lower loss, but
mostly suffer from a limited frequency range (upper and also (!!) lower frequency*) for
the HF. In any case, the DC losses are not calibrated out by the NWA, since the NWA
calibration "does not see" them (DC is not included in the NWA signal path).
Additionally, especially for silicon devices, there is a non-negligible DC contact loss,
additionally to the losses in the bias-TEE. This contact problem is especially true for
aluminum contact pads. Although, AC-wise, this loss is calibrated out (it is included in the
AC signal path!), it is shifting the DC operating point.
For device modeling, if there is a problem with the fitting of the simulated S-parameters
and the measurement curves for lowest frequencies, especially for bias conditions with
higher currents, this total DC bias voltage drop (in the AC/DC bias TEE and the contact
loss) must be considered and the simulation circuit must be enhanced with these resistors
!Typically, using the Test Circuit of the DUT in the IC-CAP ModelFile. In order to not affect
the HF S-parameter simulations, however, the resistors must be shorted AC-wise by big
255
capacitors. Typically, their value is C=100mF.
IC-CAP File: In demo_features\3_MEAS_ORGANIZE_n_VERIFY_DATA

0_MASTER_FILES\31_DEEMB_VERIFICATIONopen the file
CHECK_OPEN_SHORT_DEEMB_with_a_THRU_demodata_MASTERFILE.mdl and see DUT
check_DC_contact_Spar.
Note
For the lowest frequency, the bias-TEE (representing for the RF signal a resonance consisting of its C and
L) must have completely finished its own 360' low-frequency phase turn.
Evaluate the DC bias losses

Bias-TEE resistance Rbias and Probe Contact Resistance Rcontact
1. Contact the SHORT on the ISS calibration substrate (Gold) and measure the
resistance Rbias
2. Contact the SHORT Dummy on the wafer and measure the resistance Rcontact
How to measure Rbias and Rcontact:
1. In a first measurement setup, we characterize the ohmic losses in the bias TEEs of
the S-parameter testset (network analyzer) or in the external bias TEEs. For the
measurement, place your GSG probes on the SHORT standard of the ISS substrate
(gold contacts). Assuming an ideal contact on the golden SHORT of the ISS
substrate, we measure in this setup basically Rbias.
A typical value for Rbias is ~1Ohm (PNA), ~2Ohm (older NWAs), when using the
internal NWA S-Parameter Testset, and ~1Ohm when applying external bias TEEs.
2. In a second step, we place the HF probes on a piece of aluminum on the wafer, or we
contact a SHORT on the wafer. Either the 'de-embedding Dummy' on the wafer, or
any location on the wafer with enough aluminum area to place the probes for a good
short. Assuming 3x the same contact resistance for the GSG probe, we measure in
this setup basically Rtotal = Rbias + 1.5 * Rcontact.Rbias is known from the previous
ISS contact measurement. So we get for Rcontact:Rcontact = (Rtotal - Rbias) / 1.5
A typical value for Rcontact is ~1-2 Ohm for silicon, and nearly 0 Ohm for GaAs (gold
contacts!).
Specify The Probe Contact Resistance In IC-CAP
256
The network shown above -and specified as a 'DUT Test Circuit' in the IC-CAp ModelFile,
emulates the losses in the bias-TEEs and the wafer contact losses. Since the losses affect
only the DC path, they are shorted for the AC signals by the capacitors C_short_x
Then comes the line which calls the model, i.e. the subcircuit description of the transistor,
as defined in Tab 'Circuit' of the IC-CAP model file
IC-CAP File: In your extraction files, in the DUTs which deal with S-parameters, add the
above netlist to the Test Circuit tab.
Example:
demo_features\1_BASIC_MDLG_EXAMPLES\21_Gummel_Poon_demo\
GP_CLASSIC_NPN_Mdlg_Navigator_PELdep.mdland the Test Circuit in DUT 'Spar_mdlg'.
Tell the NWA which calibration standards will be applied "Modifying

a Calkit"
Since we will probe on a wafer, and perform a calibration down to the GSG probe tips, we
need to 'tell' the NWA about these probe tips and their corresponding ISS calibration
substrate. Otherwise, the NWA assumes we calibrate with '3.5mm connector
'cal.standards.
The above screenshot is from the Agilent PNA series of network analyzers. Enter the calkit
data, given e.g. on the inside cover of the Cascade_Microtech probes box, or from the
spec.sheet of the PicoProbe or Suss probes.
ENTERING THE CALKIT DATA INTO THE NWA

Table:Example of Standard definitions for Cascade Microtech G-S-G probes, 100um
pitch*
STANDARD TYPE C0/fF Delay Z0 Loss FREQ STANDARD Label in the
No. or psec Ohm Ohm/s min calkit
L0/pH max
1 OPEN -9.3 0 50 0 0 999 OPEN
2 SHORT 2.4 0 50 0 0 999 SHORT
3 LOAD -0,007 500 0 0 999 LOAD
4 THRU 1 50 0 0 999 THRU
These nonidealities have to be entered into the NWA. This is called defining or
'modifying the calkit'. Check the NWA manual about how to do this.
Some notes about entering the calkit data: While in the case of a CV meter, the
calibration corrects simply for an ideal offset capacitor, a NWA calibration uses
several cal standards (OPEN, SHORT, LOAD, THRU etc.) from a cal kit. These cal
standards do not represent ideal standards. They represent the real, existing
standard, including its nonidealities! It means that a SHORT is not an ideal SHORT,
257
but instead represents rather an inductance. The same applies to the THRU, which
has a non-zero delay time. The OPEN corresponds rather to a capacitor than to an
ideal OPEN. Therefore, these nonidealities of the G-S-G probes have to be entered
into the NWA before calibration. This is called 'modifying the cal kit'.
While this procedure refers to the nonidealities of the calibration standards, for all
frequencies and independent of the applied power, the subsequent calibration based
on these calkit data is related to the selected frequency range, the RF power, the
applied IF bandwidth etc. After it has been performed, the correction terms are
stored in the cal set of the NWA. In other words, the 12-term error vectors are 'filled
up'.
Afterwards, when the measurement is performed, the raw measured data arrays will
be corrected using a correction technique related to the selected calibration method,
and referring to the specified cal set. Finally, this corrected measurement result is
displayed on the NWA monitor. Therefore, after the calibration, a re-measurement of
the OPEN will not represent an ideal open, but instead exactly those parasitic
components as described in the documentation of the OPEN. In the same way, a
THROUGH shows up after calibration with its real delay time, and a SHORT exhibits
its specified inductive behavior.
Find Out The Max. Applicable RF Signal For Your Device
CHECKING THE CORRECT NWA PORT POWER SETTINGS

It is very important to not overdrive a nonlinear device like a transistor with too
much RF signal. It must be absolutely ensured that particularly the transistor's output
signal is not clipping. When this happens, the transistor output signal is no longer a
sine function, i.e. the transistor behaves nonlinear. And this is a contradiction to a
linear NWA measurements! In this case, the NWA will only measure the fundamental
frequency, and ignore the occurring harmonics!
At the same time, when harmonics occur, the DC bias is usually affected and shifted.
This is due to Kirchhoff's law: the sum of currents into a node (for all frequencies)
must be zero. So, if there appears 'energy' in the harmonics, the 'energy' in the DC is
usually affected too.
In this context, keep in mind that:-> a typical RF signal level for modeling transistors
is -40dBmi.e. the NWA port1 is set to -30dBm, plus typically ~10dB attenuation of
cables and connectors.
-40dBm corresponds to a RMS voltage of ~2.2mV, what corresponds to ~3mV peak
or ~6mV peak-peak !
Small signal condition for a transistor is usually defined as VT/10=2.7mV
As mentioned, when harmonics occur, i.e. the test device is overdriven, the DC operating
point is affected and thus, the DC bias current is affected.
We can use this effect to identify the max. applicable RF signal for NWA measurements
with a smart and simple measurement method, depicted in the slide above.
When measuring a DC output characteristics and calculating Rout out of it, the resulting
curve is very sensitive. Therefore, we can use this plot to identify possible effects of a too
big an AC power applied to the transistor.
This means, we measure the DC output characteristics, through the NWA test set, or using
external bias TEEs, and let the NWA operate in continuous mode, i.e. unsynchronized to
the DC measurement. We start with the NWA switched-off, or set to min. RF signal power
and max. port attenuations. We keep that output characteristics measurement (and the
output resistance plot) as a reference.
Then, we increase the RF signal, and measure once again the output characteristic. When
the device is overdriven, i.e. when the operating point is shifted (ib and ic for bipolar, id
for MOS), we recognize that easily by a shift of the output characteristics and a shift in the
output resistance plot. Now we know that harmonics are present. We then reduce the RF
power a bit and know the max. allowed RF power for the NWA S-parameter
measurements for that particular transistor device.
Transistor and Diode S-Parameter measurements are performed including a DC

bias.
Due to the nonlinear transistor behavior, a too big AC amplitude results in a distorted
258
output signal. These distortions will shift the DC operating point.
From the plots above, we can determine 2 important things:

the max. applicable RF signal level for the S-parameter measurements for given DC
operating points, or, inversely, the min applicable DC bias for a given RF signal. In other
words: The starting points of the measured S-par (0Hz)are determined by the DC fitting.
With RF modeling (capacitors and transit time), only the trace from the starting
points towards the end points can be fitted, but not the starting points. If this
starting point check fails, verify:
too much RF signal
self-heating
Voltage drop in S-par testset> accurate DC contact resistance> DC modeling was
performed at different bias range than the S-par modeling.
When using an Agilent PNA/ENA*, specify the NWA power

settingsdirectly in the PNA/ENA Cal.Wizzard, or applying Wincal /
SussCal
Note
See the appendix when using the older Agilent/HP NWAs
In this case, thecalibration is not performed by IC-CAP, and the InstrumentOptionsin the
IC-CAPSetup are as simpleas shown here.
259
Note
All entry fields, except the red marked ones, are ignored by IC-CAP.
Provided that the Cal Kit Data (which is provided with every GSG probes calibration
substrate) have been entered correctly into the network analyzer, and we have
determined the maximum applicable RF signal level, the calibration of the Agilent PNA
network analyzer is pretty simple. (For the older Agilent/HP NWAs, see the slide at the
end of this slide set).
The steps are (for a PNA, and pretty similar for an ENA):
after the PNA was calibrated (applying SussCal, WinCal, or ECal, PNA's built-in Cal),
manually save your CalState on the PNA by selecting File/Save_As, select '.cst' file
format, and save it to the default directory proposed by the PNA. Don't save it to
another directory. IC-CAP would not find it there.
in IC-CAP, specify the frequency range (IC-CAP Input 'freq') identical to the PNA
frequency settings-> in the InstrumentOptions of your IC-CAP Setup, when 'Use PNA
Calibration Settings=YES', then simply specify only - CalType = H, - Cal File Name.
Note: Port1 and Port2 Source Power, Port Attenuation and IF Bandwidth are now
accepted by IC-CAP as they were set by the user (directly on the PNA), or by
SussCal/WinCal.
Don't hit the IC-CAP 'Calibrate' button - it is no longer required in this case
Simply hit 'Measure'.
*IC-CAP file:* The easiest way to verify the calibration of your NWA (PNA) is to use
the following file: in demo_features, go to
3_MEAS_ORGANIZE_n_VERIFY_DATA\0_MASTER_FILES\
20_NWA_CAL_VERIFICATION, and load file
CAL_VERIFY_PNA_2port_MASTERFILE_demodata.mdl.
During the calibration, measure the Calibration Standards on

the ISS substrate
On the NWA's calibration menu, select 'Full 2-Port Calibration'.
Verify the NWA calibration by re-measuring and comparing the

result to simulated standards
260
After the calibration is finished, it is important to measure some devices with known
performance to verify the calibration quality. But, what is a known, 'golden device'?
A set of devices with exactly known characteristics are the calibration standards on
the ISS cal substrate. We remeasure all four of them, and compare the
measurements individually against the simulation of the calkit data.
A really excellent fitting of measured and simulated data is an absolute MUST for
accurate S-parameter measurements!
The goal is a fitting quality for all four measurements in the <2-5% range!
See the IC-CAP file: in demo_features, go to
3_MEAS_ORGANIZE_n_VERIFY_DATA\0_MASTER_FILES\
20_NWA_CAL_VERIFICATIONand load the file
CAL_VERIFY_genl_NWA_2port_MASTERFILE_demodata.mdl, Execute the
Macro'COPY_INSTR_OPTIONS' and then remeasure the standards applying the
Setups Calverif_short/SHORT, Calverif_load/LOAD and Calverif_thru/THRU
In the above slide, in the magnified view, is the inner transistor, which we want to
characterize. All the rest has to be de-embedded, i.e. to be stripped-off. In other
words, the NWA calibration plane, obtained by the NWA calibration so far, has to be
shifted from the GSG probe contact location down to the beginning of the device.
Prepare the De-Embedding: Measure and Model The On-Wafer

Dummies
261
This layout suggestion for Silicon on-wafer Dummy structures uses metal1 to shield
the contact pads from the lossy silicon substrate.Due to this metal1 shielding, the
measured OPEN Dummy S-parameters represent a much more ideal OPEN Dummy
schematic than the more conventional layout without the metal1 shielding. However,
the OPEN Dummy capacitance is bigger than without the metal1 shielding. But it is
representing much more an ideal capacitance.
For the SHORT Dummy, the short itself is much more ideal with the metal1 shielding
compared to without that shielding. This is because of the chain of shorting vias
down to the electrically well-behaving metal1 Ground plane. This chain is
geometrically concentrated and precisely localized, compared to the conventional
SHORT Dummy layout where the device 'hole' is simply filled up with metal1.
Note
For clarity, the layout details are simplified in the slide above. See publication T.E.Kolding, O.K.Jensen,
T.Larsen, "Ground-Shielded Measuring Technique for Accurate On-Wafer Characterization of RF CMOS
Devices", IEEE Int.Conf.on Microelectronic Test Structures ICMTS, March 2000, Monterey,CA.
*IC-CAP file:* go to demo_features, and then

3_MEAS_ORGANIZE_n_VERIFY_DATA\0_MASTER_FILES\31_DEEMB_VERIFICATION\ and
load the file CHECK_OPEN_SHORT_DEEMB_with_a_THRU_demodata_MASTERFILE.mdl
Layout Comparison:Conventional On-Wafer Dummy Structures
This represents the more 'classical' Dummy structure layouts, without using metal1 as a
shield against the lossy silicon substrate.
Modeling The On-Wafer OPEN Dummy
262
In this first step, the OPEN Dummy is modelled in order to verify that there are no
inductive, serial components present. Such inductors would be a contradiction to the
commonly applied OPEN de-embedding, i.e. the OPEN Y-matrix subtraction.
*IC-CAP file:* Go to demo_features, and then further to
load file CHECK_OPEN_SHORT_DEEMB_with_a_THRU_demodata_MASTERFILE.mdl.
See the Setup 'OPEN_dummy'.
Model The On-Wafer SHORT Dummy
Then, the SHORT Dummy is de-embedded from the OPEN and again modelled. The goal is
to verify that there are no parallel components included in the Z matrix of the SHORT. For
example, a parallel capacitive component, required to model the SHORT, is a contradiction
to the usually applied Z-matrix subtraction.
IC-CAP file: go to demo_features and to directory
3_MEAS_ORGANIZE_n_VERIFY_DATA\0_MASTER_FILES 31__DEEMB_VERIFICATION\
and load the file
CHECK_OPEN_SHORT_DEEMB_with_a_THRU_demodata__MASTERFILE.mdl
Go to Setup 'SHORT_dummy'.
263
Verify The De-Embedding With Modeling The THRU Dummy
After the OPEN and SHORT Dummy measurements have been verified by modeling, we
want to check the de-embedding procedure with a known 'golden device'. In practice,
however, like with the NWA cal.verification, the problem is the availability of an 'golden
device'. Provided we have a THRU Dummy, we can use this component for a verification of
the de-embedding procedure. We do not know the model parameters of the THRU. But we
know that the de-embedded THRU should look like a simple delay line, with a certain
characteristic impedance Z0, and a delay time TD. Additionally, the delay line may be
lossy. In most cases, due to the short geometry of the THRU, the loss is negligible.
Anyhow, the model parameters must represent physically meaningful values, and the de-
embedded trace must be simple without resonances etc. Just a simple delay line!
Some modeling hints
TD: models the phase in Sxy Z0: models the Sxx trace:
for Z0<50W, the curve starts at 50W, turns downwards and then to the left into a circle
around the line's Z0. The turning is clock-wise with increasing frequency!
for Z0=50W, the curve is a single point at at 50W in the Smith chart.
For Z0>50W, the curve starts at 50W, turns upwards and then to the right into a circle
around the line's Z0. The turning is clock-wise with increasing frequency!
Only if there were no layout bugs or measurement problems with the Dummy structures,
we will end up with a physical representation of the small, de-embedded THRU strip line!
Therefore, when we obtain a physical strip line characteristic and model, we can assume
that the de-embedding, applied later to transistors, diodes etc., will be correct as well!
Note
since the THRU strip line is passive, and not depending on bias or frequency or temperature, it can be
used much better for de-embedding verification than checking the de-embedding with a transistor, a diode
etc.
IC-CAP file: go to demo features and further to

load the file CHECK_OPEN_SHORT_DEEMB_with_a_THRU_demodata_MASTERFILE.mdl.
See Setup 'verify_with_THRU_dummy'.
An Example of bad THRU line de-embedding results
The OPEN-only-deembedded Sxx is only OK up to 2/3 of the freq. range.

The OPEN-SHORT-deembedded Sxx turns backwards (unphysical impedance Z0).
The OPEN-only-deembedded Sxy looks OK
The OPEN-SHORT-deembedded Sxy turns backwards (neg. delay time)
The slide above represents a de-embedding of Dummy structures with obviously layout
problems.
Although the inspection of the measurement results sequence undeembedded >
264
open_deemb > open_short_deemb exhibits the right trend (with each de-embedding step,
the resulting THRU becomes more and more the expected simple, short THRU), the
open_short_deemb results does not match the expected the modeling target and
represents - a strip line with negative delay time (see Sxy: a short delay line with neg.
delay time and even with a trend for MAG(Sxy)>1 with increasing frequency) - and an
unphysical Z0 (since in Sxx, it turns counter-clock-wise).
Finally, measure the device, but keep an eye on the self-heating

if you have to consider self-heating of your device (typ. >50mW DC power) and if you
measure both, first the DC curves like
transfer curve (id vs vg)

output characteristics (id vs vd)
and then the- DC biased S-parameter curves, make sure that you measure the DC
curves with the same slow speed as during the S-parameter measurements.
Verify the identical self-heating by comparing the DC-only curves with the DC bias
curves obtained during the S-parameter measurements.
DC Measurements vs. Spar DC Bias
The slide shows the S-parameter DC bias currents, measured during the S-parameter
measurements, extracted and overlaid to the conventional DC output characteristics
measurement.
IC-CAP file: go to demo_features and then further to
3_MEAS_ORGANIZE_n_VERIFY_DATA\0_MASTER_FILES\1__CHECK_DATA_CONSISTENCY
and load the file FET_MEAS_CHECKTOOLS.mdl or NPN_MEAS_CHECKTOOLS.mdl
So, How To Deal With S-Parameters?

The 10 Rules At A Glance
1. Measure the DC Bias Voltage Drop in the VNA Bias TEEs

2. Measure The Probe Contact Resistance
3. Tell the NWA which standards will be applied during calibration
4. Identifying the correct AC Power Level for measuring the transistor or diode
5. Perform The NWA Calibration
6. Verify the NWA calibration by remeasuring and comparing the result to simulated
standards
265
7. Prepare the De-Embedding: Measure The On-Wafer Dummies
8. Model the on-wafer OPEN and SHORT dummies
9. Verify the de-embedding with modeling the THRU Dummy
10. Finally, measure the device, but keep an eye on the self-heating
Last not least:Organizing the De-Embedding of S-Parameters in IC-

CAP
Hint: In the demo_features directory, go to directory

3_MEAS_ORGANIZE_n_VERIFY_DATA\0_MASTER_FILESand load the file
NPN_MEAS_MASTERFILE_demodata_PELdep.mdl Check the organization of Setup
'NWA_Spar/freq_n_bias_fwd'.
Particularly, check for:
Output 'S'
Transform 'S_deemb'
Setup Variable 'MDM_EXPORT_XFORM_DATA'=1
Setup Variable 'MDM_XFORM_LIST' = S_deemb S P1 P2 GROUND n/a B
This Setup Variable 'MDM_XFORM_LIST' will assure that the Transform 'S_deemb' is
added to the .mdm file during 'Export', and that it will be loaded back later as an
'Output' called 'S_deemb' (the default name for de-embedded S-parameter) with
Type 'S' Port1 'P1' Port2 'P2' AC Ground 'GROUND' Unit 'n/a' Type 'B'.
Organizing the De-Embedding of S-Parameters in IC-CAP
Wrap-Up Network Analyzer Calibration
Appendix
266
Performing the NWA calibration from IC-CAPfor the older HP/Agilent NWAs like HP8510,
HP87xx etc.
Specifying the NWA power settings in IC-CAP
Provided that the Cal Kit Data (which is provided with every GSG probes calibration
substrate) have been entered correctly into the network analyzer, and we have
determined the maximum applicable RF signal level, the calibration of the NWA is pretty
simple.
The steps are:
First of all, specify the frequency range (IC-CAP Input 'freq')

In InstrumentOptions (for the older HP/Agilent NWAs like 8510, 87xx), enter the
'Source Power' and the 'port x Atten'
Set the 'avg Factor' (typ. 64) - 'Cal Type' = H
Specify the 'Cal Set No'.
In InstrumentOptions (for the Agilent PNA series),
When 'Use PNA Calibration Settings'=NO, then specify
'Port Power Coupled' = No
'Port Src Power' and 'Port 2 Source Power'
to the value evaluated before
set 'Port Atten Auto' = Yes
'IF Bandwidth' (typ. 100Hz or less)
'CalType' = H
'Cal/state File Name' as specified in the PNA before.
As mentioned before, make sure to select an appropriate RF power level, in order to not
overdrive the DUT.
*IC-CAP file:* The easiest way to verify the calibration of your NWA (PNA) is to use the
following file:in demo_features, go to
3_MEAS_ORGANIZE_n_VERIFY_DATA\0_MASTER_FILES\ 20_NWA_CAL_VERIFICATION,
and load the file CAL_VERIFY_genl_NWA_2port_MASTERFILE_demodata.mdl.
For the PNA, use the ModelFIleCAL_VERIFY_PNA_2port_MASTERFILE_demodata.mdl

Then, start The NWA Calibration From IC-CAP
IC-CAP file: go to
267
demo_features\3_MEAS_ORGANIZE_n_VERIFY_DATA\0_MASTER_FILES\20_NWA_CAL_VERIFICATIONand
stay with fileCAL_VERIFY_genl_NWA_2port_MASTERFILE_demodata.mdl Go to Setup
Calverif_open/OPEN, and hit the 'Calibrate' button there.
Note
usually, you need a separate NWA calibration for every different frequency sweep. When using Agilent
PNAs, and when using 'Use PNA Calibration Settings'=YES, this step (hitting the 'Calibrate' button) is not
required.
A typical modeling mistake: fT modeling at different (higher) iC

than DC modeling
Further Data Consistency Check Checks: CV <> Spar and DC <> Spar Consistency Checks
After we have discussed in details the data consistency checking within DC, CV and S-
parameter measurements, we want to close the loop and compare that data consistency
also against the measurement domains.
Comparing CV curves from S-par measurements with CV-

measurements
When comparing CV curves converted from S-parameters, keep in mind that due to the
S_to_Y conversion and the interpretation of these Y-parameters with respect to a PI
schematic, the CV curves represent the capacitance in one of the PI branches. With a CV
meter, depending on the measurement principle, you may obtain the total capacitance
between two nodes (2-pin CV measurement method) or the individual capacitance (4-pin
method, guarded wafer chuck).
Last not least: don't merge diff. measurements of diff dies when extracting model
parameters.
268
It is obvious that when developing a model of a component, all types of measurements

have to be performed on the same device !
269
S-Parameter Utilities
Calculation the radiation losses from S-parameters
The radiation losses for passive components can be calculated as:
Note
If considerable radiation occurs, e.g. with spiral inductors, conventional lumped schematics may no longer
be valid.
270
Understanding S-Parameter Plots
The Smith Chart for Sxx Parameters
What makes Sxx-parameters especially interesting for modeling, is that S11 and S22 can
be interpreted as complex input or output resistances of the twoport (including the
termination at the opposite side of the twoport with Z0).
That's why they are usually plotted in a Smith chart. This chapter is intended to explain
the basics of such Smith charts.
The Smith chart is a transformation of the complex impedance plane R into the complex
reflection coefficient (rho) following:
with the system's reference impedance Z0 = 50 .
This means that the right half of the complex impedance plane R is transformed into a
circle in the -domain. The circle radius is '1' (see the following figure).
Relationship between Sxx and the complex impedance of a twoport
On the other hand, using an network analyzer with a system impedance of Z0, the
parameter S11 is equal to:
where v1 is the complex voltage at port 1 and v01 the stimulating AC source voltage
(typically normalized to '1'). See the following figure and the chapter, Calculating S-
parameters from Voltages (iccapmhb) for details.
About the definition of S11, (S22 is just analogous)
Under the assumption that R is the complex input resistance at port 1 and Z0 is the
system impedance, we get using eq.(2) and the resistive divider formula:
And this is the reflection coefficient from (1) !!
After all, if the reflection coefficient resp. S11 or S22 is known, we get for the complex
resistor R:
, with usually Z0 = 50
This explains how we can get the complex input/output resistance of a twoport directly
from S11 or S22, if we plot these S-parameters in a Smith chart.
Let's go back to Figure, "Relationship between Sxx and the complex impedance of a
twoport" and consolidate this context a little further:
It shows a square with the corners (0/0), (50/0), (50/j50) and (0/j50) in the
complex impedance plane and its equivalent in the Smith chart with Z0=50. Please
watch the angel-preserving property of this transform (rectangles stay rectangles close to
their origins). Also watch how the positive and negative imaginary axis of the R plane is
transformed into the Smith chart domain (), and where (50/j50) is located in the
Smith chart. Also verify that the center of the Smith chart represents Z0, i.e. for Z0=50,
271
the center of the Smith chart is (50/j0).
The following explains once again the transformation of the complex ohmic plane to the
Smith chart.
How the complex resistance plane is transformed into the Smith chart
This allows us to make the following statements:
Sxx on the real axis represent ohmic resistors

Sxx above the real axis represent inductors
Sxx below the real axis represent capacitors
Sxx curves in the Smith chart turn clock-wise with increasing frequency.
As an example for interpreting Smith charts, The following figure shows the S11 plot of a
bipolar transistor. In this case, the locus curve stars with S111 = *Z0 at low
frequencies (RBB' + Rdiode + beta*RE). For higher frequencies, the curves then tend
towards |S11|{-1...0}
for high frequencies ( the CBE shorts Rdiode, and beta = 1. Therefore, the end point of
S11 is RBB' + RE ). Since RBB' is bias dependent, and decreasing with increasing iB, the
end points of the curves represent this bias-dependency. For incrementing frequency, the
S11 locus curve turns clockwise-
S11 of a transistor with increasing Base current iB
The following figure shows the S11 curve of a capacitor located between the NWA ports.
The capacitor represents an OPEN for DC, thus S11 ~ 1 = *Z0. For highest frequencies,
it behaves like a SHORT, and we see the 50 of the opposite port2 . The transition
between the DC point and infinite frequency follows a circle, and the increasing frequency
turns the curve again clockwise.
S11 of a capacitor between port 1 and port 2
272
THE POLAR PLOTS FOR THE S12 AND S21

PARAMETERS
The polar plot for S12 and S21
The S21 parameter represents the power transmission from port 1 to port 2, if the
twoport is inserted into a matching network with characteristic impedance Z0 of e.g. 50.
This means, if no voltage is transmitted, then S21=0 (in the center of the polar plot). If
voltage is transmitted, we are on the positive X axis. The curve will be below S21=1 for
damping between the port 1 and port 2, and above S21=1 for amplification. If the phase
is inverted, we are basically in the left half-plane of the polar plot (REALS21 <0). And, for
voltage amplification, but also phase shift of e.g.-180 degrees (a transistor), below S21=-
1.
Like with the Smith chart, all S21 and S12 curves turn clock-wise with increasing
frequency.
As an example, Figure, "S21 of a transistor with varying Base current iB" shows the S21
plot of a bipolar transistor, and Figure, "S21 of a capacitor between port 1 and port 2" of a
capacitor between port 1 and port 2. While the transistor starts with S21 <-1 at low
frequencies (voltage amplification in a 50 system), its curves tend towards S21=0 for
high frequencies (no voltage transmission, the transistor capacitances short all voltage
transmission). For the capacitor it's just the opposite: no power transmission for lowest
frequencies, but an ideal short (S21=1) for highest frequencies. For more details, see the
chapter on understanding S-parameters plots.
: S21 of a transistor with varying Base current iB
S21 of a capacitor between port 1 and port 2
273
274
De-embedding Techniques
Contents
Basics of De-embedding Techniques (iccapmhb)

De-embedding Techniques Tutorial (iccapmhb)
ft Modeling of a Transistor Affected by Parasitics (iccapmhb)
Interpreting OPEN-Dummy Measurements (iccapmhb)
Twoport Matrices of Basic Schematics (iccapmhb)
Twoport Matrix Conversions (iccapmhb)
Twoport Matrix Definitions (iccapmhb)
Verifying the De-embedding Procedure (iccapmhb)
275
Basics of De-embedding Techniques
Why De-embedding ?
Example:
Chip Carrier of a Packaged Transistor and its Parasitic Components
Chip carrier
The bipolar transistor is mounted with its Collector to the carrier, Base is bonded, Double-
Emitter contact (bonded twice) goes to the backside of the carrier.
This corresponds to:
Converting the S-Parameters to Y, we can strip-off

276
the capacitors
then, converting to Z parameters
and de-embedding the inductors
gives the inner transistor
At a glance:
A note on adding of Y matrices
277
A typical representation of Matrix Yparas
A note on adding of Z Matrices
Applied to modeling, matrix Zparas represents a TEE structure:
278
This means:
Either model the parasitics by a lumped components circuit and de-embed the
individual components
or, applied to on-wafer measurements:
de-embed the whole matrix, if it correctly describes all parasitic effects that is, the Y
matrix includes only parallel parasitic components, and the Z matrix only series
parasitics
For on-wafer OPEN dummy structures, Yparas looks

like this:
The Lumped Components Values are:

Convert Y -> Z and calculate
Zxy = ( Z12 + Z21 ) / 2
A = Z11 - ZxyB = Z22 - ZxyC = Zxy
279
Physical Interpretation of the OPEN dummy
The so far proposed OPEN schematic is usually ok up to a few GHz. For higher frequencies,
the following schematic is most often appropriate:
Reminder on the de-embedding of the complete Y

matrix
280
For on-wafer SHORT dummy structures, Zparas looks

like this:
about the de-embedding of complete Y and Z

matrices
De-embedding of complete Y and Z matrices
281
Required measurements
'TOTAL' device
'OPEN' dummy device
'SHORT' dummy device.
De-embedding procedure
convert S_total -> Y_total,
and calculate Y1 = (Y_total - Y_open)
to de-embed the DUT from the parallel parasitics
convert Y1 -> Z1,

and calculate Z2 = (Z1 - Z_short ) to de-embed further the serial parasitics,
finally re-convert Z2 -> S_deemb
to obtain the 'inner' DUT.
Note:
matrix Z_short has to be de-embedded first from the outer parallel parasitics.
This is done by:
convert
S_short_incl_open -> Y_short_incl_open,
calculate
Y_short = (Y_short_incl_open - Y_open).
and convert
Y_short -> Z_short
After so much theory, a look at the reality
This slide depicts a three-dimensional representation of such a multi-finger MOS

transistor. It is interesting to note how 'high' the metal planes are relative to the active
silicon area and that the size of the transistor itself is very small compared to the size of
the pads.
The goal of de-embedding is to shift the NWA calibration plane, which is (after the NWA
calibration) at the end of the probe tips, down to the beginning of the transistor. This
beginning is one of the key points for a good de-embedding and a key point for good
dummy structures
The next slides depict the corresponding dummy structures.
282
In the above slide, in the section 'Magnification of the transistor' is the inner transistor,
which we want to model. All the rest has to be de-embedded, i.e. to be stripped-off. In
other words, the NWA calibration plane has to be shifted down here.
Note that the limits of the blue surrounded area in the OPEN dummy will become the
shifted calibration plane. All what is inside will become part of the transistor model!
283
The SHORT dummy again refers to the desired shifted calibration plane. All what is inside
this plane is now filled up with metal, so that we can consider this part to behave ideal,
while the striplines from Gate and Drain will behave like inductors.
This is the THRU dummy, for de-embedding verification purpose. See the next slides for a
proposed de-embedding verification procedure..
284
This layout suggestion for Silicon on-wafer dummy structures uses metal1 to shield the
contact pads from the lossy silicon substrate.Due to this metal1 shielding, the measured
OPEN dummy S-parameters represent a much more ideal OPEN dummy schematic than
the more conventional layout without the metal1 shielding.
Also, for the SHORT dummy, the short itself is much more ideal because of the chain of
shorting vias. This chain is geometrically concentrated and precisely localized, compared
to the conventional SHORT dummy layout there the 'hole' is simply filled up with metal1.
Note
For clarity, the layout details are simplified in the slide above. A correct layout would be composed of
conventional contact pads and a stripline each towards the component input and output, like for
conventional layouts. So, the main difference between the layout sketch above and 'conventional' layouts
is to use the metal1 as a shield fo the signal pads and lines against the lossy silicon substrate.
See publication T.E.Kolding, O.K.Jensen, T.Larsen, "Ground-Shielded Measuring Technique

for Accurate On-Wafer Characterization of RF CMOS Devices", IEEE Int.Conf.on
Microelectronic Test Structures ICMTS, March 2000, Monterey,CA
Verifying the de-embedding for on-Wafer transistor

measurements
The Idea:
Have an OPEN, a SHORT and a THRU dummy measured
Model the dummies in order to check the previously mentioned prerequisites
de-embed the THRU from the OPEN and SHORT and model it
We should get a simple lumped schematic consisting of essentially a delay line which
corresponds to the THRU on the wafer
Model The On-Wafer OPEN Dummy
285
In this first step, the OPEN dummy is modeled in order to verify that there are no serial
components included in the Y matrix of the OPEN.
Hint:
In the demo_features directory, load the file
3_MEAS_ORGANIZE_n_VERIFY_DATA\0_MASTER_FILES\3_DEEMB_VERIFICATION\CHECK_DEEMB_OpenShortThru_MASTER_PELdep.mdl
Model The On-Wafer SHORT Dummy
Then, the SHORT is de-embedded from the OPEN and again modeled. THe goal is to verify
that there are no parallel components included in the Z matrix of the SHORT.
Verify The De-Embedding With Modeling The THRU

Dummy
After the OPEN and (if available/required) SHORT dummy measurements have been
modeled, we want to check the de-embedding procedure with a known 'golden device'. In
practice, however, like with the NWA cal.verification, the problem is the availability of an
'golden device'. Provided we have a THRU dummy, we can use this measurement for a
verification of the de-embedding procedure. The idea is that the THRU should look like a
simple delay line, with a certain characteristic impedance Z0, and a delay time TD.
Additionally, the delay line may be lossy. But in reality, due to the short geometry of the
THRU, this is typically not the case. Anyhow, the model parameters must represent
physical meaningful values, and the de-embedded trace must be simple without
resonances etc: just a simple delay line!
TD: models the phase in Sxy

Z0: models the Sxx trace:
for Z0<50W, the curve starts at 50W, turns downwards and then to the left into
a circle around the line's Z0. The turning is clock-wise with increasing frequency!
for Z0=50W, the curve is a single point at at 50W in the Smith chart.
for Z0>50W, the curve starts at 50W, turns upwards and then to the right into a
circle around the line's Z0. The turning is clock-wise with increasing frequency!
Hint:
286
In the demo_features directory, load the file
3_MEAS_ORGANIZE_n_VERIFY_DATA\0_MASTER_FILES\3_DEEMB_VERIFICATION\CHECK_DEEMB_OpenShortThru_MASTER_PELdep.mdl
and model the THRU with the tuners in Setup THRU.
Only if there were no layout bugs or measurement problems with the Dummy structures,
we will end up with a physical representation of the small THRU strip line. Therefore, when
we obtain that physical strip line, we can assume that the de-embedding, applied later to
transistors, diodes etc., will be correct as well!
Note
Since the THRU strip line is passive, and not depending on bias or frequency or temperature, it can be
used much better for de-embedding verification than checking the de-embedding with a transistor, a diode
etc.
As expected, the THRU dummy can be modeled by a delay line. Therefore, we can assume
a correct de-embedding procedure, that is, the inner transistor will now be de-embedded
correctly.
This means:
After we are sure about the de-embedding, we apply:
to all measured S-parameters for a correct and verified de-embedding.
An Example of bad THRU deembedding results
The slide above represents a de-embedding of dummy structures with obviously layout
problems.
Although the inspection sequence undeembedded -> open_deemb -> open_short_deemb
exhibits the right sequence (the resulting THRU becomes more and more the expected
simple, short THRU), the open_short_deemb results does not match the expected the
modeling target and represents - a strip line with negative delay time (see Sxy: a short
delay line with neg. delay time and even with a trend for MAG(Sxy)>1 with increasing
frequency) - and an unphysical Z0 (since in Sxx, it turns counter-clock-wise).
De-Embedding Conclusions
De-Embedding is used to strip-off high frequency parasitics from the DUT
In order to determine the required de-embedding steps and to also verify the de-
embedding process, it is recommendable to also have a THRU dummy available on the
wafer.
Try yourself in IC-CAP

Hit the blue Examples icon,select the demo_features directory and go further to
subdirectory
3_MEAS_ORGANIZE_n_VERIFY_DATA/0_MASTER_FILES/31_DEEMB_VERIFICATION
and load the file CHECK_DEEMB_OpenShortThru_MASTER_demodata_PELdep.mdl
Note
You need to load the additional file DEPOTS.mdl, which is located under demo_features directly.
287
De-Embedding Techniques Tutorial
High frequency measurements are always influenced by the parasitic components (contact
pads, packaging, test fixture). This makes the determination of the model parameters of
the 'inner device' often pretty complicated. In some cases it is possible to measure and
characterize the parasitic components of the contact pads, of the packaging or of the test
fixture alone. In other cases we might have to use 'hot/cold' measurement techniques in
order to separate the parasitic components from the 'inner device'. Once the values of the
parasitic components are known, their effect on the total measurement can be eliminated
by matrix operations and we get the s-parameters of the 'inner device'. This is explained
in the following tutorial.
De-embedding tutorial
Assumed, we are going to model a transistor chip mounted on a chip carrier of a transistor
case. The chip carrier will distort the 'inner' transistor's performance.
Yet, the chip carrier can be described with series inductors of the bond wires and the
parallel capacitors of the pads (See the figure below).
Figure1: Chip carrier of a bipolar transistor case and its parasitic components
From Figure1, we obtain the simplified equivalent schematic given in Figure2.
Figure2: The equivalent schematic of the chip carrier for high frequencies.
About the de-embedding

The de-embedding procedure is sketched below. It means stripping off the parasitic
components from the outside towards the inner DUT.
Stripping off the parasitic components gives the performance of the inner transistor.
De-embedding step by step:

We have measured the S-parameters Stotal of the transistor and its package. We assume
that the components have been modeled already. (Note: de-embedding by complete
matrix subtraction and not lumped components subtraction is covered further below).
Following our assumption of having parallel capacitors as 'outer' parasitic components, we

transform the S-parameters to Y, because a Y matrix represents a PI structure of
components. A simple subtraction will de-embed the parasitic capacitor effects.
288
Figure 3: The equivalent schematic of fig.2 after the de-embedding of the capacitors
Now, the 'outer' parasitic components are the two inductors, which are in series with the
chip connections. Series parasitics can be easily eliminated by subtracting a Z matrix.
Therefore, we transform the Y-parameters from above into Z-parameters and subtract the
inductors.
Figure 4: The equivalent schematic of fig.3 after the de-embedding of the series inductors
These Z-parameters are finally transformed back into S-parameters which describe the
performance of the 'inner' chip.
MATRIX MANIPULATIONS IN DETAILS

Note: the mentioned IC-CAP model files can be found in $ICCAP_ROOT\examples
and then
demo_features\3_MEAS_ORGANIZE_n_VERIFY_DATA\0_MASTER_FILES\3_DEEMB_UTILITIES
In the following pages, we will discuss the matrix properties and manipulation schemes
relevant for de-embedding.
DE-EMBEDDING OF SERIES PARASITICS (Z-MATRIX)
Adding these matrices of parasitic components gives:

Zparas = ZParas1 + ZParas2 (2)
289
or:
what is the matrix of a TEE structure (!!).
How to proceed with the de-embedding of series parasitics:

Once the parasitic components A, B and C are known, they can be eliminated by
subtracting the Z-matrix Zparas = ZTEE of a TEE structure from the measured data
Ztotal. We obtain:
(1) (2)
ZDUT = Ztotal - ZParas1 - ZParas2 = Ztotal - Zparas
DE-EMBEDDING OF PARALLEL PARASITICS (Y-

MATRIX)
ADDING OF Y-MATRICES:
THE Y-MATRIX OF THE PARALLEL PARASITIC COMPONENT IS:
How to proceed with the de-embedding of parallel parasitics:

Once the parallel parasitic components A, B and C are known, they can be eliminated by
simply subtracting the Y matrix YPI of a PI structure from the measured data Ytotal and
we obtain:
(1)
YDUT = Ytotal - YParas
DE-EMBEDDING BY SUBTRACTING THE Z AND Y

MATRICES
without modeling their components
IC-CAP files: 3_z_y_deemb.mdl
3_pad_short_open.mdl
In case, we are sure that the connections to the DUT do not suffer from any inductive
losses and the 'inner' parasitics are all in parallel with the DUT, we can use a simple Y
matrix subtractions for the de-embedding of the DUT. In the other case, when the
parasitics are more complex, but still can be separated into a full parallel and another full
serial part, a Y and Z matrix subtraction can be applied. In order to determine the
components of these matrices, we need a measurement condition where the DUT behaves
like a short (either a transistor in full saturation or a dummy device with a SHORT) and
another one in which it behaves like an OPEN (a transistor without any bias voltage or an
OPEN dummy device).
As stated above, it is assumed that the SHORT contains no components in parallel with
the DUT (like capacitors) and that the OPEN includes no components that are in series
with the connections of the DUT (inductors).
For the case of parasitics only in parallel with the pins of the DUT, we have the following
situation:
290
The de-embedding is performed by simply subtracting the Y matrix of the parasitics from
the total measurement. This is usually o.k. up to 10GHz.
For higher frequencies, we must also consider the parasitics which are in series with the
DUT. This is illustrated in fig.2. Again, it is assumed that there are only series parasitics
included in the Z matrix, and nothing but parallel parasitics in the Y matrix. The depicted
situation with the complex resistors is a typical one for on-wafer transistor measurements.
If these conditions are true, the DUT can be embedded like follows, requiring these
measurement results:
'total' device
'short' dummy device
'open' dummy device.
The de-embedding of the 'inner' device-under-test (the DUT) from the parasitics is then:
Convert S_total to Y_total and calculate
Y_total_w_short = (Y_total - Y_open)
to de-embed the outer parallel parasitics.
Then, convert Y_total_w_short to Z_total_w_short and calculate
Z_dut = (Z_total_w_short - Z_short_wo_open)
to de-embed the inner serial parasitics.
Finally, re-convert Z_dut to S_dut to obtain the S-parameters of the 'inner' DUT.
Important Note:
matrix Z_short has to be de-embedded first from the outer parallel parasitics, before
subtracting it in the second step above.
To do this, convert S_short to Y_short and calculate
Y_short_wo_open = (Y_short - Y_open)
HOW TO DETERMINE THE SEQUENCE OF Y AND Z

MATRIX DE-EMBEDDING
In the example above, the serial parasitics are assumed to be close to the DUT, while the
parallel parasitics are around them. This is typically the case for on-wafer transistor
measurements. Here, usually, the parallel parasitics, the pad capacitors, are the outer
ones, while the small lines to the transistors, represented by series inductors are the inner
ones.
For packaged transistors, this is usually just the other way around.
In order to check the sequence of de-embedding or also to determine if an OPEN

de-embedding alone is sufficient, it is best to also have a dummy THRU structure. Such a
THRU is represented by a delay line. If the THRU after the de-embedding represent a
291
simple delay line (to compared with a simulated delay line, the de-embedding procedure
can be considered o.k.
Another method, however not always successful, is to convert the OPEN dummy
S_parameter to Z and to plot MAG(Zxx) and MAG(Zxy) versus the frequency. If there is
no rise in magnitude, it can be assumed that there are no hidden inductors in the OPEN
structure. For the SHORT, a check of MAG(Yxx) and MAG(Yxy) might give some more
insight.
DE-EMBEDDING OF LOSSLESS 50 Ohm DELAY LINES

(S-MATRIX)
Assumption: we are in the S-parameter domain and have the device-under-test (DUT)
imbedded between two delay lines:
with a: forward wave

b: reflected wave
In this case, there is b1 = s11total*a1 + s12total*a2
b2 = s21total*a1 + s22total*a2
and the S-parameter matrices of the individual delay lines are:
How to proceed with the de-embedding of lossless delay lines:

Provided we know the delay times TD1 and TD2, we can use the 'port shift' properties of
S-parameter multiplications (shift of reference planes), and obtain:
with: f : frequency
TD1: delay time at port 1,
TD2: delay time at port 2
DE - EMBEDDING USING ABC - MATRICES

IC-CAP file: 6_a_deemb.mdl
This kind of de-embedding is related to measurement conditions where we have

considerably long connection lines to the device. It can be applied for example to
insertable devices like a connector etc., but also to special components line spiral
inductors. Besides the test fixture for the DUT, which consists basically of a substrate with
a strip line for each port of the DUT, we also need a special test structure to characterize
the strip line.
This auxiliary fixture consists of a strip line on the same substrate like the test fixture, and
the strip line is exactly as long as on the test fixture. In this case, we can apply ABC-
matrices (A-matrix, cascade matrix) to perform the de-embedding. The sketch below
shows the required substrates for a connectorized situation
and for a on-wafer situation:
292
With the assumption that the Aline ABC matrix of the auxiliary structure is identical to the
connection strip lines on the wafer or on the test fixture (in a connectorized scenario, the
connector has been de-embedded !), the total performance of the device-under-test
(DUT) including the test fixture can be expressed in A matrices:
or solved for ADUT (multiplication 'at the left' with , and 'at the right' with:
How to proceed:
The measured S-parameters Stotal are transformed to Atotal using IC-CAP's TwoPort
function. The same applies to the measured Sline parameters. Then the matrix calculation
from (2) is applied in order to obtain the de-embedded ADUT.
IMPORTANT NOTE: when cutting an electro-magnetic effect into a chain of effects, make
sure to set the cut where the E-fields have a monotonous trace (where the EM effects of
geometrical discontinuities have died out). As an example, when isolating a via from a
stripline-via-stripline scenario, the via should be cut off from the strip line not at the
location of the via, but rather including a bit of the strip lines.
DE - EMBEDDING OF CHAINS OF ABC - MATRICES

IC-CAP file: 6_chain_of_a_matrices_deemb_verif.mdl
In case of the de-embedding of an 'inner' A-matrix, embedded within a chain of A-

matrices, we need to proceed as follows, regarding the inverse multiplication of A-
matrices:
Let's assume we have the following situation:
IMPORTANT NOTE:
watch the flipped sequence (with respect of the original chain of twoports !!) of the
inverse A-matrices chain !!!
DE-EMBEDDING OF LOSSY DELAY LINES WITH

GENERAL CHARACTERISTIC IMPEDANCE
This is, again, an application of ABC-matrices (A-matrices).
Knowing the de-embedding procedure for A matrices shown above, we now can apply it to
the de-embedding of striplines with general impedance Z0. Compared to the S-parameter
de-embedding above, the stripline can also be lossy in this case.
293
NOTE: Delay Lines in ADS are specified by length L, so for a lossless delay line, we can
substitute TD = L / 3E+8
SPLITTING A THRU INTO 2 SYMMETRICAL MIRROR HALVES
or
OPEN -THRU DE-EMBEDDING
APPLYING DIFFERENTIAL TWOPORT-PARAMETERS
See demo_features\3_MEAS_ORGANIZE_n_VERIFY_DATA\0_MASTER_FILES\
31_DEEMB_VERIFICATION\CHECK_OPEN_THRU_DEEMB_demodata_MASTERFILE.mdl
This de-embedding idea is after

Jon Tao, Paul Findley and G. Ali Rezvani, Novel Realistic Short Structure Construction for
Parasitic Resistance De-embedding and on-Wafer Inductor Characterization, Proc. IEEE
2005 Int. Conference on Microelectronic Test Structures, Vol. 18, April 2005
It is especially interesting for the de-embedding of spiral inductors.
The dummy structures should look like this:
Sketch of the OPEN dummy structure ('A' marks the end of line1, 'B' the end of line2)
and of the THRU dummy structure ('C' marks the center of the THRU line)
When applying differential TwoPort measurements to a THRU dummy, we obtain a virtual

Ground in the middle of the THRU dummy.
The center of the THRU becomes a virtual ground,
i.e. the differential TwoPort parameters of the THRU represent a virtual SHORT.
Measuring a THRU instead of a SHORT avoids
S-parameter measurement problems due to the near-zero impedance.

The uncertainty of the impedance measurement of the conventional SHORT dummy
structure, where it is assumed to measure from pad1 to A and also from pad 2 to B
the
series impedance of the SHORT, while the shortening from A -> Ground
and B -> Ground is assumed to be ideal, i.e. 0 Ohm.
After the conventional OPEN de-embedding, and based on the above layout sketches, the
differential impedance of the THRU structure can be calculated as
294
Zdiff = 1/2 ( Z11 + Z22 - Z12 - Z21)
i.e. the impedance of each port to the center virtual ground point (C) is:
Zc = 1/2 Zdiff
We now need to consider the electrical performance between points A and C of the THRU
(see also the sketch above):
Assuming the length of the metal stripe between point A (the later connection point of the
spiral inductor) and point C (the center point of the THRU) is L, and the metal width is W
(L and W can be measured from the layout), and T is the metal thickness, and assuming
the specific resistance of the metal layer is rho , the resistance of the line between A and
C can be calculated as
and for the special case rhoAluminum ~ 0,0264 um/um :

RAC ~ 0.0264*L[um] / (W[um]*T[um])
Similarly, the inductance between A and C can be estimated.

For a strip line width W > height of dielectric (H), and line length L, it becomes
approximately
Therefore, the Z-parameters of the virtual SHORT structure can be calculated as
By substituting the Z-parameters of the conventional SHORT dummy structure with the
Z-parameters of (1), we can apply the well-known OPEN-SHORT de-embedding, as it is
described in the chapter 'DE-EMBEDDING BY SUBTRACTING THE Z AND Y MATRICES'
further above.
Note
for strip line width W < height of dielectric (H), and line length L.
Note
the conversion of the THRU to a virtual SHORT can also be applied to verify the quality of the de-
embedding (by comparing to the real SHORT measurement). Of course, in both cases, the OPEN has to be
de-embedded from the THRU and also from the SHORT.
295
ft Modeling of a Transistor Affected by
Parasitics
Tutorial About the f t Modeling of a Transistor Affected by Parasitics
An introduction to de-embedding
In order to demonstrate the importance of de-embedding for proper s-parameter
modeling, this chapter covers the fT modeling of a bipolar transistor with and without de-
embedding.
Note
To demonstrate the de-embedding techniques in a well-defined situation, all data in this chapter come
from simulations.
The individual steps of this tutorial are:
1. Inspection of the cutoff frequency curve of a bipolar transistor without any parasitic
effects.
2. The same curve, but now distorted by package effects.
3. The de-embedding procedure
4. The de-embedded cutoff frequency curve that matches the one of step 1.
The cutoff frequency of a bipolar transistor chip is depicted in fig_1. It refers to the 'inner'
transistor and was obtained from a simulation.
In reality, we would not be able to measure the transistor under test in this manner.
Assuming a packaged bipolar transistor, fig_2 shows some typical parasitic elements of
such a package.
These parasitic elements will degrade the performance of the 'inner' DUT at high
frequencies, as given in fig_3.
Our goal is to de-embed the curve of fig._3 and to get back that one of fig_1. Provided the
parasitic elements are known (what will be explained in the chapters below), we have to
perform the following sequence of matrix operations in order to de-embed the 'inner'
transistor:
296
Referring to fig_2, we commence with these pre-considerations:
The inductors are in series with the transistor pins. Therefore they can be represented
best by a Z-matrix.
The capacitors are parallel to the DUT. They can be represented best by a Y-matrix.
This leads to the de-embedding procedure:
1. convert the measured 'distorted' S-parameters of fig_3 to Z-parameters

2. substract the known parasitic inductors of the SHORT (Z-matrix)
3. convert the Z-parameters to Y-parameters
4. substract the known parasitic capacitors of the OPEN (Y-matrix), which itself is de-
embedded from the parasitics of the SHORT!
5. finally convert the Y-parameters back to S-parameters and obtain the data of fig_1
Table_1: the de-embedding sequence
Hint: see also chapter De-embedding Techniques with Z,Y,S,A Matrices
297
Interpreting OPEN-Dummy
Measurements
This slide sequence is about the inspection of measurements of different OPEN dummy
structures and about the interpretation of possible underlying physical layout problems.
Effect of R1, C1 and R2, C2 on Sxy
Effect of R12 and C12 on Sxy
Effect of R3 on Sxy
Effect of L3 on Sxy
298
Note
The necessity to include an inductor for modeling an OPEN dummy structure is a contradiction of the fact
that all components of an OPEN dummy schematic are in parallel to the device, and therefore, can be de-
embedded by subtracting the Y matrix of the OPEN dummy.
QUIZ:
and what schematic can be assumed behind this OPEN Sxy measurement??
Note
If your OPEN measurement exhibits a trace as depicted above, it is pretty possible that the THRU de-
embedding verification will fail.I.e. that the de-embedded THRU may have a MAG(Sxy) with more than 1,
or may turn backwards (to the left), and that the Sxx may also turn backwards.
Open Dummy High Frequency Modeling
299
Note
Such a trace is pretty typically when not applying a metal-1 shield below the dummy layout structures.
Note
Such a trace is pretty typically when not applying a metal-1 shield below the dummy layout structures.
And if Sxy extends into 4th and into 3rd quadrant, its probably the bulk effect.
300
How the Silicon Substrate and the Silicon Oxide

parameters affect the 'knee' in the OPEN Dummy
Structure S11
A study on the effect of the oxide loss tangent and

the substrate conductivity
301
The signal pad of the OPEN sees basically a sequence
of SiO2 > silicon substrate > SiO2, i.e. a circuit
schematic like below
Interpreting the OPEN Sxx related to substrate and

oxide params
302
Twoport Matrices of Basic Schematics
Note
Impedance Z or Conductance Y
Z-matrix not applicable,
Y-matrix not applicable,
Z-matrix not applicable,
303
304
Twoport Matrix Conversions
H to Y Conversion
Since a TwoPort matrix describes the behavior of its circuit completely, the matrices can
be converted from one form into another. As an example, to convert a H matrix into its Y
equivalent, we just have to rearrange the matrix equations. The H matrix has the form.
and the target Y matrix

The v1 dependency, i.e. the first line in the H matrix, can be solved for i1:
, what is already the first line of the Y matrix.

We insert this result into the second line of the H matrix in order to replace the i1
dependency,
with
we can rearrange and yield,
what finally gives the relationship between the y and h parameters:
Y to Z and Z to Y Conversion
The conversion of the Y to the Z parameters is, related to the special definition of these
two matrices, very simple.
Starting with the resistance matrix form,
we can solve it for ( i ) by multiplying to the left side with ( z )-1, the inverse matrix of ( z
).
or Comparing this result with the definition of the y matrix, we get finally,
and correspondingly,
Note
Remember the matrix inversion scheme.
with ad - cb = det(matrix),
S to H Conversion
S to H conversion (S2H):
Matrix Conversions at a Glance

A Matrix
305
Z Matrix
Y Matrix
H Matrix
TEE to PI Conversion
The following conversion equations explain how to convert a TEE structure to a PI
structure.
With the impedances Z and the admittances Y of the above TEE and PI schematics, the
conversion equations are as follows:
306
307
Twoport Matrix Definitions
Terms and Definitions on Series and Parallel Circuits
Series circuit Parallel circuit
Impedance Admittance
Resistance R Conductance G=1/R

Reactance X Susceptance B=1/X
Immittance: A general term for both impedance and admittance, used when the
distinction is irrelevant.
Twoport Matrix Signal Flow Definition

Note
The following matrix conventions refer to their implementation in IC-CAP.
H Matrix
Z Matrix
Y Matrix
A matrix
S Matrix
308
K Matrix
309
Verifying the De-embedding Procedure
Verifying the de-embedding procedure is very important before applying it to the very
DUT, i.e. the transistor or the passive RF component etc.
Without this step, errors or problems associated with the de-embedding will add to the
performance of the inner, de-embedded device, and, thus, lead to a wrong device model.
This is especially critical when simply de-embedding (subtracting) the complete Y and Z
matrices of the OPEN and SHORT dummy, because, in this case, possible problems may
not show up compared to de-embedding of lumped components. In this later case, we
would have detected such problems because they would show up with non-physical
lumped circuit components of these OPEN and SHORT dummies.
Therefore, it is suggested to
Model every dummy structure before simply subtracting their total matrices in order
to verify its de-embedding prerequisites:
and to
Verify the de-embedding procedure with a well-known 'golden device'.
before applying this procedure to the very DUT.
The pre requisite for a correct de-embedding is that certain test structures are
available on a wafer together with the device under test (DUT) itself. Depending on
the selected de-embedding method, an OPEN and SHORT dummy structure is
required and must be measured. For the proposed de-embedding verification from
above, also a THROUGH dummy structure is necessary.
The principle layouts of these structures are given in the figures shown below. These
layouts are for Ground-Signal-Ground Probes (GSG).
Please check the calibration plane as marked in the figures. Every part of the DUT
included in this calibration plane will become a part of the DUT model!
In other words, you can think of de-embedding as a shift of the current calibration plane
to these new borders on the wafer.
Example of a layout of an OPEN dummy structure.
Example of a layout of a SHORT dummy structure
Example of a layout of a THRU dummy structure.
310
Note
Instead of the 'classical' layout shown above, today's layouts for silicon wafers look more and more often
like the one shown in the next figure below. It avoids the effects of the lossy silicon substrate by using the
1st metal plane to shield the contact pads and the lines to the DUT against the lossy silicon substrate.
Especially interesting is the drastically improved SHORT dummy performance, since it applies a series of
vias to ground (metal 1) at the end of the SHORT, rather than having - as with the SHORT depicted
above- a large metal plane which is considered as ideal, while the two microstrip lines of the SHORT
dummy are considered as non-ideal. Also, the OPEN is much more ideal with this alternate approach, i.e.
its measurement data do not exhibit the second half-circle effect like with the conventional layout from
above,i.e., the OPEN can now be modeled up to several 10 GHz with simply the inner schematic of the
OPEN circuit schematic as discussed in the next section.
Altogether, this layout suggestion of using the 1st metal as a shield has proofed to give
much more robust de-embedding results.
Modeling of the Open Dummy

In this first step, the OPEN dummy is modeled in order to verify that there are no serial
components included in the Y matrix of the OPEN. See fig.1.
Note
Note: if there were parasitic components in series with the DUT, the subtraction of the Y matrix cannot be
applied, and would lead to wrong OPEN-deembedding results.
311
Vector Network Analyzer - Basics for
Modeling Engineers
312
313
Measurement Errors
Systematic errors
Due to imperfections in the analyzer and test setup are assumed to be time invariant
(predictable)can be characterized (during calibration process) and mathematically
removed during measurements.
Random errors
Vary with time in random fashion (unpredictable)cannot be removed by calibration
main contributors:
instrument noise (source phase noise, IF noise floor, etc.)
switch repeatability
connector repeatability
Drift errors
These errors are due to instrument or test-system performance changing after a
calibration has been done are primarily caused by temperature variation can be removed
by further calibration(s).
Vector Error Correction

accounts for all major sources of systematic error
What to do for Vector-Error Correction?

Full 2-port calibration
(reflection and transmission measurements)
12 systematic error terms measured usually requires 12 measurements on four
known
Standards (SOLT)
Standards are defined in cal kit definition table. These cal kit definitions are entered
to the network analyzer
CAL KIT DEFINITION MUST MATCH ACTUAL CAL KIT USED!
314
Practical Considerations before starting with NWA

Measurements
Enter calkit data (non-idealities of the cal. standards) into the NWA
Enter your frequency range etc. and Calibrate the NWA and generate the calset for
the 12 term error correction (to be used for meas.data correction)
Verify the calibration
MEASURE
if required , de-embed the DUT
SOLT calibration for Ground-Signal-Ground probes

(G_S_G)
315
All standard calkit data must be accurately entered into the VNA. SOLT is sensitive to
probe placement (probe tip instance at THRU).
Other Calibration Methods

We know about Short-Open-Load-Thru (SOLT) calibration...
What is TRL, LRM etc ?
Different two-port calibration techniques

Use the same 12-term error model as the common SOLT cal
Use calibration standards that are easier to fabricate and characterize
Can give better calibration results than SOLT for frequencies >20GHzbut may have
limited bandwidth (TRL)
Two variations:TRL, LRM (require 4 receivers) and TRL*, LRM*(only three receivers
needed)
Verify the Calibration

E=m*C 3dB !!?
Verify the Calibration
Summary
Network Analyzer measurements for modeling need
Good hardware: VNA, probes, calkits

Good understanding of S-parameters
Specific chip layout
316
Nonlinear Network Analyzers
Contents
Nonlinear De-Embedding (iccapmhb)

Nonlinear RF Measurements and Device Modeling (iccapmhb)
317
Nonlinear De-Embedding
Measurements of RF power devices can best be performed using a non-linear vector
network analyzer (NVNA), also called a large signal network analyzer (LSNA). Such
instruments offer the unique capability to measure, at the input and output of any two-
port device under test, both magnitude and phase of voltages and currents in the
frequency domain:
At the fundamental frequency

and a specified number of harmonics.
From these results, the corresponding waveforms in the time domain can easily be
derived. This large-signal RF information represents the entire device RF behaviour for the
given stimulus.
While for packaged devices, when using a test fixture with in-fixture calibration standards,
the obtained large signal measurements represent - as desired- exactly the packaged
device with its inner circuit plus the package parasitics, on-wafer measurements have to
be de-embedded from the probe contact pads and the strip lines to the inner DUT.
Generally, de-embedding is associated with S-parameter measurements. Many such de-
embedding procedures are published in the literature. However, only little is found on how
to de-embed large-signal (nonlinear) vectorial RF (radio frequency) measurements.
And, in fact, the de-embedding of non-linear RF measurements is substantially different

from S-parameter de-embedding, as can be seen in the following figure:
Figure: The different (Open-Short) de-embedding methods:
De-embedding procedure for linear S-parameter measurements: TwoPort matrix

manipulations
De-embedding procedure for large-signal non-linear RF measurements
Calculation of device voltages and currents out of measured voltages and currents
case a) the de-embedded DUT is considered to behave linear (bias-dependent, but linearized in the
operating point)
case b), the de-embedded DUT is still non-linear (may be large-signal-driven in its operating points).
2-Step Nonlinear De-Embedding

In order to do device modeling for on-wafer measurements with the nonlinear network
analyzer, the measurement has to be de-embedded
from the on-wafer contact pads

and the strip lines from these pads to the device itself.
Different to the de-embedding process for S-Parameters, we cannot simply apply the
conventional Y-parameter or Z-parameter subtractions. The reason is that S-
parameters are linear, and that all the twoport matrix theory is also based on linear
circuits. For non-linear devices like transistors or diodes, this means a small-signal
excitation around a DC operating point, and no harmonic frequencies.
When applying a nonlinear network analyzer, these prerequisites of the mentioned linear
twoport theory are not fulfilled. We need to handle the complex (magnitude and phase)
voltages and currents. However, since the OPEN and SHORT dummies represent linear
circuits, we can use their Y- and Z-matrix elements in the de-embedding calculations.
Note that these matrix elements are represented by complex numbers, versus frequency.
And these frequency points must match the fundamental frequency and its harmonics
from the nonlinear network analyzer measurement.
Also note, that we can apply them 'as they are' and do not need to care about their
underlying inner schematic in the de-embedding procedure presented below.
The proposed 2-step de embedding procedure of the nonlinear measurements is based on

the conventional OPEN and SHORT dummy structures.
318
Figure: OPEN-SHORT de-embedding for large signal RF measurements.
The de-embedded large signal currents and voltages (i1dut, i2dut, v1dut, v2dut) are
calculated from the measured ones (i1, i2, v1, v2) after:
and re-arranged:
where y ij are the admittance parameters of the OPEN dummy Y-matrix, and z ij are the
impedance parameters of the SHORT dummy Z-matrix after the de-embedding of the
OPEN dummy effects.
Note
Both, OPEN and SHORT structures, are reciprocal, i.e. y12 = y21 and z12 = z21. Therefore, we can apply
the mean values.
and
to reduce measurement noise.
IC-CAP demo file

demo_features\3_MEAS_ORGANIZE_n_VERIFY_DATA\0_MASTER_FILES\30_DEEMBEDDING\5_nonlinear_deemb_OPEN_SHORT.mdl
Special acknowledgements regarding the de-embedding procedure: Giovanni
Crupi, Dipartimento di Fisica della Materia e Ingegneria Elettronica, University of
Messina, 98166 Messina, Italy.
Giovanni Crupi, Dominique Schreurs, Electronic Engineering Department,
Katholieke Universiteit Leuven, 3001 Leuven, Belgium
Publications
G.Crupi, D.Schreurs, D.Xiao, A.Caddemi, B.Parvais, A.Mercha, S.Decoutere:
"Determination and Validation of New Nonlinear FinFET Model Based on Lookup
Tables", IEEE MicrowaveAnd Wireless Components Letters, Vol. 17, No. 5, May
2007.
3-Step Nonlinear De-Embedding

Note
This chapter is based on section D of the publication, E. Vandamme, D. Schreurs and C. van Dinther,
"Improved three-step de-embedding method to accurately account for the influence of pad parasitics in
silicon RF test-structures," IEEE Trans. Electron Devices, vol.48, no.4, pp.737-742, 2001.
The 3-step de-embedding is based on four Dummy structures which are all used in the de-
embedding process. The quality of the obtained de-embedding relies on the fact if the
individual dummies can be represented by the underlying, assumed schematics. The
dummies and their assumed schematics are shown in the following figure.
Note
For silicon, the layout using metal1 as a ground shield against the lossy Silicon substrate is preferred over
the conventional layout using a single metal layer. The short circuit of the Short Dummy is achieved by a
chain of vias from the upper metal to metal1 (ground).
319
Figure: Layout of the DUT and details of the de-embedding dummy layouts
The non-linear 3-step de-embedding procedure in details:
Figure: 3-step de-embedding procedure for large-signal non-linear RF measurements.
The large-signal currents and voltages in the de-embedding scenario shown above in fig.4,
are defined in the frequency domain, with the fundamental frequency and a certain
number of harmonics. The same frequency stimuli are applied to the S-parameter
measurements of the parasitic elements as well. In order to model the large-signal RF
behaviour at the level of the d ut, currents i 1dut and i 2dut , and voltages v 1dut and v 2dut
are calculated after:
where,
and the voltages v 1, v 2 and the currents i 1, i 2 are measured by the nonlinear network
analyzer.
320
Nonlinear RF Measurements and Device
Modeling
Topics
Introducing Nonlinear RF
Linear S-Parameter Measurements
Nonlinear VNA Measurements
Harmonic Balance (HB) Simulations
X-Parameters
PHD Model
Device Modeling Aspects for Nonlinear VNA On-Wafer Measurements
From DC > CV > NWA
This slide depicts the standard device modeling approached based on DC -> CV -> S-
Parameters:The DC modeling is non-linear, while the CV modeling and the S-Parameter
modeling are linear in the (many) operating points, i.e. tangents to the measured curves
at the operating points.In other words, the modeled 'tangents' to the bias points are a
substitute for large-signal non-linear RF excitation (trajectory curves).
Lightwave Analogy to RF Energy
321
One of the fundamental concepts of RF power transmission involves incident, reflected and
transmitted waves traveling along transmission lines. It is helpful to think of traveling
waves along a transmission line in terms of a light wave analogy. We can imagine incident
light striking some optical component like a clear lens. Some of the light is reflected off
the surface of the lens, but most of the light continues on through the lens. If the lens was
made of some glossy material, then some of the light could be absorbed within the lens. If
the lens had mirrored surfaces, then most of the light would be reflected and little or none
would be transmitted.
This concept is valid for RF signals as well, except the electromagnetic energy is in the RF
range instead of the optical range, and our components and circuits are electrical devices
and networks instead of lenses and mirrors.
Linear Vector Network Analyzer
In fact a vector network analyzer (VNA) is a complete measurement system > a signal
generator, providing sequentially the a1 (forward) and the a2 (reverse) waves:
a narrow-band S-Parameter testset, measuring (filtering out) the bi waves in forward

and reverse,
and a model extraction capability, calculating the S-Parameters, which can be
considered as a model of the DUT.
In other words, the VNA experiment consists of a forward and reverse excitation and
measurement.The model extraction refers to the equation-solving using the
measurements, resulting into S-Parameters.
Considering a 2-port: stimulated at port1, then at

port 2
322
Note
During S-Parameter measurements, the power sources provide the RF signal to the twoport. The
voltages v1, i1, v2 and i2 are therefore a consequence of that stimulating power source. On the
other hand, the above plot explains how they can be calculated out of the measured or simulated S-
Parameters (vi and ii as a function of the ai and bi ). In the formula above, voltages vx and the
currents ix are 'rms' values, i.e. peak/SQRT(2). The obtained waves ax and bx are 'rms' too. Their
unit is SQRT(Watt).
For the final set of S-Parameters, an overlay of forward and reverse measurement is applied. This is
possible for the 'linear' S-Parameters. But this is not possible for the nonlinear case, the nonlinear
network analyzer (NVNA).
The conditions a1=0 and a2=0 are fulfilled with an impedance of Z0 (typ. 50 Ohm) in
series with the stimulating voltage source at the input, and with an output
termination of again Z0.
S-Parameters = linear theory

323
S-Parameters are based on linear circuit theory. Therefore, the forward (a2=0) and
reverse (a1=0) measurement of a network analyzer, performed sequentially, can be
overlaid to give the 4 S-Parameters.
Also, in linear circuit theory, matrix conversion (from e.g. S-Par to Y-Par, to do PI-
schematic-based Y-Parameter modeling, is possible.
Last not least, these matrix conversions also allow the easy de-embedding methods
which subtract Y-matrices and Z-matrices.
As will be seen later, this all is not possible for nonlinear behavior, and therefore,
other methods need to be applied there.
S2Y, S2Z etc. matrix conversions are possible
S-Parameters are based on linear circuit theory. Therefore, the forward (a2=0) and
reverse (a1=0) measurement of a network analyzer, performed sequentially, can be
overlaid to give the 4 S-Parameters.
Also, in linear circuit theory, matrix conversion (from e.g. S-Par to Y-Par, to do PI-
schematic-based Y-Parameter modeling, is possible.
Last not least, these matrix conversions also allow the easy de-embedding methods
which subtract Y-matrices and Z-matrices.
As will be seen later, this all is not possible for nonlinear behavior, and therefore,
other methods need to be applied there.
Beyond S-Parameters
Beyond the S-Parameters, we have nonlinear device excitation: instead of a tangent

to the operating point, the large-signal excitation leads to a revolving 'satellite' curve
around the operating point: the trajectory curve.
This means there are 2 ways of looking at the signals:
Either in the time domain, or
Regarding the spectrum (DC, fundamental frequency and harmonics) in magnitude
and phase.
Extending linear to RF-power-dependent signals
324
Linear versus nonlinear behavior

Linear Behavior
Input and output frequencies are the same (no additional frequencies created)
Output frequency only undergoes magnitude and phase change
In order to understand nonlinear measurements, let's review the differences between

linear and nonlinear behavior. Devices that behave linearly only impose magnitude
and phase changes on input signals. Any sinusoid appearing at the input will also
appear as a sinusoid at the output at the same frequency, of course with different
magnitude and phase. No new signals are created.
Nonlinear behavior:
Additional frequencies created (harmonics, inter-modulation)
Non-linear devices convert single-frequency stimuli into output signals which exhibit
harmonics or intermodulation products. -> New signals are created.
Note
Many components that behave linearly under most signal conditions can exhibit nonlinear behavior if
driven with a large enough input signal.
325
Many network analyzers have the ability to perform power sweeps as well as
frequency sweeps. Power sweeps help characterize the nonlinear performance of an
amplifier. Shown above is a plot of an amplifier's output power versus input power at
a single frequency. Amplifier gain at any particular power level is the slope of this
curve. Notice that the amplifier has a linear region of operation where gain is
constant and is independent of power level. The gain in this region is commonly
referred to as small-signal gain. At some point as the input power is increased, the
amplifier gain appears to decrease, and the amplifier is said to be in compression.
Under this nonlinear condition, the amplifier output signal is no longer sinusoidal
some of the output power is present in harmonics, rather than occurring only at the
fundamental frequency.
Brief Explanation: Nonlinear device characterization and X-

Parameters
The NVNA features two methods for measuring the nonlinear effects of the device under
test:
326
Nonlinear device characterization (measuring the Ai and Bi waves)
X-Parameter measurement
Nonlinear device characterization measurements provide calibrated, vector corrected wave

forms of the device under test. Measurements such as amplitude and phase of harmonics,
group delay, phase between the harmonics and vector corrected time domain waveforms
can be made. Nonlinear component characterization measures the incident, transmitted
and reflected waves of the DUT. This data can then be displayed in frequency, time or
power domains as well as user defined ratios such as I/V to display dynamic loadlines.
Additionally, this information can be imported into Agilent's ADS simulator or into IC-CAP
for applying conventional models (BSIM, PSP, Hicum, VBIC, Angelov, Statz etc.) to
nonlinear RF measurement data. For an amplifier, the designers can measure the full
spectral output of the device in both amplitude and phase.
X-Parameters, on the other hand, are the logical, mathematically correct extension of S-
Parameters into a nonlinear operating environment. X-Parameters measurements require
an additional source which is used to drive the DUT with both a large signal tone and a
small signal tone at the same time. With the device information derived from such
measurements, the coefficients of the X-Parameter model are calculated. The X-
Parameters provide information such as the device gain, match, ... while the device is
operating in a linear or nonlinear state. The X-Parameters can then be used like S-
Parameters, but now for the non-linear, large-signal case. Since the X-Parameters have
power and harmonic dependencies, there are many more X-Parameters than S-
Parameters. Possibly one of the strongest benefits of X-Parameters is the ability to
accurately cascade the X-Parameters from individual devices using the PHD (Poly-
Harmonic-Distortion) model in ADS.
Detailed Explanation
X-Parameters are to nonlinear linear measurements what S-Parameters are to linear
measurements. S-Parameters were developed as a method to analyze and model the
linear behavior of RF components. They also play a key role in analyzing, modeling, and
designing more complex systems which cascade multiple individual components. They are
related to familiar measurements: S11 input match, S22 output match, S21 gain/loss,
S12 isolation, and can be easily imported into electronic simulation tools such as ADS.
While S-Parameters are extremely useful and powerful, they have limitations and are
defined only for small signal, linear systems.
The usefulness of X-Parameter information can be seen in the process of designing a

power amplifier. The designer is compelled to drive the amplifier into the nonlinear region
to get the maximum output power as well as to extract the maximum efficiency. Some
form of a feedback circuit is then used to compensate for the nonlinear effects and make
the output behave like a high power linear device. A typical approach is to suppress the
harmonic outputs of a power amplifier using filters and other components. If the input
match of the filtering components does not match the output match of the specific
harmonic of interest generated by the amplifier, then the attenuation level of that specific
harmonic could be significantly different than anticipated. This leads to a time consuming
experience of trail and error at best. With accurate phase and amplitude information from
the X-Parameters and using simulation tools such as ADS, designers can design the most
robust systems possible in the shortest amount of time, with the highest degree of
accuracy.
S-Parameters measure half of the parameters on the forward sweep and half on the
reverse sweep. Nonlinear characterization measurements gather all the reflected and
transmitted waves at once, and include the measurement of the un-ratioed, power
calibrated receiver measurements and the poly harmonic phases. For X-Parameter
measurements, both a large signal and small signal at the same frequency is presented to
the DUT at the same time. Knowing the phase of each signal and changing this
relationship enables the extraction of the measured coefficients.
327
The NVNA requires, additional to the conventional Vector Calibration known from linear
network analyzers, an Ampitude Calibration and a Phase Calibration, using a phase
standard.
Phase Reference
Phase: The critical element
Phase calibration is the important point of NVNA calibration. Due to the down-converision
of the signals into the IF domain, the phase information of the harmonics is lost. Phase
calibration preserves this information.
NVNA measurement
Similar to a linear VNA,the transmitted and reflected NVNA PowerWaves are defined as:
328
Note
In the formulae above, voltages vx and the currents ix are 'rms' values, i.e. peak/SQRT(2). The waves ax
and bx are 'rms' too. Their unit is SQRT(Watt).
Note
The above equations explain how the equivalent currents and voltages can be calculated out of the
measured a and b waves, in order to be compared in e.g. IC-CAP with ADS Harmonic Balance simulations
(currents and voltages) of any selected device model (PSP, HISUM, Gummel-Poon, VBIC, Hicum, Angelov,
Statz etc.).
The calculated peak voltages and currents at both ports are for DC, fundamental and
harmonics. This allows to compose the time-domain signal.
329
The slide above depicts the difference between a linear measurement (and its
corresponding S-Parameter simulation), and a non-linear measurement (and its
corresponding RF simulation).
It shows a diode DC characteristic (blue) and the trajectory curve (red) of a certain DC
bias point, vD=0,9V, stimulated with an RF signal of -40dBm (small signal) and of -10dBm
(large signal).
The Nonlinear Case (right):The curve represents an ellipse for small RF signal, but for
incrementing RF signal levels, this ellipse changes to a more complex trace, due to
evolving harmonics (which are a consequence of the nonlinear DC trace!).
330
I.e. instead of following the measured DC traces in a tangent at a certain operating point,
the trajectory curve of the large RF signal begins revolving along the DC characteristic due
to the harmonics. It widens up due to the diode's capacitances and the transit time. And
the DC operating point is typically also shifted, due to the energy provided by the large-
signal RF source!
The Linear Case (left):
Also for large-signal RF stimulation, a linear VNA filters-out (and an S-Parameter
simulation only considers) just the fundamental frequency (like it did it also for -40dBm),
and thus, the trajectory curve will still exhibit a tangent to the operating point also
for large RF signals. It will keep its shape, just become bigger. And it will not
represent the real-world complex, satelliting curve following the DC trace.
In other words:due to the linear approach (no harmonics, just the fundamental
frequency), the S-Parameter curves can only represent a trajectory curve in circle or
ellipse form (Lissajoux curves). It keeps its shape independent of the applied signal
power. It will only scale with it. This is equivalent to a linear VNA.
A NVNA measurement or an HB simulation, however, featuring fundamental and harmonic
frequencies, gives realistic trajectory curves, which can realisticly follow the nonlinear DC
curve.
What is Harmonic Balance Simulation?
Kirchoff's current law: sum of currents for all frequencies at each node is zero.
Such a system of nonlinear equations is solved using e.g. the Newton-Raphson
method
Note
Since most nonlinear devices are described by (Spice) models in the time domain, the simulator has
to transform (inverse FFT) the intermediate voltage spectrum into the time domain, evaluate the
response of the device and then transform it back to the frequency domain (FFT).
Harmonic Balance Simulation Flow Chart
Harmonic Balance Iterations
331
For the circuit given above, the following iterations are performed:
1. First, a DC simulation is executed to obtain the DC operating point of the circuit.

2. A linear AC analysis (Harmonic Balance Fourier Analysis) is applied to analyze the
linear and passive components in the frequency domain, stimulated by the
3. Nonlinear device models are usually defined as i=f(v). Therefore, the actual voltage
frequency spectrum across the nonlinear component is Fourier-transformed into the
time domain. This resulting time domain signal is applied to the nonlinear device with
a Spice-type simulation. The resulting time domain current is Fourier-converted back
to the frequency domain.
4. The sum of the actual current spectra at each circuit node is calculated and it is
checked, if this sum is below a certain tolerance limit for all frequencies.
5. Steps 1-4 are iterated until Kirchhoff's current law is satisfied for all frequencies at all
nodes. The same is true for fulfilling Kirchhoff's voltage law. Iterative, matrix-driven
techniques are used (Newton-Raphson algorithms etc.).
6. The obtained solution is a satisfaction of Kirchhoff's current law (sum of all currents,
for all frequencies, is zero for each node) as well as the voltage law. The
mathematical analysis has converged. The harmonics are now balanced.
For low RF signal levels, in the linear range of the amplifier characteristics, the harmonic
frequencies can be neglected.
332
For RF signal levels in the nonlinear range of the amplifier characteristics, the harmonic
frequencies play a role and can no longer be neglected.
Linear vs. Non-Linear: modeling-wise
Related to S-Parameters
S-Parameters are a vehicle to describe the performance of a device for high frequencies.
S-Parameters are not the 'real world!'
However,
The real world is: non-linear currents and voltages vs. time!
A quick intermezzo a look at Device Modeling.
Example: MOS-Transistor
what happens to the measurement (and modeling) traces when going from DC -> S-Par -
> nonlinear RF ?
333
This slide depicts the input, transfer and output characteristics of a MOS transistor. These
plots will be investigated further in the next slides to study the effects of network analyzer
measurements and NVNA measurements.
Beyond 100MHz, the phase shift between the stimulating voltage and the measured
current becomes bigger, and what was a tangent before during the CV measurements,
becomes now an ellipse. In the slide above, see especially how the input current of the
MOS transistor becomes now an ellipse around the stimulating vG operating point value.
334
Applying a large RF signal creates harmonic frequencies. Time-domain-wise, this is

represented by complex satelliting curves ("cucumber" or "potato" shapes !).
Applying the NVNA features realistic self-heating conditions for device modeling (for measurements and
simulations).
When performing NVNA measurements, the transistor (or an amplifier etc.) can be
operated in the conditions of its later application. The RF signal is satelliting around this
operating point. Therefore, the self-heating is automatically covered correctly when
measuring with the NVNA.
Besides measuring the Ai and Bi waves, a NVNA can also measure X-Parameters.
X-Parameters
Doing for nonlinear components under large-signal conditions what S-Parameters do for
linear systems.
S-Parameters
Calculated by the VNA out of the measured A1, A2, B1 and B2 power waves
(fundamental freq. only)
They represent a linear black-box-model of the DUT
X-Parameters
Calculated by the NVNA out of a stimulation at port1 and at port2
They represent a nonlinear black-box-model of the DUT (fundamental freq. +
harmonics + (optionally) DC bias)
X-Parameters are a rigorous, mathematically correct linearization of device under test
(DUT) behavior as represented by a spectral map from incident to scattered pseudo-
waves.
Unlike classic S-Parameters, which capture only linear device behavior and ignore
nonlinear behavior such as harmonic/intermod generation and even same-frequency
higher order mixing effects, X-Parameters capture linear device behavior and
linearize nonlinear behavior about a large signal operating point (LSOP).
Both, S-Parameters and X-Parameters can be interpreted as a 'model' of the DUT !
Note
Ai and Bi measurements, as discussed before, are executed extremely fast by the NVNA. Measuring
X-Parameters, applying the same stimuli takes much longer. Reason: a (small) stimulus signal is
applied also to port 2.
S-Parameters: Nonlinear System X-Parameters
Different to S-Parameters (simple matrix), X-Parameters are represented by a rather

non-linear matrix system, split into 3 terms: F-term, S-term and T-term. The F-term
can be down-stripped to linear S-Parameters, - the S-term is similar to Hot S22
measurements, a very special application of (linear) S-Parameters to non-linear RF
power devices - and the T-term takes into account all the non-linearities.
The terms X (F), X (S) and X (T) depend on the fundamental frequency and its
harmonics and the power transfer between port m and port n, and (optionally)
335
depend also on the DC bias.
What are HOT S-Parameters?

The output signal of nonlinear devices is input power level dependent. Most
amplifiers, however, are operated close to the 1dB compression point. HOT S-
Parameter measurements emulate this scenario, using a conventional linear VNA.
For HOT S-Parameter measurements, the input signal at port1 is at frequency f1 ,
and its power level is so that the amplifier's output signal is close to the 1dB
compression point. I.e. the amplifier is operated in its later application condition.
At the same time, a second small test signal is applied at port2 with f2 = f1+Df, and
S22 is measured with this 2nd stimulus. NOTE: The power level of this is typically -
30dB smaller than the 1dB compression point output power. NOTE: Df should be
bigger than the IF bandwidth of the VNA.
References
The following references are available for additional information on X-Parameters and
the PHD framework:
1. D. E. Root et al., "Broad-Band Poly-Harmonic Distortion (PHD) Behavioral Models

From Fast Automated Simulations and Large-Signal Vectorial Network
Measurements," IEEE Trans. MTT, vol. 53, no. 11, pp. 3656-3664, November 2005
2. J. Verspecht and D. E. Root, "Poly-Harmonic Distortion Modeling," in IEEE Microwave
Theory and Techniques Microwave Magazine, June, 2006.
3. J. Verspecht, D. Gunyan, J. Horn, J. Xu, A. Cognata, and D.E. Root, "Multi-tone,
Multi-Port, and Dynamic Memory Enhancements to PHD Nonlinear Behavioral Models
from Large-Signal Measurements and Simulations," 2007 IEEE MTT-S Int. Microwave
Symp. Dig., Honolulu, HI, USA, June 2007
4. J. Horn et al, "X-Parameter Measurement and Simulation of a GSM Handset
Amplifier", accepted for publication at European Microwave Conference, 2008
336
Ideal X-Parameter Extraction Steps (Perfect
match, no source harmonics)
The slide above describes the measurements necessary to extract X-parameters.

At each operating point (large signal level):
One measurement is required with no extraction tone to get the XF terms

Two measurements with extraction tones are required at each port/frequency of
interest to get the XS and XT terms. The extraction tones in the two measurements
have different phases (ideally 90 degrees offset).
337
X-Parameter Measurement Principle
Wrap-Up:
1st measurement (red underline in the formula, red vector in sketch 'output bij'): ->
XF
2nd and 3rd measurement (blue and green underline in the formula, blue and green
vector in sketch 'output bij'): -> XS and XT
X-Parameters
The NVNA measures X-Parameters about a single-tone, large signal operating point.
This is achieved by stimulating the DUT with a single large tone at port 1 and, while
continuing to apply the large tone, by injecting additional small tones at both ports 1 and
2 at all harmonics of interest. The level of the extraction tone must be small enough to
ensure a spectrally linear response, but large enough to ensure that the response is
measurable.The extraction tones will be injected into the DUT for the extraction of XS and
XT.
At least two phase-offset small tones must be injected at each port/frequency of interest
in order to extract the corresponding X-Parameters. Although two phases are sufficient to
allow X-Parameter extraction, both the accuracy and robustness of the extraction are
improved by using additional phases. A setting of five phases is recommended to provide
robust X-Parameter extraction in most measurement configurations.
All resulting waves are measured for each stimulus, and the X-Parameters are calculated
directly (if only two phases are used) or using regression (in the case of 3 or more
phases).
Intermezzo: NVNA X-Parameter Display
338
Port,Harmonic indexing in general:
p=i,k
q=j,l
See the indexing in the previous slides.
Intermezzo: Number of X-Parameter

measurements
In addition to XS and XT, there is also an XF term.XF is the X-Parameter component
of the output due to the large signal input (a11). It is indexed only to the receive
(port, harmonic) XFp. For each possible selection of p, there is exactly one XF term.
Therefore, in total, there are 13 X-Parameters for each possible (port, harmonic)
selection of p. For example, for Port2 and the fundamental frequency, i.e. p=(2,1),
and assuming 3 harmonics of interest:
(XF21)
(XS21,11), (XT21,11)
(XS21,12), (XT21,12)
(XS21,13), (XT21,13)
(XS21,21), (XT21,21)
(XS21,22), (XT21,22)
(XS21,23), (XT21,23)
For a 2-port device, with 3 harmonics of interest at each port, there are 6 possible
(port, harm) combinations for p. Each (port, harm) combination yields 13 parameters
as shown above, and therefore, we end up with a total of 6*13 = 78 X-Parameters
to be measured.
Example of an X-Parameter MDIF file, created by

the NVNA
339
The size of such an MDIF file is typically several MByte.

NOTE about file types which can be saved by the NVNA after an X-Parameter
measurement:
The X-Parm measurement file is the raw (still vector corrected) measurements before
applied to the X-Parm extraction algorithms.
The X-Parms file is the extracted X-Parms but without normalization.
The PHD model file is the normalized X-Parameters with normalization.
Note
The PHD model file is the one you should use in ADS.
Simulating with X-Parameters

Although subsets of X-Parameters can be useful design tools on their own, their true
predictive power comes from using the entire set together. Since the X-Parameters
include both magnitude and phase information, the computed b terms can be
combined, along with the XF terms to accurately predict device response to a wide
range of input signals.
This includes properly accounting for upstream harmonics and downstream match.
The complete equation for scattered wave prediction is as follows:
This powerful predictive ability has been combined with:

Interpolation on frequency and input power as well as any additional independent
variables such as DC bias
Optionally load-pulling,
Volterra constraints (to ensure physically correct extrapolation below measured
power levels in the PHD framework of ADS).
ADS:X-Parameters become the PHD Model

What is the PHD Model ?
PHD Model (Poly Harmonic Distortion Model) means: using the X-Parameters as a 'non-
linear blackbox model' in ADS, and applying special interpolation algorithms
The NVNA can calculate and directly export the nonlin RF measurement into an ADS
PHD.mdf data file (save the data as 'mdif file for PHD).Volterra constraints for small signal
levels are enforced, ensuring that the harmonic levels go to zero at the appropriate rate.
340
Also the data are re-ordered to be monotonic and fully compatible with ADS.
ADS can directly read this file.
Note
ADS can also handle cascaded PHD blocks.
PHD Model
The size of such an MDIF file is typically several MByte.

The file will include- the applied RF signal level (in SQRT(Watt)) in the data column
AN_1_1.
The conversion to dBm is after
Power_dBm = 30 + 10*log10((AN_1_1)^2)
FB-x-x for the X-Parameter's F-term, T-x-x-x-x for the T-term and S-x-x-x-x for the
S-term.
There may also be DC values present in this file, supposed the user has declared a
sequence in the NVNA for setting different VDC levels by GPIB commands for a DC
power supply.
You find these DC bias informations as variables VDC_p where p is the port number
on which the DC bias appears. If the bias variables are named VDC_1 and VDC_2,
the bias information is covered by shared DC/RF ports 1 and 2. In this case, you will
need DC_Feed and RF_Feed components in ADS around your PHD component, to
separate DC from AC. If VDC_3 and VDC_4 are used, DC-only ports 3 and 4 will be
created on the component. You should consider this when naming the DC bias
voltages in the DC instrument driver section of the NVNA. The DC bias currents can
be found in the data columns FI_p.
341
Note
To properly incorporate DC bias behavior into the PHD file, the following naming conventions must
be followed (p is the port number):
Swept or Measured Variable Naming Convention
Swept DC Voltage at port p VDC_p
Swept DC Current at port p IDC_p
Measured DC voltage at port p Vp
Measured DC current at port p Ip
ADS does not allow device names to start with a number. Therefore, give your .mdif file a name
starting with a character, and not with a number.
How to load a PHD file into ADS:
1. In the NVNA, save the data as 'mdif file for PHD'.

2. In your ADS project directory, create a new directory called 'PHD_Files'.
3. In your ADS session, select (as shown above) the menu pick 'PHD (X-
Params)/Create_single_PHD_Component'. A file browser opens up, and you should
be able to select your X-Parameter .mdif file(s), located in the project sub-directory
'PHD_Files'.
4. To see what has happened: In ADS/Main, open a new 'Schematic', and select 'Open
Design'. Select the previously created PHD Project. Then close this Design.
5. Open a new Design (or use one of your existing ones), and select the PHD-subcircuit
(Design) from 'Component Library/Designs'. Drag your PHD Design into your
Schematic.
342
Besides creating PHD models from NVNA measurements, PHD models can also be
created out of (parts of) a schematic in ADS.
Using PHD .mdif files, relevant information can be exchanged without disclosing
technology information.
Wrap-Up NVNA Measurements

S-Parameters: a linear matrix, calculated by the VNA out of the measured linear Ai
and Bi waves.
NVNA Harmonics Measurement: The nonlinear Ai and Bi power waves with a RF
signal applied to port-1 (corresponding to a HB simul.). Measures Magnitude and
Phase of a speficied nr. of harmonics.
X-Parameters: A highly nonlinear matrix system, calculated by the NVNA out of the
nonlinear Ai and Bi waves when applying signals to both ports. Is called PHD Model
when used in ADS.
Agenda
Introducing Nonlinear RF

X-Parameters
PHD Model
Performing the NVNA on-wafer Measurements
For our application, the NVNA measures the transmitted (b) and reflected (a) PowerWaves
and calculates the ix and vx.
Note
The NVNA also triggers the DC Analyzer which provides the DC bias for the nonlinear RF measurement.
NVNA Calibration
343
The NVNA requires, additional to the conventional Vector Calibration known from linear
network analyzers, an Ampitude Calibration and a Phase Calibration, using a phase
standard.
NVNA Calibration Plane
On the NVNA, probe de-embedding is supported either during measurement or during

calibration through the NVNA software. The deembedding settings are accessed under the
Response->Cal->Fixturing menu.
A connectorized (3.5mm) cal, plus an OPEN, SHORT and LOAD measurement of the
ISS standards.
Note
Measurement of the GSG probe S-parameters can be done with the Adapter Characterization macro in the
PNA-X VNA firmware or using 3rd party software such as WinCal.
Importing the NVNA (DC, i1, v1, i2, v2) mdif data file
into IC-CAP
Pls. contact the author for the IC-CAP file 'read_NVNA_mdif.mdl'.
344
As discussed so far, the NVNA features an integrated down-to-the-probe-tips

calibration:implemented as:
a connectorized (3.5mm) cal,

plus an OPEN, SHORT, LOAD measurement of the ISS standards and their de-
embedding.
For on-wafer diodes and transistors with contact pads, we need to apply Nonlinear De-
Embedding
The measured nonlinear currents and voltages cannot be converted to e.g. S-

Parameters or Y-Parameters
De-Embedding by matrix subtraction is not possible !!!
We have to deal with the voltages and currents (Kirchhoff's Law)
The de-embedded large signal voltages and currents (v1dut, i1dut, v2dut, i2dut) are
obtained from the measured ones (v1, i1, v2, i2) after:
NonlinearDe-Embedding
In the formulae,Yij are the S-to-Y-converted S-Parameters of the OPEN Dummy Zij the S-
to-Z-converted S-Parameters of the SHORT Dummy (de-embedded before from the OPEN
Dummy)
Pulication: G.Crupi, D.Schreurs, D.Xiao, A.Caddemi, B.Parvais, A.Mercha,

S.Decoutere: "Determination and Validation of New Nonlinear FinFET Model
Based on Lookup Tables", IEEE MicrowaveAnd Wireless Components Letters,
Vol. 17, No. 5, May 2007
IC-CAP Demo File: nonlinear_deemb_Tutorial.mdl
Case Study Example

Non-linear de-embedding of a transistor from a big Capacitor
345
In this case study, a transistor plus an extra package capacitor (with a big capacitance
value, to make the effect more obvious) is considered.This extra package capacitor shall
be de-embedded.
When we simulate the transistor plus the (pretty big) package capacitor, and sweep the
RF input signal from -60dBm to +10dBm, and display the fundamental and the harmonics
at the pins of the transistor, the fundamental follows the input signal, but at a much
lower level. This is because most of the stimulating input signal is lost to ground via the
(big) capacitor. As a consequence, the harmonics at the very transistor pins are >100dB
below the fundamental:
The transistor behaves linear, although the RF input signal is very big.
When we simulate the transistor without the capacitor, all of the RF input signal
goes across the transistor. The fundamental at the output again follows the input
signal, with 20dB above the input signal level, and is clipped at an input signal level
of ~ -20dBm. And the harmonics are not negligible at all.
Nonlinear de-embedding affects the operating condition of the DUT: in our example,
the transistor changes from linear operation (non-de-embedded) to non-linear (de-
embedded).
Linear deembedding, i.e. subtracting Y- and Z-matrices, keeps the DUT in its
operating mode: also after de-embedding, it is linear.
Note
Therefore, when measuring S-Parameters of devices like transistors, we take so much attention to not
overdrive the transistor, and to always keep the transistor in linear operation.
Notes about De-Embedding

S-Parameter: device was linear before, is linear after de-embedding
NVNA: device's linearity/non-linearity is affected by the de-embedding
We have now
Shifted the cal. plane to the very DUT,and are ready to

Compare these measurements against simulations, to finally do
Device modeling
In order to compare the NVNA measurement with an HB simulationin the time domain, the
phases of measured and simulated signal need to be normalized.
Intermezzo:Phase normalization is achieved by dividing a signal by its phasor
346
Usually, v 1 is used as reference signal. Therefore, all measured v x and i x are divided by
the phasor of v 1meas and all simulated v x and ix are divided by the phasor of v 1simul.
This means: v 1meas has 0 phase, and v 1simul has 0 phase as well.
Calculate the Fourier components of the vx and ix
Creating the Dynamic Transfer Trajectory
347
Now that we have the time-dependent voltages and currents of the fundamental and its
harmonics, we can even plot current vs. voltage (trajectories).
Handling NVNA measurements and standard device

modelslike PSP or Hicum, Gummel-Poon, Angelov etc.
in IC-CAP
Once the NVNA data have been imported into an IC-CAP session, a conventional device
modeling can be applied, using a BSIM, PSP, Hicum, Angelov, Curtice model etc. Physics-
based device models are a useful application of NVNA measurements, when considering->
scaling rules for compact models that allow you to measure a device of many sizes. ->
process variation is another important application for compact models.
A possible Modeling Sequence:

DC
S-Parameters
Convert to Y-Parameter for PI-schematic modeling
Extract, verify and fine-tune model parameters using the NVNA data in time and
power domain
To ease your learning of nonlinear RF measurements, simulations and modeling, an IC-

CAP tutorial ModelFile has been developed.
Note
Pls. contact the author for a copy of the IC-CAP nonlinear RF demo file.
The IC-CAP Nonlinear RF Tutorial ModelFile (1)
348
349
Conclusions
Introducing Non-Linear RF

X-Parameters
PHD Model
Modeling Tips and Tricks

Effect of L's and C's on the Dynamic Transfer Curve
For low RF signal, the dyn. transfer curve follows the DC trace perfectly, if all C's = 0F.
When adding the C's with the ADS simulation, the tangent opens up to a 'cucumber'-
shaped trace following up and down the DC trace, and widening up with increasing C
values. For a trajectory curve perpendicular to the DC trace at low RF signal, we need to
apply inductors (package) in addition to the capacitors.
vDS vGS in the time domain

LEFT: The time-domain-converted ADS harmonic balance simulations of a simple
transistor model: just a voltage-controlled current source between Drain and Source (gm
= 50mS), no capacitors, no transit time.
The data refer to Pin=-30dBm. As it can be expected, vDS is counter-phase to vGS.
RIGHT: Again vGS and vDS, but now including the model capacitors. And again the
lowest power applied in the measurements, i.e. no clipping, just the fundamental
frequency.
Since the capacitances are not 0, the vDS is behind or comes later than the ideal vDS in
the simulations. (Although may look like vDS comes before vGS).
350
351
Time Domain
Contents
TDR Measurements Basics (iccapmhb)

TDR Calibration Techniques (iccapmhb)
TDR Plots Tutorials (iccapmhb)
Differential TDR Measurements (iccapmhb)
352
TDR Measurements Basics
This chapter is intended to make you familiar with TDR measurements and the
interpretation of TDR plots. Let's commence with some basic TDR theory: Assuming we
have the following measurement, an unknown resistor Zx to ground between two 50 Ohm
delay lines.
At any point within the tested device, there is a backreflected and an on-going voltage:
with Z : impedance seen in forward direction at the actual location.

Example:
Be Vstep = 1 (normalized), then Vin = 0.5 when the step hits the location X at t=0.
At location Y , we have after time t1 and with Rx = 50 Ohm using equ.(1):
This voltage Vrefl is reflected back to the TDR port and seen there after time 2 * t1
overlaying the on-going Vin=0.5 as
The ongoing propagating wave beyond location Y is then,
what is observed with the scope input at the open end at the time t1 + t2 as
Where factor 2 comes from equation (2) for Z = infinite (open end).
As another example, we can calculate the
impedance Z at any location of the backreflected TDR graph:
From (1), we obtain
or solved for Z:
--------------------------------------------(5)
It is,
and solved for,
-------------------------------------------(6)
Introducing (6) into (5) gives finally the local impedance along the TDR measurement:
with Vstep typically 400mV and 0 < VTDR < Vstep
353
Modeling the Complex Load Out of TDR Plots
Equations for ideal TDR response to complex loads can be derived, allowing to determine
R, L, and C values of a circuit. Response time constant and final value, resp. incremental
step, are usually the most important factors for such calculations. Some typical TDR
responses can be seen in figure 1 below. Its sketches are from the HP application note 62-
3 'Advanced TDR techniques'.
For example, with a series RL circuit, with a measured time constant of 20ps, and a final
value of 150mV from a 200mV input Ei: R=30 Ohm, and L=1.6nH.
354
Figure: TDR responses to complex loads
About the Context Between Time and Location

The signal delay TD is given by the length L of the line, and the dielectric constant of the
dielectric materiel. The delay is not dependent on the diameter of the conductor.
For a conductor, surrounded homogeneously by a dielectric (r), it is
with
0: 8.854*10-12 F/m
r: dielectric constant of the isolator
0: 1.257*10-6 H/m
L: conductor length
or
As a rule of thumb, the speed of a step function on a ceramic substrate strip line (r~3) is
roughly:
Note
On a TDR, the displayed time represents the distance to the location and back.
This gives also an idea about when the connecting components between electronic devices
have to be modeled too. This is necessary when the forth and back traveling impulse time
is longer than the impulse rise time. This means that the still rising edge of the impulse
will be interfered by the back reflection (glitches in digital signals).
355
TDR Calibration Techniques
TDR (time domain reflectometer) measurements meet ideally the needs of modeling
components with geometric extension. This type of devices include packages, connectors,
strip lines on PC boards (printed circuit) etc. The advantage of TDR measurements is that
the individual components of the DUT appear at separate time slots in a direct relation to
their location. The NWA dilemma, namely that all sub-components of the DUT add
basically to the 'turning to the right' of the Smith chart, and thus overlay each other's
effect, is easily avoided by this method.
Basically, a TDR is a very fast GHz oscilloscope including a fast pulse generator in the
picoseconds range. This fast pulse with an amplitude of usually 400mV is applied to the
DUT, and the backreflections of that voltage are displayed on the oscilloscope. Therefore,
the time axis on that plot corresponds to the physical location of the event on/in the DUT.
For example, when an OPEN ended line is connected to the TDR output, the resulting
oscilloscope reading shows a 200mV amplitude (voltage divider: 50? internal resistance
and 50 characteristic impedance of the line), as long as the impulse is present in the
line. Once the open end is reached, there is a jump to 400mV. Otherwise, if the line is
SHORTed at its end, the impulse again sees a 50 characteristic impedance (while still in
the line), yet disintegrates to 0V when hitting the SHORT at the end of the line.
If there is a capacitor between two 50 lines, the reflected signal with its level of 200mV
(in line 1 in fig.1) will briefly drop down to 0V, 'seeing' the capacitor as a SHORT for the
first moment, and then ascend back to 200mV (line 2) with an exponential rise that is
proportional to the capacitance value. In case there is an inductor instead of the capacitor,
the reflected signal will jump to a full 400mV (because the inductor behaves like an OPEN
in the first moment) to once again come back to the 200mV level of line 2.
Figure: time domain reflectometry (TDR)
A connector between two lines is commonly seen by a TDR like a short overshoot followed
by a short undershoot. We now know what this would mean for modeling: the
overshooting represents an inductor L, while the undershooting hints to a capacitor C. The
following figure shows such a TDR measurement.
Figure: TDR measurement of a connector.
Note
The basic principles of understanding TDR measurements are summarized in the chapter entitled
'Interpreting TDR Measurements'.
Calibration Techniques
When measuring and modeling a packaged device in the GHz region, it must be placed in
a test fixture. Therefore, when using a TDR, we have to do a calibration of the reflected
signal first. This is done by the following two steps:
356
1. Model the step impulse of the TDR
(either using the SPICE step function with the parameters slope, start and end value,
or more accurately, using a sum of PWL function (piecewise linear) of the simulator)
2. Determine the length of the connection cable
(modeled using a delay line)
After that, we connect the cable to the test fixture, model it and, finally, model the
component itself. The following figure illustrates these steps.
Because of the sequence of modeling steps, it is always very easy to determine the part of
the equivalent schematic related to the test fixture or the DUT itself.
Figure: The individual steps to calibrate a TDR and to perform modeling measurements
TDR Calibration under IC-CAP

Examples about how to calibrate the TDR step function and the test cable in IC-CAP, can
be found in the following IC-CAP model file
$ICCAP_ROOT/examples/demo_features/4extraction/specific_PEL_routines/tdr_cal.mdl
Publications
Advanced TDR techniques, Hewlett-Packard Application Note 62-3, Publication HP5952-
1141, April 1990.
Evaluating Microstrip with Time Domain Reflectometry, Hewlett-Packard Application Note

1304-1, Publication HP5968-0007E, August 1998
Time Domain Reflectometry Theory, Hewlett-Packard Application Note 1304-2, Publication

HP5966-4855E, May 1998
357
TDR Plots Tutorials
This chapter is intended to give you some ideas on how electronic components look like in
the time domain.
Introduction
The characterization of parasitic devices can be done in the time domain or in the
frequency domain. But time domain may be smarter, since the parasitic devices show-up
with respect to their physical location. In the frequency domain, using s-parameters, we
always have an overlay of curves that 'turn to the right' for capacitors, inductors or delay
lines.
Note
For frequencies below 3GHz, a simple parasitic device schematic is often sufficient (two delay lines or
capacitors at port1 and port2 and a cross-coupling capacitor, no inductors).
Using time domain, it is best using a time domain reflectometer (TDR) like the hp54xxx
series oscilloscopes with a TDR plug-in on channel_1. A voltage step of typically 200mV is
triggered by IC-CAP on channel_1 and the back reflections are then measured by
channel_1 also. Using other channels of the TDR, one could also measure the cross-
coupling from one port of the test object to another.
TDR measurements are best for the modeling of geometrically large or spread structures.
As an example, we consider the modeling of the test fixture:
This can be done using delay lines, inductors (positive spikes in the back-reflected signal)
and/or capacitors (negative spikes). Changes in the amplitude of the reflected signal are
modeled using parallel (negative amplitude step) or series resistors (positive amplitude
step). After the the measured curve has been fitted, the parameters of the test fixture
parasitics are known and the 'inner' transistor can be de-embedded from S-parameter
NWA measurements using Y-, Z- and/or S-matrix manipulations.
Expected Reflectograms of Typical Parasitic

Components
IC_CAP file: tdr_plots.mdl
The following plots are intended to help you with the development of equivalent circuits
from reflectograms.
TDR Basics
358
359
360
361
362
TDR Plots for Experts
363
See also the chapter 'Skin Effect' in the Modeling Handbook.
Note
Parallel strip lines, coupling across with mutual inductors (even with high coupling factors), are usually not
much affecting the TDR measurement of the single strip line. But they can affect the corresponding S-
parameter measurements heavily! So, for their modeling, either a THRU measurement of the cross-
coupled signal using the oscilloscope input of the TDR will be necessary or additional measurements using
a network analyzer.
364
Differential TDR Measurements
Since IC-CAP 5.4, also differential TDR measurement capabilities are available. This
follows the growing demand for the modeling of the behavior of connectors, packages etc.
when used with differential signals. Therefore, two new entries have been added to the
Agilent 54750 Instrument table:
1. Differential Mode: Set the instrument in differential mode. Channel 1 and 2 are the
TDR channels. The differential stimulus on channel 1 and 2 can be Differential (DIFF)
or Common (COMM). Default is no differential stimulus (NONE).
2. Differential Response Mode: Once the instrument has been calibrated in
differential mode, the response reading can be set Differential (DIFF) or Common
(COMM). Note that this field is active only when the Normalize Flag of the response
channels is set to yes. Default is DIFF.
To make TDR differential measurements place the Agilent 54754A plug-in in the first two
instrument slots (channel 1 and 2). In the IC-CAP measurement page insert one input of
type TDR (Unit TDR1 or CH1). Insert one input of Mode T (Type LIN) and set the time
interval and the number of points. Insert two outputs of Mode V monitoring channel 1 and
2. In the 54750 Instrument Option Table, set the Differential Mode to DIFF or COMM.
To measure raw data simply set the Normalize flags of CH1 and CH2 to N and run the
measurements. To measure normalized data, one needs to perform the TDR normalization
before running the measurements. Follow the instructions in the 54754 manual on how to
calibrate in differential TDR mode. Once the instrument has been successfully calibrated,
set the Normalize Mode to TDR, set Differential Response Mode to DIFF or COMM. To
measure the normalized response simply set the Normalize flag of channel 1 and 2 to yes.
Summary Differential TDR

- Differential Mode Differential Response Mode Response Mode CH1 Norm CH2 Norm
Raw DIFF/COMM Not Relevant Not Relevant N N
Norm DIFF/COMM DIFF/COMM TDR Y Y
To make TDT differential measurements place one Agilent 54754A plug-in in the first two
instrument slots (channel 1 and 2) and second 54754 plug-in in the third and fourth slots.
When measuring differential TDT, the driver assumes that Channel 1 and 2 supply the
differential stimulus (input). In the IC-CAP measurement setup page insert one input of
type TDR (Unit TDR1 or CH1). Insert one input of Mode T (Type LIN) and set the time
interval and the number of points. Insert four outputs of Mode V monitoring channel 1 to
4. In the 54750 Instrument Option Table, set the Differential Mode to DIFF or COMM.
To measure raw data simply set the Normalize flags of CH1,CH2,CH3 and CH4 to N and
run the measurements. To measure normalized data, the user needs to perform the TDT
normalization before running the measurements. Follow the instructions in the 54754
manual on how to calibrate in differential TDT mode. Once the instrument has been
successfully calibrated, set the Normalize Mode to TDT, set Differential Response Mode to
DIFF or COMM. To measure the normalized response simply set the normalized flag of
channel 3 and 4 to yes.
Summary Differential TDT

- Differential Differential Response Response CH1 CH2 CH3 CH4
Mode Mode Mode Norm Norm Norm Norm
Raw DIFF/COMM Not Relevant Not Relevant N N N N
Norm DIFF/COMM DIFF/COMM TDT N N Y Y
365
Noise
Contents
Noise Measurement Golden Rules (iccapmhb)
366
Noise Measurement Golden Rules
Verifying the Measurement Data of the IC-CAP 1/f
Noise Toolkit
The Applied 1/f Measurement Setup
Note
Website of Stanford Research Systems:http://www.thinksrs.com
1Hz Lowpass Filter

The filter is required for biasing the transistor input. It offers -60dB attenuation at 50Hz.
Required Output Impedance Rout (to avoid shortening the 1/f transistor noise):
Rout =330 k for bipolar (in any case Rout > Rbe_noise_bias),
50 for MOS transistors
In a bipolar transistor, the 1/f noise is generated in the B-E region. If the output
resistance Rout of the 1Hz filter was smaller than the dynamic, iB-dependent value of rBE,
it would AC-wise shorten the rBE (by the big capacitort in the 1Hz filter). Therefore, select
Rout >>rBE. For MOS, Rout=50 Ohm is sufficient.
Importance of the Filter Rout for Bipolar Transistors
Since the filter capacitors represent a SHORT for the 1/f noisecreated mainly by rbe , Rout
of the bias filter must be bigger than any rbe(vbe) during the 1/f noise measurements.
Otherwise, the measured 1/f noise will be smaller than the real 1/f noise.
.subckt bip_noise 1=C 2=B 5=L 99=AUX

*satisfy SPICE for a noise simulation
Raux 99 0 1
*emulate the bias lowpass filter at the Base
Rfilter1 2 21 1k
Cfilter1 21 0 100u
Rfilter2 21 22 1k
Cfilter2 22 0 100u
*here is the filter output resistor (high value for bipolar transistors)
367
Rout 22 23 330k
*here comes the call to the transistor in the Model file's Circuit tab
X1 11 23 0 noise_1_f_bip
*add a dummy 0V source to sense the transistor noise current
V1 1 11 0
*convert the Collector noise current into an identical voltage for
* further use in the modeling toolkit
H1 5 0 V1 1
R1 5 0 1
.ends
Hint
You can study the effect of Rout to a bipolar 1/f measurement also when varying the Rout in the
simulation deck.
Checking the resolution of the Dyn. Sign. Analyzer 35670A
In other words, the device noise has to be amplified (by the SR570)above the resolution
level of the 35670A.
The Low Noise Amplifier SR570

some hints (from the SR570 Manual)
1. The Instrument must be warmed up for about 60 minutes.

2. For best performance, the input current should produce an output voltage of about 1
V or less. This eliminates problems with slew rate limiting in the various amplifier
stages.
3. Make sure the source impedance is greater than the inverse of the sensitivity (e.g.
with a sensitivity of 1 nA/V use a source impedance greater than 1 GOhm).
4. Use short lengths of high quality coaxial cable to connect to the amplifier input.
5. Keep the amplifier output below 1 VRMS to avoid slew rate limiting at high
frequencies.
6. If you want to ground the chassis, use the green connector on the back side, but do
not connect the chassis to the amplifier ground (white connector).
7. For noise measurements, disconnect the power cord and use the internal batteries.
Note
If you experience problems with the 1/f toolkit measurement setup, apply an oscilloscope.
Simply connect the output of the SR570 amplifier to the oscilloscope.
Typical problems with 1/f measurement setup
Do not use 1Hz filters with unipolar electrolyte-capacitors (instead of cascaded antiparallel
electrolyte-capacitors). This may lead to very bad popcorn noise (!!) with frequencies at ~1/10Hz
for negative biasing (PMOS), what lead to non-reproducible measurement results.
Do not connected the voltmeter common (to check the SR570 voltage offset settings) to amplifier
common and not to chassis ground of the SR570 either. (These connectors are at the rear side of
the SR570). Instead, fully disconnect the voltmeter during the noise measurements.
The Cascade wafer prober monitor may affect the measurement at several 10kHz: switch it off
during 1/f measurements
Like the 1Hz filter, the SR570 exhibits also a settling time of ~10sec, since it is operated in
bandpath mode .
The SR570 sensitivity (1/gain) should be >= Gout of the transistor. However, a too high sensitivity
increases the SR570 noise contribution. For details, see the SR570 manual for the resolution plot
"Current Noise as a function of Frequency for several sensitivity settings (typical).
The offset current settings of the SR570 can add SR570-1/f-noise .
In the same way, the voltage offset of the SR570 can add SR570-1/f-noise up to 1E-17 A/Hz. It is
partly cancelled out with an appropriate offset current setting. Note: This was verified with resistor
measurements, but may be different with transistor measurements.
Recommendation: apply the SR570 current offset carefully, and if you are not sure, measure again
with a different current offset.
Verifying the SR570 Measurements
368
From the SR570 Manual:

Current Noise as a function of Frequency for several sensitivity settings (typical).
In order to verify the SR570 specs, a sequence of measurements
at different Sensitivities (SR570 amplifications) was performed using IC-CAP and the
Agilent 35670A. Both, voltage and current bias were OFF at the SR570 and the source
resistor was set to R=1/Sensitivity.
Measured resolutions for a sequence of SR570 Sensitivity Settings.

Sensitivity = 1/R metal_film_resistor
In every case, the calculated metal film resistor's white noise level(straight blue line) was
well below the internal SR570 noise level. Therefore, the measured noise levels (red)
represent the resolution limit of the SR570, due to the selected SR570-sensitivity.
at a glance: SR570 resolution levels as a function of its Sensitivity (Gain)
It was observed, that the SR570 bias voltage adds 1/f-like noise to the measurement.
But, that's an effect we have to live with. At Sensitivity=10u, a 100 resistor was
measured with and without 2V Voffset
Ioffset = 0
This effect was identical for a metal film and a wired resistor of the same value. Therefore, the effect
stems from the SR570 and not the source resistor.
369
Adding a current bias to compensate the resistorcurrent reduces

the 1/f-like noise addition of the offset voltage,but does not fully compensate it.
At Sensitivity=10u, a 100 resistor was measured - without any offset- with 2V 0A Offset-
with 2V 5uA Offset.
Note
this effect was identical for a metal film and a wired resistor of the same value
Measurement test sequence for fixed gain (Sensitivity) and varying current offset.
Measurement test sequence (PMOS Transistor)for fixed gain (sensitivity) but varying
offset current [required offset current: 360uA].
370
Observation: offset currents below desiredvalue do not add noise to the measurement
system.
Measurement test sequence for fixed current offset and varying gain (Sensitivity).
Observation: the SR570 manual states but when the Sensitivity is

increased, the noise level of the SR560 increases too.
Measurement test sequence (PMOS)for offset OFF but varying gain (sensitivity)
[required sensitivity: 50uS = 1/Rout_PMOS]
Observation: sensitivity (gain) above desiredvalue shifts the noise level of the
measurement system
Conclusions for obtaining best results with the SR570 amplifier:
on the SR570,
1. Set bias offset voltage

2. Set current compensation
3. Set Sensitivity = 1/R4. make sure the SR570 output is above the 35670 noise floor
otherwise increase the SR570 amplification (smaller value of the SR570 sensitivity
setting)
4. If overflow lights up, adjust the current compensation
Note
If you are not sure whether you measure the DUT's noiseor the SR570's noise, vary the sensitivity, the
voltage and current bias and check the effect to your measurement.
Best Practice Notes, March 2006: Falk Korndrfer
Always apply lowest sensitivity (highest SR570 amplification) for best resolution (1E-23
A/Hz). Adjust bias compensation to avoid overload alarm.
Last not least, do not forget to check these SR570 instrument settings:
371
A Note on Data Consistency Checks for MOS

Transistors
For MOS-Transistors, vg_noise is independent of the applied bias. Therefore, calculating
vg_noise=id_noise/gm^2 should give overlying curves .
See: M.T.Yang et.al., Characterization and Model of On-Chip Flicker Noise

With Deep N-Well Isolation for 130nm and Beyond,ICMTS 2005, Leuven,
Conference Proceedings.
Acknowledgements:
special thanks to Edel Griffith of Analog Devices in Limerickand to Falk
Korndrfer at IHP-Microelectronics/Frankfurt-Oderfor many interesting
discussions !
Appendix
Parts List
Small Parts w/o Supplier
1ea 1Hz Filter with switch bipolar-MOS, details see in the slides above
1ea test fixture for packaged MOS or bip. transistors (to verify the system
performance)
The Following Equipment Is Usually Available At The Customer:
1ea triax-triax cable 80cm (SMU->Filter)

2ea short Coax->3.5mm cables (Filter->GSG probes, GSG probes ->SR570)
1ea short Coax-Coax cable (SR570->HP35670A)
optional: 2ea Triax(m)->Coax(f) adapters, Trompeter part number ADBJ20-E3-PL75
Measurement Instruments
1ea Stanford Research Amplifier SR570 (http://www.srsys.com )

1ea Agilent 4142 or 415x DC analyzer, other Agilent DC analyzers upon request
1ea 35670 dynamic signal analyzer (no special options are required)
Note: when using a new 35670, make sure that it is NOT set to system controller!
IC-CAP Modules
1ea 85199G IC-CAP Noise Measurement Drivers

1ea 85199A,B,D or 85190A bundle (typically already available at the customer)
1ea 85195BL 1/f noise toolkit (includes 1 Hz filters, however, only applicable for NPN
and NMOS transistors)
Wafer Prober Setup
372
Meas.Setup for Packaged Transistors
Meas.Setup for Resistors, Diodes etc.
Note
Special acknowledgements to the modeling engineers at ON-Semiconductors in Roznov/Czech Republic,
for this smart application.
How to fine-tune the SR570 offset current settings

Important copy from the SR570 Manual
Setting The Input Offset Current
The SR570 can provide a DC current offset to suppress any background currents at the
input.
The offset range can be changed from 1 pA to 5mA (both positive and negative) in
373
discrete increments. Use the up/down arrow keys in the Input Offset section to change the
current level.
In addition to these fixed settings, the user may specify arbitrary currents through the
UNCAL feature. To set an uncalibrated offset current, the user must press both up and
down buttons simultaneously, lighting the UNCAL LED. In this mode, by pressing the up or
down pushbuttons, the user may reduce the calibrated current in roughly 0.1%
increments from 100% down to 0% of the selected offset value. In contrast to other front-
panel functions, when in UNCAL the instrument's key-repeat rate will start slowly and
increase to a limit as long as either button is depressed.
Simultaneously pressing both Offset buttons once again will restore the unit to the
previously calibrated current setting, and turn off the UNCAL light
The same applies to the Sensitivity settings.
374
Spectrum Analyzer
375
Data Verification
Data Verification and Data Consistency Checks
Prologue
Besides good measurements, the measured data should be consistent from Setup to
Setup. Therefore, before performing the first parameter extractions, the data consistency
should be proofed.
Possible Checks
DC data among themselves (e.g. output vs. transfer chacteristics)
S-parameter bias points vs. DC measurements
S-parameter starting points vs. simulations of fitted DC parameters
CV data against S-parameters (which have been converted to CV)
and, using ADS, also linearity of S-parameter measurements
DC vs. DC
In this example, an IC-CAP macro has converted the currents of the foutput measurement
of a bipolar transistor into a pseudo-measured Gummel plot .
IC-CAP Example
1. Select the demo_features directory

2. Goto sub-directory
3_MEAS_ORGANIZE_n_VERIFY_DATA/0_MASTER_FILES/1_CHECK_DATA_CONSISTENCY
3. and load the model file NPN_MEAS_CHECKTOOLS_demodata.mdl and execute the
Transforms 'check_data_consistency' in the different Setups.
Note
The IC-CAP macro applies some data management manipulations like-> save foutput data to an ASCII
.mdm file-> read the data back into IC-CAP, but now with inverted sweep orders. This gives a pseudo-
measured Gummel plot. Link the data of the original Gummel plot into the same plot.
376
Of course, DC-DC data checks can also be performed in the other way:In this example,
the forward Gummel plot data are converted into pseudo-measured foutput data, and
plotted together with the originally measured foutput data.
DC vs. Spar Bias
In the next step, we compare the DC bias conditions of the S-parameter measurements
with the DC measurements.
Note
How the IC-CAP macro manipulates the data, export the S-parameter data including the also measured
DC bias currents-> import back only the DC bias currents-> compare to the originally measured DC
curves.
DC Measurements vs. Spar DC Bias
Spar DC bias data mirrored into foutput
377
To make the test complete, the DC foutput measurement and the DC biases of the S-
parameter measurements are compared too. If there is a difference in the DC currents of
the S-parameters and the standalone DC measurements:
Too much RF signal, too small DC bias power (compared to applied RF signal power)
Self-heating
Voltage drop in S-par testset, 1MOhm shunt resistors to Ground in the S-par testset
Unphysically modeled DC performance
DC modeling was performed at different bias conditions than S-par modeling
CV vs. Spar converted to CV
We are now ready to check the data consistency for the frequency range. Commencing
with CV, the example above shows the comparison of CV data obtained with a standard
LCRZ meter, with pseudo-CV data obtained from converted S-parameters.
Note
Here a quick overview of the IC-CAP macro:
1. De-embed the measured S-parameters

2. Convert the de-embedded S-parameters to Y-parameters
3. Calculate the capacitance out of the imaginary part of the Y-parameters
4. Compare these pseudo-CV data with the original CV meter data.
When comparing CV curves converted from S-parameters, keep in mind that due to the S_to_Y conversion
and the interpretation of these Y-parameters with respect to a PI schematic, the CV curves represent the
capacitance in one of the PI branches. Measuring with a CV meter, depending on the measurement
principle, you may obtain the total capacitance between two nodes (2-pin CV measurement method) or
the individual capacitance (4-pin method, guarded wafer chuck).
Spar Starting Points vs. Existing DC Modeling
378
Since the starting points of the simulated S-parameters, or better said, the extrapolated
points for 0Hz frequency, are determined by nothing but the DC parameters, we should be
able to get a good fit for the low-frequency range even when we haven't yet extracted the
high-frequency parameters like the capacitances or the transit time of transistors.
Important Note and Recommendation

In your daily modeling work, you should perform this test every time you are done with
fitting the DC (and CV) curves and begin with S-parameter modeling.
The starting points of the measured S-par (0Hz)are determined by

the DC fitting
Example Transistor
With RF modeling (capacitors and transit time), only the trace from the starting points
towards the end points can be fitted, but not the starting points.
If this starting point check fails, verify this: too much RF signal> self-heating> voltage
drop in S-par testset> DC modeling was performed at different bias conditions than S-par
modeling.
Discussion
If the starting points of the S-parameter curves do not match the simulations
with the so far extracted DC parameters:
Too much RF signal-> self-heating

Voltage drop in S-par testset
Unphysically modeled DC performance
Solve the problem before you continue with S-parameter modeling. The S-
parameter points for 0Hz can be extrapolated from the two lowest measured S-
parameter frequency points.
Spar linearity checks
379
Last not least, regarding our consistency checks, we can even verify the linearity of the
performed S-parameter measurements with the NWA.
Provided ADS is available, we can enter the so far extracted model parameters, perform
an harmonic balance simulation for a certain fundamental frequency, and convert the
simulation result into 'S-parameters as a function of RF power'. (This is done by
converting the fundamental frequency simulation results into S-parameters. We compare
this result with a linear simulation (what emulates the linear NWA measurement).
We increase the applied RF signal level from e.g. -40dBm up to -10dBm, and check, if the
pseudo-S-parameters from the harmonic balance simulation stay constant, or if they
begin to become a function of applied RF power. As long as they are independent of the
RF power, the device behaves linear. This way, we can backwards check the max.
applicable RF signal for the NWA measurement and the NWA calibration.
Example in IC-CAP:
load the file
demo_features/3_MEAS_ORGANIZE_n_VERIFY_DATA/0_MASTER_FILES/4_RF_LINEARITY/large_signal_RF_NPN_PELdep.mdl
and execute Transform 'TUNE_ANALYZE_xxx' in the harmonic balance Setups.
When S-parameters are affected by the RF signal, then fT is also affected.
Note
Too big RF signal during NWA measurements distorts the fT curve to the left of MAX(fT).
It is thermal self-heating what affects the fT curve at high DC bias, even FOR low RF
signal levels.
Self-heating 'pushes down' the ft for IC bias beyond max(ft).
380
And here's the answer to the previous slide's quiz:
as a general rule, for iC bias conditions beyond max(ft), it is the thermal self-heating what
pushes down the ft points towards lower ft values. Depending on the type of transistor,
also the max(ft) can be pushed down by self-heating.
Self-heating affects the S-parameters (shift of the operating point).
The S21-parameters in this slide depict a transistor measured in pulsed mode (1us pulse
for DC and NWA) compared to a conventional CW measurement. This data is the base for
the fT curves of the previous slide.
A Typical Modeling Mistake
fT modeling at different (higher) iC than DC modeling
Last not least: don't merge diff. measurements of diff dies when extracting model
parameters.
It is obvious that when developing a model of a component, all types of measurements

have to be performed on the same device.
381
Data Management
Measurement Standardization and Data Consistency
Checks
The Concept of Separating Measurements From Extractions
Separate measurements from extractions
1 MASTER MEASUREMENT file

1 MASTER EXTRACTION file
Data Exchange by mdm Files

The separation of measurements from extractions requires as a first step to use 2 model
files:- one for the measurements- another for the extractions.
Both files act as strategy files. The measurement file contains the measurement strategy,
and the extraction file the modeling and parameter extraction strategy.
Therefore, both files can be saved without measured or simulated data!

The measured data are stored in separate .mdm files instead, and the extracted model
parameters in SPICE, ADS, Spectre, ELDO etc. decks.
Required Naming Conventions for the concept of

separated measurements and extractions
As a consequencs, it becomes mandatory to standardize on some naming conventions.
Otherwise, the data exchange by exported (Measurement_Master.mdl ) and then imported
(Extraction_Master.mdl) .mdm files will not work.
Diodes: va, vc, ia, ic (anode - cathode), additionally for varactor diodes: vp, vw (poly -
well)
Inductors: see under S-parameters
Transistors DC and CV:
MOS and MESFET:
DC: vg, vs, vd, vb, ig, is, id, ib

BIP:
DC: vb, ve, vc, vs, ib, ie, ic, is
CV: vbe, cbe, etc. with the first index representing the + node, the second the - node
S-PARAMETERS: freq, S (undeembedded data), S_deemb (deembedded data) dummy1,

dummy2 for 0V dummy Inputs (to satisfy SPICE)
MEASUREMENT_MASTER <> EXTRACTION_MASTER: naming conventions
With this concept, it becomes mandatory to define some naming conventions. Otherwise,
the data exchange by exported (MEASUREMENT_MASTER.mdl) and then imported
(EXTRACTION_MASTER.mdl) .mdm files will not work.
For the data import, an ImportCreate is usually performed, which means that the names
of the Inputs and Outputs in the mdm files become the names of the Inputs and Outputs
of the corresponding master_extraction files. This means, the transforms in this file have
to use these names.
Suggestion for standardized Input and Output names::

DIODES (DIODE_MEAS_MASTER.mdl) va, vc, ia, ic (anode - cathode)
VARACTOR DIODES (MOS_VARACTOR_MEAS_MASTER.mdl) additionally to diodes: vp, vw

(poly - well)
INDUCTORS (SPIRAL_MEAS_MASTER.mdl) see under S-parameters
TRANSISTORS DC and CV
MOS (BSIM4_DC_CV_MEASURE.mdl) and MESFET: DC: vg, vs, vd, vb, ig, is, id, ib
BIP (NPN_MEAS_MASTER.mdl):
DC: vb, ve, vc, vs, ib, ie, ic, is
CV: vbe, cbe, etc. with the first index representing the + node, the second the -
node
S-PARAMETERS freq, S (undeembedded data), S_deemb (deembedded data) dummy1,

382
dummy2 for 0V dummy Inputs (to satisfy SPICE)
In order to re-use measurements in different model extraction files, it is important to
agree on these naming conventions. So, once the data are measured, the modeling
engineer can select the model, without remeasuring the device.
When separating measurementsfrom extraction,we can easilyadd a 3rd stepfor important

datainspections and data consistency checks:the VERIFY.mdl
verify measurement data

rearrange measurement data (vs. L, W, TEMP, wafer location..)
Taking also the most important data consistency checks into account, requires an
intermediate model file layer, i.e. a third model file.
In this VERIFY section, we check for data outliers, scaling effects, network analyzer
calibration verification and de-embedding verification etc. We further verify the
consistency of the measured DC currents with the DC bias currents of the S-parameter
measurements.
It is also in this intermediate layer where non-parametric statistics could be applied to
identify the measurement data of the 'golden device' and the 'boundary devices' for the
subsequent parameter extraction in the model file EXTRACT.mdl
Going more into details, this slide shows the realization of the previously presented
scheme, using
a master model file for wafer and thermochuck control

a master model file for performing the measurements
mdm files as the measurement data base,
several model files for data verification
non-parametric statistics for golden device identification and finally
the master model file for extraction.
Project Flow Suggestion for Reusable Model Files

This is an example of a complete project with measurements separated from extractions,
applying data consistency checks, plus re-use of centralized PEL programs (DEPOTS).
With this concept, the measurement data are stored in mdm files,while the model files are
all saved without measurement data!
IC-CAP examples file: NPN_MEAS_MASTER.mdl

NPN_MEAS_CHECKTOOLS.mdlVBIC_EXTRACT_NPN.mdl
CAL_VERIFY.mdl
CHECK_DEEMB_OpenShortThru.mdlcombine_mdms.mdl
383
In this example, we will discuss the use of a centralized PEL program called 'DEPOTS' plus
the separation into a MASTER_MEASUREMENT model file, several
MASTER_DATA_CONSISTENCY_CHECK files and a MASTER_EXTRACTION model file.
The project flow for the above sketched scenario is as follows:
1. NPN_MEAS_MASTER (MASTER_MEASUREMENT)
This is a model-independent, but device specific file for performing all required
modeling measurements plus the additional documentation measurements like
contact resistance, NWA calibration quality, dummy structure S-parameters
etc.During this task, there is an
Export of the NWA calibration verification measurements, i.e. the (re-)measured

OPEN, SHORT, LOAD and THRU calibration standards after the NWA calibration into
model file CAL_VERIFY
export of the S-parameter dummy structure measurements into model file
CHECK_DEEMB_OpenShortThru
call to re-usable central de-embedding PEL routines in model file DEPOTS.
1. Then, we export the measurement data into a mdm file data base
2. We import the mdm files into model file NPN_MEAS_CHECKTOOLS where for example
the DC curves are compared to the DC bias curves of the S-parameters (identifying
too big RF signal, self-heating, contact problems etc.), or where CV measurements
are compared to S-parameter measurements. It is also here where special scaling
effects are analyzed. For example, for MOS transistors, .mdm files of the
ransconductance vs. L and W are composed here for later VTH extraction. Also,
outliers in the .mdm data files are identified and eliminated in this step.
3. Finally, we import the mdm files finally into model file VBIC_EXTRACT_NPN. This
VBIC extraction re-uses central diode modeling extractions (like for IS, N, CJO, VJ
and M) for VBIC from model files DEPOTS.
Note
Because of the completeness of the measurements in the NPN_MEAS_MASTER file, we also could import
the data to any other bipolar MASTER_EXTRACTION model, Gummel-Poon, VBIC or Hicum. With this
concept, the measurement data are stored in mdm files, while all the model files are all saved without
measurement or simulated data. This is why there is a small button called 'Save without
Measured/Simulated Data' when saving model files. Besides the discussed project flow for the bipolar
toolkits for Gummel-Poon*, VBIC*, and Hicum*, this concept is also used in the IC-CAP BSIM and
PSP toolkits*, as well as in the toolkits for spiral inductors, varactor* and diode* modeling.
IC-CAP examples files of the discussed bipolar project flow:
NPN_MEAS_MASTER.mdl
DEPOTS.mdl
cal_verify_MASTER.mdl
check_deemb_OpenShortThru.mdl
NPN_MEAS_CHECKTOOLS.mdl
combine_mdms.mdl
GP_EXTRACT_NPN.mdl
VBIC_EXTRACT.mdl
Contacts:
Franz Sischka, Agilent EEsof, [email protected]

www.xmodtech.com
www.admos.de
Examples for Data Consistency Checks

BEFORE MODELING:
cross-checking for DC bias conditions

verification of NWA calibration
verification of de-embedding
verification of S-Parameter bias currents vs. DC curves
comparing CV measurements versus CV from S-Parameters
AFTER MODELING:
check nonlinear RF performance of obtained parameter set

IC-CAP examples file:
Before modeling:
demo_features\3_MEAS_ORGANIZE_n_VERIFY_DATA\0_MASTER_FILES\10_CHECK_DATA_CONSISTENCY\NPN_MEAS_CHECKTOOLS.mdl
After modeling:Verification of the linearity of the NWA measurements (after the DC model
parameters have been extracted):
demo_features\3_MEAS_ORGANIZE_n_VERIFY_DATA\0_MASTER_FILES\40__RF_LINEARITY\
large_signal_RF_diode.mdl
large_signal_RF_xtors.mdl
Example: DC vs. DC Measurements mirror foutput into fgummel
384
Example:DC Measurements vs. Spar DC Bias
Advantages of Standardization
no double PEL programs any more
ease of use
clear structure of strategies
However...
it takes some effort to convert existing 'hacked' solutions into standardized, reusable
modeling schemes
385
Extraction
Contents
Curve Fitting Techniques (iccapmhb)
386
Curve Fitting Techniques
Introduction
Contents
Synthesizing Measurement Data for Verification of Extraction Methods On the Web

(iccapmhb)
Regression Analysis (iccapmhb) Articles
Applying Linear Regression Analysis to Modeling (iccapmhb) Knowledge Center
Direct, Visual Parameter Extraction (iccapmhb) Technical Support
General Modeling Techniques (iccapmhb) Training
Data Interpolation (iccapmhb) Videos
387
Synthesizing Measurement Data for
Verification of Extraction Methods
Using IC-CAP for the extraction of model parameters offers a lot of flexibility in terms of
creating user-defined models and implementing the corresponding extraction routines.
But when developing a new extraction strategy, we may run into two major problems:
do the routines extract the parameters correctly?
is the model able to fit the measured device at all?
Related to the first problem: are the extraction routines correct ?

Assume that we know the precise parameter values in advance. If we were able to
generate quasi-measured data out of these parameters, then the extracted parameters
would have to provide exactly these parameters back again!
Using IC-CAP, it is simple to perform such a check. The trick is to 'synthesize' quasi-
measured data out of a set of parameters and to apply then the extraction routines to
these data. This can be done as follows:
1. Define a measurement Setup in IC-CAP, for which the extraction routines shall be
tested. Example: an bipolar output characteristic for an Early-voltage extraction.
2. Select a 'typical' set of parameters (no default values like 'zero' or 'infinite', but
instead real realistic values!)
3. Change the 'Output' data type to 'S' (simulated only). The array behind that Output
is now simulation data only.
4. Simulate this Setup with these known parameter values.
5. Change the 'Output' data type back to type 'B'.
IC-CAP doubles now the data field to measurement and simulation data.
This means: the simulated data of step 4. is now converted to measured data!
6. Reset the model parameters by clicking 'Reset to Defaults' and simulate the Setup
using the default parameters.
7. Apply the extraction routine-under-test and check the values of the extracted
parameters.
Provided we get the parameter values back within a good tolerance, we can be sure that
the extraction works correctly. If we now apply the extraction to real-world measured
data, we should obtain the right parameters. This is true if the measured data have the
same shape like the model equations! If not, we might have to choose another model or
go for subcircuit modeling.
This means, we can now decide whether the model is able to fit the meas. data !
Note
The problem of whether the actual device can be described by the selected model or not, can also be
checked by using the method of 'Direct, Visual Parameter Extraction'. If this method does not produce a
parameter plateau, the DUT does not behave in the manner described by the model.
388
Regression Analysis
Regression Analysis is a technique which provides a best curve fit for given data sets.
Mathematical Basics of Curve Fitting

Let's assume we made 'N' measurements yi at the stimulating points xi.
I.e. we obtained the array [ xi, yi ], which is then shown in a plot.
A curve Y( x ) shall be fitted to this array of measured data points using least square
curve fitting technique.
Referring to an individual measurement point, the fitting error is:
and for all data points:
This error shall be minimized.
The fitting will be done by varying the coefficients of the fitting curve of equation (2). The
minimum of the total error E depends on the values of these coefficients. This means, we
have to differentiate E partially versus the curve coefficients and to set the results to zero.
We obtain a system of equations, solve it, and get the values of the coefficients for a best
curve fit. This is known as regression analysis.
Note
This regression analysis is simple for a straight line fit. But in general, measured data is non-linear.
Unfortunately, a non-linear regression analysis can be quite complicated. This problem can be solved if we
use a suitable transformation on the measured data. This means that the measured data is transformed to
a linear context between the yi- and the xi-values. As will be seen in the diode example later, this is a
pretty smart way to get the curve fitting parameters easily without much calculations.
Linear Curve Fitting

IC-CAP File:
demo_features/5_PEL_PROGRAMMING/3_PARAM_EXTRACTIONS_CURVE_FITTING/1_basic_PEL_extractions/1fit_line.mdl
Provided we have got an array of N measured data points of the form [ xi,yi ].
A linear curve with the equation
shall be fitted to these points. This situation is depicted below.
Linear regression applied to measurement points
The error of the i-th measurement is:
Using the least means square method following equ.(2) yields:
Partial differentiation versus slope 'm' gives:
and versus y-intersect 'b':
389
We obtain from (5) after a re-arrangement:
and from (6):
Multiplying (7) by -N and (8) by xi and adding these two equations allows the elimination
of the coefficient 'a', and we can separate the slope 'b':
or:
and from (8) for the y-intersect 'a':
with 'b' according to (10).
With equations (10) and (11), we determine the values of the two coefficients of the line
which fits best into the 'cloud' of measured data.
Finally, a curve fitting quality factor r2 [ regression coefficient ] is defined.

Its value ranges from [ 0 < r2 < 1 ].
The closer it is to 1, the better is the fit of the linear curve.
with 'b' from (10)
General Formula for Regression Analysis

Linear Regression:
Quadratic Regression:
Cubic Regression:
continue with the evolution of matrices like above.
390
Application: Calculating the quadratic polynomial
parameters from measurement data
IC-CAP File:
demo_features/5_PEL_PROGRAMMING/3_PARAM_EXTRACTIONS_CURVE_FITTING/1_basic_PEL_extractions/1fit_quadratic.mdl
Starting with equ.(14) from above
we get the following system of 3 equations:
Equations (15a) .. (15c) are of the form
for which we have already a solution for a, b and c in chapter 6.Data_Interpolation.
Therefore, with the following abbreviations:
we can re-use the quadratic data interpolation formulae of the chapter Data Interpolation:
and
and are ready to get the quadratic fitting function y( x )
In the plot below, the red curve (dummy_data) has been fitted by the magenta curve
(QUADFIT), inside the marked box.
391
392
Applying Linear Regression Analysis to
Modeling
For IC-CAP examples, see under directory
$ICCAP_ROOT/examples/demo_features\5_PEL_PROGRAMMING\3_PARAM_EXTRACTIONS__CURVE_FITTING
1_basic_PEL_extractions
and also .../specific_PEL_routines
DC curves:
Applying linear regression analysis to DC modeling data is pretty obvious. Examples
include the Early voltage VAF of a bipolar transistor, the threshold voltage of a MOS or the
saturation current IS and the exponential coefficient N of a DC diode characteristic.
For more details see the mentioned IC-CAP demo files, or the chapter on diode modeling.
As a brief overview, we consider the diode DC current modeling:
with VT: temperature voltage 27mV at 25C
or
DC characteristic of the diode under forward bias and the determination of the DC parameters IS and N
Provided vD > 0, i.e. neglecting the term ( -1) in equ.(1), and applying a logarithmic
conversion yields:
This is an equation of the form:
In order to interpret (2b) linearly, we have to substitute:
And this explains how to manipulate the measured data: after the logarithmic conversion
of the measured values of (2c), they are introduced, together with the still linear
values of vD (2f) into the regression equations (10) and (11) of the previous chapter as
yi- and xi-values. We obtain the y-intersect b and the slope m of the linear regression
function. Solving (2d) for Is and (2e) for N we finally are able to calculate these two
parameters out of b and m as follows:
393
CV curves:
When modeling non-linear components such as diodes or transistors, space charge
capacitor curve fitting is always an issue. These are voltage dependent capacitors (CV
curves). And since this is a central modeling issue, some different methods of applying
regression techniques to this problem are discussed in this chapter.
The general space charge capacitance Cs formulation for voltages vD < FC * VJ is:
and else
with
CJ: space charge capacitance at vD = 0V

VJ: diffusion voltage (built-in potential), models the pole of equation (1a) (typ. 0,7V)
MJ: space charge exponential factor (models the slope of the CV curve)
(abrupt pn transition (<0,5um): MJ = 1/2)
(linear pn transition (> 5um): MJ = 1/3)
FC: switching coefficient in forward mode, default 0,5, see left picture in figure 1.
Parameter FC switches from the hyperbolic model equation to a linear continuation for vD>FC*Vj
For most simulators, the capacitances are rather described by the charges.
This means for vD<FC*Vj
General CV modeling
For simplicity, the capacitor voltage is always less than FC*VJ for parameter extraction.
i.e. we start with equ. (1a), repeated here again for convenience
A logarithmic conversion of equation.(1a) yields
This equation can be linearized following
394
when substituting:
In order to apply regression techniques, we have to transform the stimulus voltage values
vD[ i ] following (5) and the measured capacitance values Cs[ i ] following (2). Under the
assumption that ~0.3 < VJ < ~1 is valid, we use for the transformation of vD[ i ] a
starting value for VJ, e.g.. 0.3V. These two data arrays x[ i ] and y[ i ] are applied to a
regression analysis, and we get a certain slope m(VJ) and a y-intersect b(VJ). This fitting
is also related to a certain fitting factor r, the regression coefficient. The values of m, b
ad r are stored and then the parameter VJ is incremented to e.g. 0.31V. Then, the same
data transform is applied to the stimulating and measured data x[ i ] and y[ i ] , and
another regression analysis performed. From that we get another triplet of coefficients m[
VJ ], b[ VJ ], and r. If r is now better than before, we continue by stepping VJ, if it is
worse, VJ is stepped in the other direction. When r reached its maximum, the best fitting
coefficient r is obtained and the corresponding VJ_opt, m(VJ_opt) and b(VJ_opt) are now
known.
This allows to re-substitute:
from (4): MJ = - m[ VJ_opt ]
and from (3): CJ = exp ( b[ VJ_opt ] )
In practice we often have the problem that the measured space charge capacitance is
overlaid by an offset capacitor (package, test structure etc.). This leads to the problem of
determining that offset capacitance as well. This is covered by the following two
suggestions.
CV modeling with MJ=const, but including offset

capacitance C offs
In this case, the fitting formula is
This equation can also be made linear, yielding

y=mx+b
if the following substitutions are made:
y = Cs
m = CJ
with
MJ = constant
b=
I.e.: we preset MJ to a fixed value:

abrupt PN junction (<0,5um): MJ = 1/2
linear PN junction (> 5um): MJ = 1/3
Using the same procedure as in the previous chapter, VJ is incremented/looped until the
best regression factor r is found.
Reverse substitution gives:
CJ = m(VJ_opt)
and
= b(VJ_opt)
where, as noted above, MJ has been specifically predefined.
Total CV modeling including offset capacitance

Assuming that the CV value at vD=0 is known, there is a more 'refined' substitution
method that will determine all the CV parameters as well as the offset capacitance Coffs.
Again:
395
First, to eliminate the offset capacitance C offs, Cs is derived by vD:
Next, the logarithm is determined:
then a linear substitution is again performed:
in such way that
Here again, VJ is incremented until the maximum regression quality, r 2, is reached. The
following reverse substitution is made for this VJ_opt:
Finally, Coffs is calculated to:
Regression Analysis for S-Parameters

When modeling S-parameters, the problem of fitting curves to circle sectors again is a
common problem. Here too, the modeling equation of a circle with its center on the real
axis of the Smith chart can be transformed linearly ( ! ) using the following formulation:
Circle equation with the center on the real axis:
or:
This form can also be linearized ( ! ), into
with the following settings:
In other words: the measured values x[ i ] and y[ i ] are employed in equation (2a) and
(2d). Then, the vector ylin[ i ] = x i [ i ] 2 + y i [ i ] 2 is plotted against xlin[ i ] = x i [ i ],
and a linear regression line fitting is applied.
From the slope m we obtain (2c)
and from (2b) solved for r
Summary
This section illustrated how linear regression can be applied to various modeling problems.
396
In practice, this method has proven simpler than applying special regression rules for
nonlinear curve fitting. This is because device modeling, the equation to be fitted is
known. A that remains is to determine a nonlinear transform for that modeling equation to
which linear regression can be employed.
397
Direct, Visual Parameter Extraction
As an additional possibility for parameter extraction, this chapter presents a method for
which measured values are directly transformed into parameter values based on a known
model equation.
Example, the MOS transistor:

The threshold voltage of different MOS transistor sizes, plotted versus length or width,
shows a certain dependency. In old MOS models, the model parameter VTH was the only
mean to model VTH. In other words, no dependency of the threshold voltage vs. geometry
was included. Applying the visual parameter extraction method, we can develop an
improved MOS model by replacing (in the model equations !) the constant parameter VTH
by a function of VTH = f(L, W), as obtained by our visualized parameter plots. L and W are
the length and width of the MOS transistor.
By the way, this is how company-internal models are developed and improved very often.
Another example: the bipolar transistor:

You know that the x-intersect of lines fitted to an output characteristic of a bipolar
transistor should hit always the same point, the Early voltage. If we apply an IC-CAP PEL
(parameter extraction language) program to calculate the x-intersect of a line that is fitted
to two adjacent measured points, and if we display the result of this operation versus the
collector voltage (first order sweep), we will obtain a plot of the 'equivalent' Early-voltages
of adjacent measurement points.
The advantage of using this visual extraction method is that we can see clearly, if the
model is able to fit the measured data at all. We only have to check if there is a flat region
in the transformed data domain or not. If it is there, we can extract the parameter very
simply by calculating the mean value of the flat region. And we know at the same time, in
which range the parameter is dominant and can therefore be used for fine-tuning with the
optimizer. If there is no flat range, the model cannot fit the measured data. We could vary
the parameter as much as we like and would not achieve a fit of the simulated to the
measured data.
For the application of this method, we start with some basic equations that refer to
following figure:
About the definition of some basic equations for the direct visual parameter extraction
Assumed we have
where m: slope
y0: y-intersect
Then it is:
and:
This finally gives the basic formulas for this modeling method:
starting with
398
we get:
The following plots give some examples on how to apply this idea to the parameter
extraction of a bipolar transistor using the Gummel-Poon model. It should be mentioned
that this method can be applied to all the parameters of this model, as well as to other
models like Statz, Curtice, BSIM etc.
IC-CAP demo files about how to apply 'direct visual parameter extraction' to the Gummel-
Poon model are available from the author.
The first example is about modeling the parasitic resistors of a bipolar transistor.
RE is the derivative of the flyback measurement
The "visualized" RE is the y-intersect of two adjacent data points from that flyback plot.
In this figure, the proposed flyback measurement method and its interpretation to
represent the behavior of the Emitter resistor RE (assumed to be an ideal ohmic resistor)
seems correct for higher Base currents. Therefore, we extract its value out of the range
marked with the box sign.
In the next example, we will see how to apply the visual parameter extraction method to
output characteristics of a bipolar transistor.
the Early voltage is calculated from the x-intersect of a line through two adjacent data points
399
In the example of above figure, we see that the Gummel-Poon model is nicely modeling
the increasing slope of the output characteristics with its parameter VAF. Its value is the
mean of the data inside the marked box range.
Let's study also the application of the "visual" method for the Gummel-Poon plot:
interpretation of the Gummel-Poon plot for parameter extraction
The transformation of measured data directly into the parameter domain works well also
for the examples depicted in figures 6 and 7.
In our example, it can be clearly seen that the Base current of that particular transistor
does not exhibit the recombination effect (NF in Fig.6 is flat for low vB values).
Note: This is the famous 'knee' in the iB curve visible at lower Base currents.
400
General Modeling Techniques
When developing an equation for a new model, this equation has to follow the
measurement data. In order to develop such an equation from simple partial equations,
we can distinguish two situations: Simple adding of partial equations ('summing') or
inversely adding ('paralleling').
Adding of functions allows to model a curve steepening
While modeling a steepening of a curve with adding of two functions is pretty straight
forward (fig.1), the modeling of a declining function is more complex. However, applying
'paralleled' functions, following the equation form
allows to also model more complex curve traces, by inversely adding their max. limiting
partial functions, see fig.2.
Paralleling of functions allows to model a curve decline
Diode DC modeling
When measuring the DC characteristics of a diode, we have basically 4 regions to model.
With increasing bias voltage, this is
1. the recombination effect at low currents

2. the 'conventional', ideal diode region
3. a saturation effect
4. the ohmic region.
Compared to the conventional, ideal diode region, we typically have another linear trace in
a semi-logarithmic plot for the low-current region, i.e. the recombination effect. For a
given low voltage vD, there is a higher current flowing than expected from the
'conventional' ideal diode model. This additional current (at the same applied voltage bias)
is therefore modeled with a parallel diode DRECOMB.
Then, with increasing vD, comes the 'ideal' diode range, modeled by diode DIDEAL.
In other words, with increasing vD, DIDEAL now conducts more and more of the current
iD.
Before the ohmic resistor affects the curve, there is often a saturation diode visible in the
transition range between the ideal diode and the ohmic range. Since this corresponds to a
voltage shift (for a certain, applied current bias), this is a series diode, in series to the so
far described parallel diodes. Its effect can be described like this: when the total diode
current curve begins to decline, we see an additional voltage drop for a given current. For
currents below that value, the voltage drop is zero. Therefore, the IS parameter of the
DSAT diode is exactly the 'knee' of the diode DC curve. If, above the knee, the slope is
half of the ideal diode, the N parameter of DSAT equals N of DIDEAL. If it is flatter, N of
DSAT is even bigger.
401
After that DSAT region, a series resistor is used to model the ohmic losses.
This means, we end up with a schematic like below:
Schematic for modeling the DC performance of a diode
Note
if you do not want to use DSAT for the CV modeling, treat this diode as an RF short, i.e. set DSAT
parameter CJO=100m.
The following measurement result and the modeling steps illustrate the idea for the DC
modeling of a typical diode.
MOS Gate Oxide Capacitance (simplified model)

MOS transistor. In this case, the measured trace (see fig.3) can be interpreted as two
series capacitors,
one of constant capacitance C oxide

and a second one of varying C inversion, after
Overlying C oxide with C inversion yields the Gate Oxide Capacitance of a MOS transistor
402
403
Data Interpolation
IC-CAP File:
demo_features\3_MEAS_ORGANIZE_n_VERIFY_DATA\
Interpolate_DC_CV_by_QuadrPolynomial____Get_inner_device_volt.mdl
Fitting a quadratic function to three data points:
For three data points [ x0, y0 ], [ x1, y1 ], [ x2, y2 ], we start with a set of 3 quadratic
equations
Goal: We need to get the coefficients a, b and c as a function of xi and yi

From (1):
(4) into (2):
In the same way, from (3):
(6) into (2):
Finally, (7) into (5):
Conclusion:
(6), (7) and (8) represent the equations for the polynomial coefficients a, b and c.
404
Simulations
Introduction
Contents
History (iccapmhb) On the Web

Simulator Interface (iccapmhb)
Netlist Syntax (iccapmhb) Articles
CV and S-Parameters (iccapmhb) Knowledge Center
Harmonic Balance (iccapmhb) Technical Support
Training
Videos
405
History of Circuit Simulators
The origin of Berkeley SPICE
SPICE (Simulation Program with Integrated Circuit Emphasis) was introduced in May 1972
by Prof. D.O.Pederson at the University of Berkeley, California. It simulates the electrical
behavior of circuits based on discrete components In particular, these include:
Resistors, capacitors, inductors

Coupled inductors ideal lines and cables
Independent and controlled current and voltage sources
Semiconductor components
It soon became a standard for electronic circuit simulation, used by both, the industry and
the universities. This fact is based on the technical expertise of the developers of SPICE:
especially Prof. Pederson, and Prof. Nagel.
After 17 revisions of SPICE, SPICE2 was introduced in 1975. The main improvements were
the dynamical RAM allocation capabilities and the automatic time stepping for time-
domain simulations. Especially SPICE version 2g6, was a specially reliable program
version, written in Fortran. This version has become the root of many industry in-house
simulators and also of many commercially available simulators. This explains why they
typically understand the Berkeley syntax (Spectre, Hspice, Pspice, Eldo).
Finally, SPICE3, introduced by the end of the 80s, was the first version written in the C
language.
One of the reasons for the success of SPICE is also the implementation of good transistor
models. This is especially true for the bipolar model. The Gummel-Poon model has only
become popular after it was implemented in SPICE. (However, some simplifications had
been made to the original paper of Gummel-Poon.) The same is true for MOS transistor,
however, after MOS3, and the introduction of the BSIM series of model, SPICE's model
quality had suffered a lot compared to MOS2 and MOS3. Finally, with BSIM3v3, Berkeley
SPICE gained back its authority of model definition.
The origin of ADS

The origin of MDS simulator goes back to the late 1960's when Hewlett-Packard
introduced the simulator OPNODE which was a linear s-parameter simulator. In 1984 HP
wrote a linear simulator called Time-Domain Tutorial which could be used to take a
transmission line topology, compute the S-parameters, and use the 8510 to get a time
domain reflectometer plot. This time domain tutorial program eventually evolved into the
MDS Microwave Linear Simulator MLS in MDS A.01.00.
Ken Kundert worked on MLS until 1985 when HP funded him to get his Ph.D at U.C
Berkeley working on nonlinear microwave simulation based upon harmonic balance
technology. HP was co-developing harmonic balance in conjunction with Ken Kundert and
Berkeley. The internal name for this simulator was "Spectra" which was later named the
HP Microwave Nonlinear Simulator or MNS. MNS essentially evolved from Berkeley's code.
The Berkeley code was named Spectre ( a subtle difference ). After Ken Kundert received
his doctorate, he left HP to work at Cadence where he built a new time domain simulator
based upon the UC Berkeley Spectre code, hence the Cadence Spectre simulator was
born.
In 1996, Hewlett-Packard bought EEsof, and MDS and the EEsof simulator products were
merged. The new product was called ADS (Advanced Design System).
References
1. Nagel, L.W.; Pederson, D.O.: SPICE. Berkeley, Univ. of California, Electronic
Research Laboratory, ERL-M 382, 1973
2. Nagel, L.W.: SPICE2: A Computer Program to Simulate Semiconductor Circuits.
Berkeley, Univ. of California, Electronic Research Laboratory. ERL-M520, 1975
406
Simulator Interface of IC-CAP
IC-CAP always links to separate, standalone simulators. However, to make this concept
work independently of the customer's simulator capabilities, three types of simulators are
added to the CD-ROM containing IC-CAP:
UCB Spice2g6, (University of California, Berkeley) which is the base of most

customer-specific simulators
UCB Spice3e2
HPSpice
Since IC-CAP 2006, the license for using ADS linear (DC, CV, S-par) and transient
(TimeDomain) simulations is included to IC-CAP. This allows the user to use the latest
versions of the industry-standard models like PSP, Hicum, VBIC etc.
Spice3 is no longer supported by UCB since the early 2000, and therefore, does no longer new
models introduced to the marked since then.
ADS needs to be installed additionally to IC-CAP.
This means that, provided the user has no local simulator available, he can always
perform simulations based on ADS or on these 3 types of Spice simulators.
On the other hand, IC-CAP can also interface to other types of simulators. Links, included
in the IC-CAP Analysis Module (see IC-CAP file $ICCAP_ROOT/lib/iccap/usersimulators),
are depicted in fig.1 below. It should be mentioned that these simulators need not
necessarily be available on the local IC-CAP workstation, but that also remote simulation
is supported (see again the IC-CAP file $ICCAP_ROOT/lib/iccap/usersimulators).
Using IC-CAP's PEL language, or the userC functions, it would be also possible to include
the model equations directly into IC-CAP. However, this would imply the following
disadvantages:
if the model equations change, esp. true for user-specific models, this has to be
reflected in the IC-CAP model equations.
if an additional component has to be added to the specified model in IC-CAP, one
would have to redefine all model equations. Example: an extra series resistor for a
diode model would reduce the voltage drop across the implemented model equations,
thus affecting the diode equations.
In order to avoid this dilemma, it is desirable to always use external simulators. In this
case, and coming back to the example from above, an additional series resistor is simply
added to the simulator sub-circuit. Then, there is no need to modify any model equations
inside IC-CAP. The only modification would be to add a series-resistor parameter
extraction module in order to determine the value of this additional parasitic resistor.
Because IC-CAP also supports simulators which allow the user to specify own proprietary
model formulas like ADS (Verilog-A, sdd), Eldo (Verilog-A, HDLA, ADVance-MS) etc., it is
possible to also use user-specific models.
For examples, see
demo_features\6_SIMULATORS\2_ADS\5_NETLIST_TEMPLATES\0_USER_DEF_MODELS
How IC-CAP interfaces with the simulators.
407
Step-by-step:
1. After the component has been measured, the user specifies a suitable sub-circuit in
IC-CAP's Circuit description window as shown in (fig.3)
2. Then, the model parameters are extracted and exported from the IC-CAP Transform
hierarchy into the Model Parameter window (which, on the other hand, is a
consequence of the user-specific Circuit Window), see fig.3 again.
3. When a simulation is requested by the user (or the optimizer), IC-CAP takes the
circuit deck of the Circuit Window, replaces the default values defined there with the
actually extracted model parameters from the Parameter List Window, adds the
stimulus information of the Setup Inputs and also the requested type of simulation
specified with the Setup Output(s) and sends the whole information as a simulator
input deck to the simulator.
4. The simulation is executed, locally or remote, using the syntax 'simulator input_deck
output_deck'
5. The simulator perform the calculations and exports its simulation results into two
files: one ASCII file containing the execution time and the error messages, and a
binary file containing the very simulation result.
6. IC-CAP reads the binary file and places the simulation results under the Output icons
of the Setup window. The ASCII file is only read if the Simulation Debugger window is
open. Otherwise, this file is ignored. (See fig.4)
Fig.3: First, a suitable sub-circuit is selected and defined for the simulations. Then, the exacted
model parameters are exported into the Model Parameters window. The model parameters showing
up in this window are a consequence of the user's sub-circuit definition.
408
409
Netlist Syntax for ADS and SPICE
Contents
Comparing Spice and ADS Syntax (iccapmhb)
Common Parameter List (iccapmhb)
UCB Spice Netlist (iccapmhb)
Spice3 User Defined Model (iccapmhb)
ADS Simulator Netlist (iccapmhb)
ADS User Defined Model (iccapmhb)
Spice ADS Header (iccapmhb)
410
Comparing UCB Spice and ADS Netlist
Syntax
Table 1: summary of the most commonly used
components
Component Syntax for Berkeley Syntax for ADS
SPICE
Resistor R1 1 2 1k R:R1 1 2 R=1k
Capacitor C1 1 2 1p C:C1 1 2 C=1p
Inductor L1 1 2 1n L:L1 1 2 L=1n
Coupling between inductors L1 and L2 K1 L1 L2 .8 Mutual:M1 K=0.5 Inductor1="L1"
Inductor2="L2"
Ideal Delay Line T1 1 0 2 0 Z0=50 #uselib "ckt" "TimeDelay" TimeDelay:T1 1 2
TD=10n ZRef=50 Delay=10p
Independent Voltage Source between V1 1 2 5 V_Source:V1 1 2 Vdc=5
node 1 and 2, with 5V
Diode D1 1 2 myDiomodl myDiomodel:d1 A C
.MODEL myDiomodel model myDiomodel Diode
D Tnom = 27
+ IS = 1p Is = 1p
... ...
Gummel-Poon bipolar transistor Q1 1 2 3 myGPmodel myGPmodel:Q1 C B E
.MODEL myGPmodel model myGPmodel BJT\\NPN=yes
NPN Tnom= 27
+ IS = 1p Is = 1p
... ...
Subcircuit Call X1 1 2 3 mySub mySub:X1 1 2 3
Subcircuit .SUBCKT mySub 10 define mySub 10 20 30
20 30 ...
... end mySub
.ENDS
Note: most simulators except Berkeley SPICE permit the use of node names instead of numbers
411
Common Parameter List
How to have a common Parameter List for Spice and
ADS
If you plan to do device modeling with exchanging simulators, but maintaining the
extractions and optimization steps in your IC-CAP model file, you should set us the
netlist(s) so that the Parameter list is the same. In this case, your extraction routines and
optimizer settings stay the same, even when exchanging the netlists.
SPICE ADS IC-CAP Modelfile Parameter Table

.subckt test 1 2 define test 1 2
R1.R 1 11 .2 R:R1 1 11 R=.2 R1.R 0.2
L1.L 11 12 50p L:L1 11 12 L1.L 50p
L=50p
C1.C 12 2 31f C:C1 12 2 C=31f C1.C 31f
.ENDS end test
And here an example for the corresponding, common parameter extraction PEL code for
SPICE and ADS:
R1.R = 4711
412
UCB SPICE Simulators Netlists
(detailed)
For UCB Spice (University of California Berkeley), start every netlist with a comment line,
for example
*this is my great circuit
Table1:SPICE Element Component Specifications
Component General Form Example

Resistor RXXXXXXX N1 N2 VALUE R1 1 2 1000
<TC=TC1<TC2>> TC=0.001,0.015
Capacitor CXXXXXXX N+ N- VALUE <IC=INCOND> COSC 15 2 10U IC=3
Inductor LXXXXXXX N+ N- VALUE <IC=INCOND> LSHUNT 3 29 10U
IC=15.7m
Mutual Inductor KXXXXXXX LYYYYYYY LZZZZZZZ VALUE K43 LAA LBB 0.999
Transmission Line TXXXXXXX N1 N2 N3 N4 Z0=VALUE T1 1 0 2 0 Z0=50
<TD=VALUE>+ <F=FREQ TD=10NS
<NL=NRMLEN>>
<IC=V1,I1,V2,I2>
Linear Voltage- Controlled Current GXXXXXXX N+ N- NC+ NC- VALUE G1 2 0 5 0 0.1M
Source
Linear Voltage- Controlled Voltage EXXXXXXX N+ N- NC+ NC- VALUE E1 2 3 14 1 2.0
Source
Linear Current- Controlled Current FXXXXXXX N+ N- VNAM VALUE F1 13 5 VSENS 5
Source
Linear Current- Controlled Voltage HXXXXXXX N+ N- VNAM VALUE HX 5 17 VZ 0.5K
Source
Independent Voltage Source VXXXXXXX N+ N- <<DC> DC/TRAN VIN 12 0 DC 6
VALUE>
+ <AC <ACMAG <ACPHASE>>>
Independent Current Source IXXXXXXX N+ N- <<DC> DC/TRAN ISRC 23 21 AC
VALUE> 0.333 45.0
+ <AC <ACMAG <ACPHASE>>>
+ SFFM(0 1 10K 5 1K)
Table2:SPICE Semiconductor Component Specifications
Component General Form Example

Junction DXXXXXXX N1 N2 MNAME DCLAMP 3 7 DMOD 3.0 IC=0.2
Diode + <AREA><OFF><IC=VD>
BJT QXXXXXXX NC NB NE <NS> MNAME Q2A 11 26 4 20 MOD1
+ <AREA> <OFF> <IC=VBE,VCE>
JFET JXXXXXXX ND NG NS MNAME J1 7 2 3 JM1 OFF
+ <AREA> <OFF> <IC=VDS,VGS>
MOSFET MXXXXXXX ND NG NS NB MNAME M1 2 9 3 0 MOD1 L=10U
+ <L=VAL><W=VAL><AD=VAL><AS=VAL> W=5U
+<PD=VAL><PS=VAL><NRD=VAL><NRS=VAL>
+ <OFF> <IC=VDS,VGS,VBS
Note: see also the UCB Spice Maunal in section Spice/Publications.
413
User-defined Model Equations in
Berkeley SPICE 3
see the ModelFile: demo_features\6_SIMULATORS\1_UCB_Spice\user_def_equat.mdl
Nonlinear resistor in Berkeley SPICE syntax:
The UCB SPICE3 syntax is:

.SUBCKT nonlinR HI=1 LO=2
Bnonlinr 1 2 I = A*(V(1)-V(2)) + B *(V(1)-V(2))^2 + C * SQRT(V(1)-V(2))
NOTE:
(Bxxx is a user-defined SPICE model, which may contain either a voltage (V=...) or a user-defined current
(I=...))
.ENDS
However, the IC-CAP circuit parser has some problems with this line, and prints an error.
Therefore, in the spice deck line below, #echo switches the IC-CAP parser off for the
whole line.
As a consequence, however, IC-CAP will not detect the parameters
A, B, C etc. and export them into the Model Parameters area.
This means, they cannot be extracted or optimized from within IC-CAP !
Therefore, the $mpar statement tells IC-CAP to export the specified parameter into the
Model Parameters table what is between the parenthesis of this command.
.SUBCKT nonlinr HI=1 LO=2
#echo Bnonlinr 1 2 I=$mpar(A=1.5m)*(V(1)-V(2))
$mpar(B=5m)*(V(1)-V(2))^2+$mpar(C=2m)*SQRT(V(1)-V(2))
.ENDS
Gummel-Poon iC characteristic in Berkeley SPICE syntax:
.SUBCKT fgummelC C=1 B=2 E=3
#echo BfgumC 1 3 I=$mpar(IS=1E-25)exp((V(2)-V(3))/($mpar(NF=1)

$mpar(VT=1)))
.ENDS
414
ADS Simulator Netlist Syntax (detailed)
Due to its HF application, ADS uses a somewhat modified circuit deck syntax compared to
SPICE.
For example, an inductor in SPICE is a pure, ideal inductor, while in ADS, an inductor for
example may include also a resistance, a sophisticated temperature model etc., as visible
in the ADS screenshot below:
Therefore, the ADS syntax for such an inductor reads:
L:myInductor node1 node2 L=1n
while it is in Spice simply:
L1 1 2 1n
Getting a list of all ADS components/models:

In order to get an ASCII text file with all ADS model descriptions included, apply the
following script
FOR UNIX:
#!/bin/sh
##################################################################################
# Author: Rene Stoll, Agilent EEsof
# created: Jan.5, 2000
# Description: generate an ASCII file containing the syntax for ADS components
# How2use: copy this script to your operating system and run it
##################################################################################
#
# Set path and environment
# Check if OS is SunOs 4.1.3, Solaris 2.5.1, HPUX9.x or HPUX 10.x
# and set the EESOF_DIR variable accordingly so that the proper code is used.
HPEESOF_DIR=/opt/ads
LM_LICENSE_FILE=/opt/licenses/license.dat
#
export HPEESOF_DIR LM_LICENSE_FILE
PATH=$HPEESOF_DIR/bin:/bin:/usr/bin:$MATLAB/bin:/opt/aCC:/opt/aCC/bin
export PATH
SHLIB_PATH=$HPEESOF_DIR/lib/hpux10:$HPEESOF_DIR/hptolemy/lib.hpux10
export SHLIB_PATH
#execute a 'hpeesofsim -h all to see all components available
hpeesofsim -h all > /tmp/ADS_syntax.txt
For WINDOWS:
REM ##################################################################################
REM Author: Rene Stoll, Agilent EEsof
REM created: Dec.18, 2007
REM Description: generate an ASCII file containing the syntax for ADS components
REM How2use: copy this script to your operating system and run it
REM ##################################################################################
set LM_LICENSE_FILE=C:\agileesof\licenses\license.lic
set AGILEESOF_LICENSE_FILE=%LM_LICENSE_FILE%
set HPEESOF_DIR=C:\agileesof\ADS2006A
PATH=%HPEESOF_DIR%\bin;%HPEESOF_DIR%\adsptolemy\lib;%HPEESOF_DIR%\adsptolemy\lib.win32;%PATH%
hpeesofsim.exe -h all > C:\temp\ADS_ModelsSyntax.txt
See also the IC-CAP ModelFile

demo_features\6_SIMULATORS\2_ADS\5_NETLIST_TEMPLATES\00_ADS_Models_and_Netlist_Syntax.mdl
415
See also the list at
demo_features\6_SIMULATORS\2_ADS\5_NETLIST_TEMPLATES\00_ADS_Models_and_Netlist_Syntax.txt
416
User-defined Model Equations in ADS
Using ADS sdd's (Symbolic defined devices)
For an IC-CAP example, see
demo_features\6_SIMULATORS\2_ADS\5_NETLIST_TEMPLATES\0_USER_DEF_MODELS\1_sdd\1_Diode_sdd.mdl
Example: a diode, but DC model only:
Below the full diode sdd netlist, including charges, as it has to be specified in the IC-CAP
ModelFile Circuit tab:
define diode_SDD_tutor (A C)
;defining some constants for later use in the SDD equations
temp=27
vt=(8.62e-5*(temp + 273.15))
;parameters for custom sdd
IS =100a
NF =1
CJO=200f
VJ =.7
MJ =400m
FC =500m
TT =2p
;standard devices as part of this subcircuit
R:RS A A1 R=.6
L:LS A1 A11 L=1p
C:CAC A C C=1f
; here starts the details of the SDD
vdc =(_v1 - _v2)
;define DC current
id = (IS * (exp(vdc / (NF*vt))-1))
;define diode depletion charge (like most simulators, ADS prefers charges instead of capacitances)
depl_charge(cj,v,mj,vj,fc) = (if (v > fc*vj) then depl_charge2(cj,v,mj,vj) else
depl_charge1(cj,v,mj,vj) endif)
depl_charge1(cj,v,mj,vj) = (cj*vj/(1-mj)*(1-(1-v/vj)^(1-mj)))
depl_charge2(cj,v,mj,vj) = (cj*vj/(1-mj)*(1-(1-FC)^(1-mj))+cj/(1-FC)^mj*(v-FC*vj) \
+.5*cj *mj/vj/(1-FC)^(mj+1)*(v-FC*vj)^2)
;define diode diffusion charge
; NOTE: diff.cap = TT* gm = TT * IS / (N * vt) * exp ( v / (N * vt)),
; but we need the charge for the sdd diff.charge = TT*id(vdc)!
diff_charge(v,is,n,vt,tt) = (tt*is*exp(v/(n*vt)))
charge(cj,v,mj,vj,fc,is,n,vt,tt) = depl_charge(cj,v,mj,vj,fc) + diff_charge(v,is,n,vt,tt)
;finally, calling the SDD feature in ADS, with the properties defined above
SDD:diodeDCnCV A11 0 C 0 I[1,0]=id I[2,0]=-id I[1,1]= charge(CJO,vdc,MJ,VJ,FC,IS,NF,vt,TT)
\
I[2,1]= -charge(CJO,vdc,MJ,VJ,FC,IS,NF,vt,TT)
end diode_SDD_tutor
Using Verilog-A
For an IC-CAP example, see
demo_features\6_SIMULATORS\2_ADS\5_NETLIST_TEMPLATES\0_USER_DEF_MODELS\3_VERILOG_A\diode_VerilogA.mdl
The slide below shows what has to be specified in the IC-CAP ModelFile Circuit tab:
417
Below the contents of file C:\tmp\diode.va
// Make reference to verilog-A discipline 'electrical' (nodes)

`include "disciplines.vams"
module diode_va(anode,cathode);
electrical anode,cathode;
parameter real Is=1e-14 from [0:inf];
parameter real N=1 from [0:10];
parameter real Cjo=0 from [0:inf];
parameter real M=0.5 from [0.1:1];
parameter real Vj=0.7 from [0.1:10];
parameter real Fc=0.5 from [0.4:0.95];
parameter real Tt=1p from [0:inf];
real Vd, Id, Qd;
real f1, f2, f3, fcp;
analog begin
f1 = (Vj/(1 - M))*(1 - pow((1 - Fc), (1 - M)));
f2 = pow((1 - Fc), (1 + M));
f3 = 1 - Fc*(1 + M);
fcp = Fc * Vj;
Vd = V(anode,cathode);
// intrinsic diode.
if (Vd < 0)
Id = - Is;
else
Id = Is * (exp( Vd / (N * $vt)) - 1);
// capacitance (junction and diffusion).
if (Vd <= fcp)
Qd = Tt * Id + Cjo * Vj * (1 - pow((1 - Vd / Vj), (1 - M))) / (1 - M);
else
Qd = Tt * Id + Cjo*( f1 + ( 1 / f2 ) * ( f3 * ( Vd - fcp ) + ( 0.5 * M / Vj ) * ( Vd * Vd -
fcp * fcp)));
I(anode,cathode) <+ Id;
I(anode,cathode) <+ ddt(Qd);
end
endmodule
For an IC-CAP examples see:

demo_features\6_SIMULATORS\2_ADS\5_NETLIST_TEMPLATES\0_USER_DEF_MODELS\3_VERILOG_A\table_models\resistor_tablemodel_VerilogA.mdl
418
419
Netlist Syntax for ADS and SPICE
Comparing Berkeley UCB SPICE and ADS Netlist Syntax
How to have a common Parameter List for Spice and ADS
Berkeley UCB SPICE Simulator Netlists (detailed)
User-defined Model Equations in Berkeley UCB SPICE 3
ADS Simulator Netlist Syntax (detailed)
User-defined Model Equations in ADS
420
CV and S-Parameters
CV Curves form SPICE (iccapmhb)
S Parameter Curves form SPICE (iccapmhb)
421
How to Obtain CV Curves from SPICE
Calculating CV Curves from Complex Currents
IC-CAP file: demo_features\6_SIMULATORS\0_GENERAL\calc_cap_vs_volt_SPICE.mdl
From an inspection of the Simulation Debugger for spice2, IC-CAP surrounds your Circuit
Netlist by the following extra circuitry:
The simulated current result is then interpreted like this:

IC-CAP assumes a situation like with the default LCRZ meter setup (!):
-> a capacitance in parallel with an conductance.
From the simulated current of the schematic above, we obtain the complex admittance
Y = iac / v.
Since v=1, the equation is simply
Y = iac
Therefore, the capacitance value is calculated within IC-CAP from the imaginary part of
the current after
C = IMAG(Y) / (2*PI*1MHz) ----------- (1)
and the conductance is
G = REAL(Y) ------------- (2)
IC-CAP applies formula (1) when 'Mode' is set to 'C', and formula (2) when 'Mode' is set to
'G'.
422
How to Obtain S-Parameter Curves
from SPICE
Calculating CV Curves from Complex Voltages
IC-CAP file: demo_features\6_SIMULATORS\0_GENERAL\calc_Spar_vs_freq_SPICE.mdl
As a fact, most SPICE-like, time-based simulators cannot simulate S-parameters. But they
can simulate the frequency behavior of circuits in magnitude and phase (or real and
imaginary numbers). And IC-CAP can calculate the S-parameters out of these numbers.
How this is done is explained below:
Note: two subsequent simulations are needed in order to obtain one set of S-parameters.
1st simulation: Z1 = Z2 = Z0, e.g.Z0 = 50 Ohm
2nd simulation: Z1 = Z2 = Z0, e.g.Z0 = 50 Ohm
This is the reason why IC-CAP has to insert the user-defined circuit (2-port DUT) as a
subcircuit into a hyper-circuit and to finally send this total circuit to the simulator. In this
manner, the simulator output deck will provide the complex voltages that are required by
IC-CAP to calculate the S-parameters.
To speed-up simulations, IC-CAP puts these two schematics together into one big circuit.
This 'big' circuit is depicted below in fig.3 using SPICE syntax. Its structure is valid for all
other simulator circuit descriptions in IC-CAP, only the syntax may be different. For
MDS/ADS, however, these complex calculations are not required, because both simulators
support S-parameter simulations directly.
The example shows a S-parameter simulation for a bipolar transistor with inputs
VB,VE,VC,VS and frequency. For simplicity, the bias is kept constant and only the
frequency is swept.
see IC-CAP file spar_from_volt.mdl
THE USER_DEFINED SETUP DESCRIPTION LOOKS LIKE THIS:
423
Notes on the names of the circuit elements of the following pages:

The user-defined 'Circuit' description consists simply as shown above of a bipolar
transistor with 4 connections E,B,C and S. Therefore nodes 1 .. 4 (or 1 .. n in a more
general case) as well as 5 ... 8 (or n+1 .. 2n) of the total circuit generated by IC-CAP are
used to link the user-defined circuit into the 'overhead circuit'.
Nodes 9 .. 11 (or generally 2n+1 .. 2n+3) as well as 12 .. 14 (or 2n+4 .. 2n+6) of the IC-
CAP circuit are used to calculate the voltages required for equ. (1) - (4)
(or in other words to 'emulate' the 'S-parameter test set' of the network analyzer).
The simulator node names as defined by the user in the input fields determine the
effective name of the elements in the IC-CAP 'main' circuit, e.g. VCGRO: a voltage from
user-defined node C to GROund or LBGRO: an inductor L from user-defined node B to
GROund.
node numbers valid for data/bjt_npn.mdl file
The values of LxGROx and CxGROx are defined by the IC-CAP system variables
TWOPORT_L and TWOPORT_C, all resistors in TPxCKT that have 50Ohms can be set to a
different value when defining the IC-CAP system variale TWOPORT_Z0.
The corresponding UCB SPICE input file generated by IC-CAP and sent to SPICE:
424
.MODEL npn NPN IS=2.704E-16 BF=86.16 NF=0.979 VAF=86.95

QCKT 1 2 3 4 npn
QCKT2 5 6 7 8 npn
* START SOURCES
.SUBCKT TP1CKT 1 2 10 40 50
CB1 1 30 1e-05
R1 30 10 50
R2 0 40 1
R3 0 50 1
CB2 2 20 1e-05
R4 20 0 50
G1 0 40 30 0 2
G2 0 50 20 0 2
VIN 10 0 AC 1
.ENDS
.SUBCKT TP2CKT 1 2 10 40 50
CB1 30 2 1e-05
R1 10 30 50
R2 40 0 1
R3 50 0 1
R4 20 0 50
CB2 1 20 1e-05
G1 0 40 30 0 2
G2 0 50 20 0 2
VIN 10 0 AC 1
.ENDS
XTP1CKT 2 1 9 10 11 TP1CKT
XTP2CKT 6 5 12 13 14 TP2CKT
LBGRO 2 15 0.0001
VBGRO 15 0 DC 0.91
LBGRO2 6 16 0.0001
VBGRO2 16 0 DC 0.91
LCGRO 1 17 0.0001
VCGRO 17 0 DC 1.5
LCGRO2 5 18 0.0001
VCGRO2 18 0 DC 1.5
VEGRO 3 0 DC 0
VEGRO2 7 0 DC 0
VSGRO 4 0 DC -3
VSGRO2 8 0 DC -3
* END SOURCES
.AC LIN 11 5e+07 3.5e+08
.PRINT AC V(10) V(9) V(11) V(14) V(13) V(12)
.END
A word on 'Noisy S-parameters' for Spice and spice-like simulators:
As was explained above, IC-CAP needs to emulate a network analyzer and its bias-TEE
when requesting complex voltages from spice-like simulators, so that IC-CAP can calculate
the S-parameters by itself.
A special problem might now occur with the limited precision of numbers in the computer
chip. The default value of Twoport_C and Twoport_L is 100 each in IC-CAP. In the
example above, these components became LB_GRO, LC_GRO, and LB_GRO2, LC_GRO2.
Provided the components capacitance is very small, e.g. some 10fF, such a processor
rounding effect can happen and as a result, the simulated S-parameters may look like
'noisy', as depicted below. This usually happens in the SPICE simulator solver matrix,
where the big values (100F, 100H) together with these small DUT capacitance values lead
to numerical rounding effects. To prevent this, specify the two System Variables (under
IC-CAP/Main) and set their values to lower ones, e.g. 100m each.
However, do not set them to too small values (!) . Both, TWOPORT_L and TWOPORT_C
must be big enough, so that their influence (phase shift due to the resonance of
TWOPORT_L and TWOPORT_C) is over for the lowest simulation frequency. This may be
critical for low-frequency NWAs who's lowest frequency is in the 30kHz range.
425
Keep in mind that adding bias-TEEs means a resonance at a few 10 Hz (the L and C of the
bias TEE), and that the phase shift of this resonance must be over at the lowes
measured/simulated NWA frequency. In other words, what you measure with a NWA and
bias TEEs (internal or external ones), represents the S-parameter trace of your DUT after
a complete 360' phase shift.
426
Harmonic Balance
ADS
ADS, as an example for a powerful RF simulator, exhibits a group of simulators, which are
called automatically according to the simulation task
The total tool chain of ADS plus the IC-CAP part is given below:
The ADS simulator tool chain and the positioning of IC-CAP
Comparing to conventional SPICE-like simulators, ADS' high-frequency SPICE offers these

features:
uses high-frequency models for microstrip lines, bends, gaps and other
discontinuities
used to analyze steady-state response of mixers, oscillators, amplifiers etc.
frequency dependent components are modeled with approximations that neglect
some of the frequency dependent effects, such as dispersion and high frequency loss.
This results in faster simulation results, which are nevertheless accurate enough for
electrically small components on an IC.
requires the linear simulator license
The Convolution simulator offers:
includes dispersion effects and high frequency loss

results in accurate high frequency results at the expense of longer simulation time
handles circuits containing distributed elements and S-parameter data used for
components
can accurately analyze circuit start-up and transient conditions at low and high
frequencies, where the effects of dispersion and discontinuities are significant (for
electrically big components on printed circuit boards
The AC S-parameter simulator allows linear S-parameter simulations. Being part of a RF

simulator, enhanced RF models are available.
What is Harmonic Balance (HB) Simulation?

Harmonic balance (HB) is a frequency-domain simulation technique. Signals are treated as
DC plus summations of finite numbers of sine waves, i.e. as periodic signals. Harmonic
balance simulations use one or two sine-wave stimuli to determine (for example)
conversion loss vs. drive level, distortion vs. frequency, and gain compression.
Harmonic Balance and Time-Domain Simulation

Harmonic balance calculates voltages and currents in much the same way as time-domain
simulators such as SPICE do. Frequency-wise, it solves for the magnitudes and phases of
all spectral lines simultaneously. In effect, the entire waveform is being solved for at once.
Harmonic balance is extremely efficient if the signals are simple in the frequency domain-
as they are, for example, when an amplifier is driven by a sine wave.
Circuits that have non-repeating transient behavior are best analyzed using ADS-Impulse,
a time-domain (transient) simulator. In general, harmonic balance is not very useful for
analyzing signals that might be analyzed using an arbitrary waveform generator or a
wide- bandwidth oscilloscope. ADS-Impulse is also useful for verifying the dynamic
performance of a circuit after it has been designed and optimized using harmonic balance.
What happens when a nonlinear harmonic balance

simulation is executed?
427
For the circuit given above, the following iterations are performed:
1. First, a DC simulation is executed to obtain the DC operating point of the circuit.

2. A linear AC analysis (Harmonic Balance Fourier Analysis) is applied to analyze the
linear and passive components in the frequency domain, stimulated by the
3. Nonlinear device models are usually defined as i=f(v). Therefore, the actual voltage
frequency spectrum across the nonlinear component is Fourier-transformed into the
time domain. This resulting time domain signal is applied to the nonlinear device with
a Spice-type simulation. The resulting time domain current is Fourier-converted back
to the frequency domain.
4. The sum of the actual current spectra at each circuit node is calculated and it is
checked, if this sum is below a certain tolerance limit for all frequencies.
5. Steps 1-4 are iterated until Kirchhoff's current law is satisfied for all frequencies at all
nodes. The same is true for fulfilling Kirchhoff's voltage law. Iterative, matrix-driven
techniques are used (Newton-Raphson algorithms etc.).
6. The obtained solution is a satisfaction of Kirchhoff's current law (sum of all currents,
for all frequencies, is zero for each node) as well as the voltage law. The
mathematical analysis has converged. The harmonics are now balanced.
NOTE: Changing the value of an individual component changes the current and the harmonic
content of the signal through the other components. It also changes the output voltage. But it does
not affect the sum of the currents. This sum is always zero.
In the Agilent Advanced Design System (ADS), harmonic balance simulations assume that
all voltage and current waveforms in the circuit are summations of sinusoidal signals. In
other words, finite spectrums composed of individual spectral lines.
Energy is not allowed to exist continuously throughout a spectrum. It is discretized into

ideal spectral lines that are the harmonics of more basic fundamental frequencies and
intermodulation (sum and difference) products of those fundamental frequencies.
This section offers a brief introduction to the parameters that affect the frequencies in the
analysis: FUND, ORDER, and MAXORDER.
Signal Sources
In harmonic balance simulations, the voltage, current, and power sources in the circuit
specify the amplitude of the signal, not the frequency (at least not directly).
In harmonic balance simulation, frequency information is specified in the simulation setup,
as part of specifying the number of fundamental frequencies that are being used and their
values.
Signal sources for harmonic balance simulation have a parameter FUND that relates the
harmonic balance simulation information to the signal source.
FUND is a small integer, and is likely to be 1 or 2.
428
If FUND is set equal to 1 on a VLS large-signal source, the source produces a sine wave
whose frequency is the same as the first fundamental frequency of the harmonic balance
simulation component (for example, 1 GHz).
Fundamental Frequencies, or Tones

Spectral lines in harmonic balance simulation results are almost always linear
combinations of one or more fundamental frequencies:
FREQ of one spectral line = (m)*FREQUENCY1 + (n)*FREQUENCY2 + ...
The number of frequencies is important mostly because each additional fundamental

frequency means a significant increase in memory usage, simulation time, and dataset
storage size. Choose the smallest number that still gives the correct results.
In ADS, fundamental frequencies are also known as tones. Thus, in a one-tone harmonic
balance simulation, only one fundamental frequency (FREQ), multiples of that frequency
(its harmonics), and DC (the harmonic of order 0) are used.
One-tone harmonic balance is the most common simulation type for amplifiers.
ORDER
In the simulation setup, the parameter ORDER specifies for each fundamental frequency
the highest harmonic (largest multiple) of that fundamental frequency that will be
considered by the simulator.
The Harmonic Balance Variable MAXORDER

While the parameter ORDER affects the number of harmonics for each fundamental, the
parameter MAXORDER controls the number of intermodulation (mixing) products that will
be computed in a two-tone or multi-tone simulation.
For example, if MAXORDER is set to 3, only first-order, second-order, and third-order
mixing products will be computed.
Dimensionality of Results
In presenting the results of harmonic balance simulations, keep in mind the fact that the
node voltages and currents are complex spectrums, not single numbers.
With no swept variables, a simple harmonic balance simulation produces a result with
a frequency axis. Therefore, a sweep of (n) variables on the circuit page results in
(n + l)-dimensional data.
Nonlinear High Frequency Characteristics

There are 4 main nonlinear high frequency characteristics
Third Order Intersect (IP3)

1-dB Gain Compression (1 DBC)
Saturated Power (Psat)
Gain Compression at Saturation (GCS), which are explained with the sketches below:
1dB compression point, a measure for the 1st order harmonic distortion
429
Definition of the 3rd order intersect, IP3, a measure for 3rd order harmonic distortion
Special Discussion of the Third-Order intersect analysis techniques

The third-order intersect of an element or network is a widely accepted system design
parameter because it indicates the degree of nonlinearity of a nonlinear element. The
output to input voltage relationship for an electrical component can typically be described
as a polynomial relationship like:
where
a0 output dc offset
al small signal gain
an higher order coefficients; n > 1
Fig.2, above, shows a typical nonlinear element output power versus input power curve
for the fundamental and third-order output tones for a single tone input test signal. In the
linear operating range of the amplifier, the small signal fundamental curve varies with a
1:1 slope. The small signal third-order curve varies with a 3:1 slope. The third-order
intersect is that point where the extrapolated small signal fundamental and third-order
curves intersect. At this third-order intersect one may be interested in the input power
level, or in the output power level.
These curves are of main interest when performing a distortion analysis on a circuit. Here,
430
we can distinguish between a 1 tone and a 2 tone signal input. While the 1 tone result
looks like the one given in figures 1 and 2, fig.3 shows the result for a 2 tone distortion
analysis.
Distortion analysis for a two-tone input
Harmonic Balance Simulations for Modeling

The importance of harmonic balance simulations and its application for modeling can best
be seen when considering the measurement principle of a vector network analyzer.
The measurement principle is assuming a linear behavior of the DUT. However, this is only
true for passive components. For an active DUT, like a transistor, this is not quite true.
Even for very low RF signal levels, i.e. in the -20dBm level, harmonics occur. This is
sketched below:
431
This nonlinear frequency behavior particularly becomes important when considering

today's communication systems, where pulsed signals are used rather than CW sinusoidal
signals like in the past.
Therefore, a nonlinear vector network analyzer should be used for characterization instead
of the conventional S-parameter VNA measurements.
However, such nonlinear network analyzers are very expensive and sophisticated to
handle. In many cases, modeling engineers therefore replace such systems by a
conventional small-signal VNA measurements (-30dBm RF signal source power) plus
additional measurements using a spectrum analyzers. The later, as a disadvantage, allow
only to measure the magnitude of the harmonics, and not also the phase. In this case, the
true time domain signal cannot be reconstructed out of the harmonics any more, what is
possible when using a nonlinear network analyzer.
IMPACT OF HARMONIC BALANCE SIMULATIONS FOR CONVENTIONAL S-

PARAMETER MODELING
432
When using a conventional linear vector network analyzer, only the base frequency is
measured, and the frequency harmonics are ignored, due to the very small bandwidth of
such a VNA. This means, that when harmonics occur, they are not part of the measured
S-parameters and therefore can lead to a modeling problem which has not much to do
with the real device under test.
Taking this into account, it has been observed, that the transit frequency modeling of
transistors can be performed much more accurately when taking this measurement
situation into account. This means that no S-parameter SPICE simulations should be
applied for curve fitting, but rather a harmonic balance simulation, from which only the
base frequency is considered and compared to the measured VNA S-parameters.
Furthermore, the harmonic balance DC operating point conditions can be compared in the
same step with the measured DC Collector current or Drain current of the transistor.
With these HB simulations applied to S-parameter measurements, the real measurement
conditions and their limitations are reflected, and therefore, accurate modeling results for
the ft modeling can be achieved.
It must be mentioned, that for such an application, the harmonic balance simulations has
to be applied to a frequency sweep which is identical to the VNA measurement frequency
steps.
Again, a SPICE simulation is a linearization and does not take the real existing non-
linearities into account. It assumes no harmonics, and can therefore lead to wrong
simulation results for ft modeling.
Final remark on using HB simulations for S-parameter modeling:

in this case, no OPEN or OPEN-SHORT de-embedding can be applied, because these
twoport operations assume a linear circuit behavior. When applying HB simulations, the
OPEN and the SHORT have to be modeled and added to the transistor circuit. Only in this
case, the HB simulation will give a correct result.
After the measured S-parameter curves have been modeled using HB simulations, in a
next step, the harmonics spectrum is modeled. Based on the mentioned spectrum
analyzer measurements, a small fine-tuning of the RF model parameters will help to fit the
harmonics vs. RF input power measurements.
433
With such a complete modeling, from DC, via CV curves, to S-parameters and
corresponding HB simulations to finally fitting the harmonics versus RF power, the
transistor is modeled as accurately as possible. Its parameter set can now be considered
as fulfilling the requirements of its later application, i.e. its operation in the mW RF signal
range (~0dBm).
Last not least, the following two sketches depict nonlinearities showing up when a
sinusoidal signal or a two-tone sine function is applied to an amplifier, or a single
transistor:
434
History
History of Circuit Simulators
The origins of Berkeley SPICE
SPICE (Simulation Program with Integrated Circuit Emphasis) was introduced in May 1972
by Prof. D. O. Pederson at the University of Berkeley, California.
It simulates the electrical behavior of circuits based on discrete components

In particular, these include:
resistors, capacitors, inductors and coupled inductors

ideal lines and cables
independent and controlled current and voltage sources
semiconductor components
It soon became a standard for electronic circuit simulation, used by both, the industry and
the universities. This fact is based on the technical expertise of the developers of SPICE:
especially Prof. Pederson /1/, and Prof. Nagel /2/.
After 17 revisions of SPICE, SPICE2 was introduced in 1975. The main improvements were
the dynamical RAM allocation capabilities and the automatic time stepping for time-
domain simulations. Especially SPICE version 2g6, was a specially reliable program
version, written in Fortran. This version has become the root of many industry in-house
simulators and also of many commercially available simulators. This explains why they
typically understand the Berkeley syntax (Spectre, Hspice, Pspice, Eldo).
Finally, SPICE3, introduced by the end of the 80s, was the first version written in the C
language.
One of the reasons for the success of SPICE is also the implementation of good transistor
models. This is especially true for the bipolar model. The Gummel-Poon model has only
become popular after it was implemented in SPICE. (However, some simplifications had
been made to the original paper of Gummel-Poon.) The same is true for MOS transistor,
however, after MOS3, and the introduction of the BSIM series of model, SPICE's model
quality had suffered a lot compared to MOS2 and MOS3. Finally, with BSIM3v3, Berkeley
SPICE gained back its authority of model definition.
1. Nagel, L.W.; Pederson, D.O.: SPICE. Berkeley, Univ. of California, Electronic

Research Laboratory, ERL-M 382, 1973
2. Nagel, L.W.: SPICE2: A Computer Program to Simulate Semiconductor Circuits.
Berkeley, Univ. of California, Electronic Research Laboratory. ERL-M520, 1975
The origins of ADS

The origin of MDS simulator goes back to the late 1960's when HP introduced the
simulator OPNODE which was a linear s-parameter simulator. In 1984 HP wrote a linear
simulator called Time-Domain Tutorial which could be used to take a transmission line
topology, compute the S-parameters, and use the 8510 to get a time domain
reflectometer plot. This time domain tutorial program eventually evolved into the MDS
Microwave Linear Simulator MLS in MDS A.01.00.
Ken Kundert worked on MLS until 1985 when HP funded him to get his Ph. D at U.C
Berkeley working on nonlinear microwave simulation based upon harmonic balance
technology. HP was co-developing harmonic balance in conjunction with Ken Kundert and
Berkeley. The internal name for this simulator was "Spectra" which was later named the
HP Microwave Nonlinear Simulator or MNS. MNS essentially evolved from Berkeley's code.
The Berkeley code was named Spectre ( a subtle difference ). After Ken Kundert received
his doctorate, he left HP to work at Cadence where he built a new time domain simulator
based upon the UC Berkeley Spectre code, hence the Cadence Spectre simulator was
born.
In 1996, Hewlett-Packard bought EEsof, and MDS and the EEsof simulator products were
merged. The new product was called ADS (Advanced Design System).
435
Device Modeling
Contents
Diode (iccapmhb) On the Web

Transistors (iccapmhb)
RF Passive Linear Components (iccapmhb) Articles
Low Frequency Noise (iccapmhb) Knowledge Center
Thermal Modeling (iccapmhb) Technical Support
On Target Modeling (iccapmhb) Training
Scaled Modeling (iccapmhb) Videos
436
Diode
Contents
Modeling a Diode (iccapmhb)

ADS Junction Diode Model (iccapmhb)
Step Recovery Diode (iccapmhb)
437
Modeling a Diode
Introduction
Referring to the Modeling Handbook's chapter on curve fitting, regression analysis was
introduced as a method for linear curve fitting. It was a pretty simple and straight-forward
method. We had to solve the partially differentiated equations for the parameters m and b
of the equation
y = m*x + b
It was mentioned that fitting more complex measurement curves leads to much more
complex problems for solving the set of equations for the parameters. And pretty often,
this set of non-linear equations cannot be solved without numerical methods.
But there is a way around:

A simple way to fit non-linear curves to measured non-linear data is to transform them to
a linear world. But the question is, how to do it. A look at the 'target function' of the
model equation gives a hint. An example: in case of an exponential function like
,
the transform needed is a simple logarithmic conversion. This automatically gives the non-
linear transformation for the measured data. Once the measured data are transformed to
this linear range, a linear regression analysis is applied. And so we yield the slope and y-
intersect of the fitted line. The model parameters are finally calculated out of these two
values.
This will become much more transparent in the following diode modeling example.
The SPICE equivalent schematic for a diode is shown in figure 1. It consists of the ideal
diode D representing its non-linear DC characteristic plus two voltage dependent
capacitors for taking care of the space charge (CS) and delay effects (CD) as well as a
series resistor RS for the high-current effects. The series inductor (bonding effect ...) is
neglected as well as a parasitic capacitor (housing effects ...).
438
DC Characterization Parameters IS, N AND RS

Neglecting high current effects, i.e. RS=0 or vD = vD internal, and also neglecting
recombination effects for low biasing voltage, the diode current in the forward active
region is modeled using:
Determination of the DC parameters IS and N

Provided vD >0, i.e. neglecting the term ( -1) in (1),
and applying a logarithmic conversion yields:
vD
log(iD) = log(Is) + ------ log(e)
N VT
= log(Is) + [1 / (2,3 N VT)] vD (2a)
This is an equation of the form:
y = b + mx (2b)
In order to interpret (2b) linearly, we have to substitute:
y = log(iD) (2c)
b = log(Is) (2d)
439
m = [1 / (2,3 N VT)] (2e)
x = vD (2f)
This explains how to manipulate the measured data: after the logarithmic conversion of
the measured values of iD (2c), they are introduced with the still linear values of vD (2f)
into the regression equations (10) and (11) of the previous regression analysis chapter
4.1 as yi- and xi-values. We obtain the y-intersect b and the slope m of the linear
regression function. Solving (2d) for Is and (2e) for N we finally are able to calculate these
two parameters out of b and m as follows:
Is = 10 b (3)
and
N = 1 / (2,3 m VT)
with VT from (1a) (4)
Validity of this extraction:

The parameter extraction described above is valid only in that range of measured data,
where the assumptions are true. This means: eq.(3) and (4) are valid for vD > 0 (data
above the measurement resolution (non-noisy data), typ. >0,2 V). The diode current
should not be dominated by recombination effects (the weaker slope at low bias voltages)
but also below high-current effects (no ohmic effects, the famous knee in the half-
logarithmic diode characteristic above typically 0.8 V)
Parameter RS
After the parameters Is and N are extracted, the value of RS can be found from the two
highest bias points (biasmax_index and biasmax_index-1 of a sweep from vD_min to
vD_max) as follows:
vD[max_index] - vD[max_index-1]
RS = --------------------------------- (5)
iD[max_index] - iD[max_index-1]
This method gives only accurate results for RS if the applied vD was high enough so that
the ohmic effect (of RS) dominates over the nonlinear, exponentional diode curve iD =
IS*exp(vD/(N*vt) -1)
Another method to determine the ohmic part of a diode characteristic, independent of that
limitation from above, is to consider the voltage drop along RS, between the ideal diode
characteristics and the applied outer voltage. Since we know the applied iD, and since we
can calculate the voltage drop across the inner diode (IS and N are known ! ), we can
calculate RS.
This is done by firstly determining the maximum current iD from the sweep by
i_RS=ia.m[max_index]
This is the current flowing through RS.

Then, we calculate the voltage drop across RS following
v_RS = measured_voltage - ideal_diode_voltage
or
v_RS = va.m[max_index] - vt*N*ln(i_RS / IS)
what finally gives
RS = v_RS / i_RS
Of course, as mentioned, the diode DC parameters IS and N have to be determined first.
Pre-requisite for a good RS extraction

the ohmic effect dominates the diode characteristics. Referring to fig.2, the decline for
high bias voltage must be clearly visible in the extraction range.
CV Characterization
The frequency behavior of a semiconductor can be modeled by a space charge capacitance
(dominant at reverse biasing) and a diffusion capacitance (dominant at forward bias) that
models the time delay effects. The first capacitance is typically measured with a negative
bias using a CV meter (capacitance versus voltage) while the latter is commonly measured
440
using a network analyzer.
This chapter covers the modeling of the space charge capacitor. Another method is using
a network analyzer, measuring the s11 curve with a negative bias. This is not covered
here.
Space charge capacitance in general, extracting parameters Cj0, Vj,

m
The behavior of the space charge capacitor is given by:
For more details see /Antognetti/.
Determination of the CV parameters
We only use the negative bias region for parameter extraction. The logarithmic conversion
of (6a) gives:
This equation can be interpreted again as a linear function:
y = b + m * x
with
y = ln(CS) (8a)
b = ln(CJO) (8b)
m = - M (8c)
and
x = ln[1 - vD / VJ ] (8d)
How to proceed
The measured values of CS are converted logarithmically according to (8a). Then,

following (8d), the stimulating data of the voltage sweep vD are converted too. Since the
parameter VJ has a physical meaning, its value should be in the range of 0,2 .. 1V.
Therefore we select 0,2V as a starting value for VJ. These two arrays are now introduced
into equations (10) and (11) of the chapter on regression analysis as yi- resp. xi-values. A
linear curve is fitted to this transformed 'cloud' of stimulating and measured data and we
get the y-intersect b and the slope m for the actual value of VJ. These two values are the
best choice for the given VJ. In the next step, this procedure is repeated with an
incremented VJ, and we get another pair of m(VJ) and b(VJ) . But now the regression
coefficient r2 will be different from the earlier one: depending on the actual value of VJ,
the regression line fits the transformed data 'cloud' better or worse. Once the best
regression coefficient is found, the iteration loop is stopped and we get VJ_opt as well as
441
the corresponding b(VJ_opt) and m(VJ_opt).
The final parameter values are then from (8c):
M = - m(VJ_opt) (9a)
and from (8b):
CJO = exp [ b(VJ_opt) ] (9b)
Validity of this extraction
The parameter extraction for the space charge capacitor is valid only for stimulus voltages
below FC * VJ , FC_default = 0,5 .
In practice, there is always an overlay of this capacitance with some parasitic ones, e.g.
packaging or bond pads. If they are not known and therefore cannot be de-embedded
(eliminated from the measured data by calculations), the three modeling parameters may
have values that have no physical meaning. This is especially true for VJ and M.
Nevertheless the fitting of the proposed method is generally very good and pretty easy.
In order to also determine the parasitic offset capacitance, see the examples in the
chapter about 'regression analysis applications'.
HF Modeling: Parameter TT
The small signal equivalent schematic of the diode for high frequencies is given in fig.4.
When comparing it to fig.1, it can be seen that the element D, representing the non-linear
DC transfer curve of the diode has been replaced by the linearized small signal
conductance gD.
Let's start with gD , the slope to the DC diode characteristics at the operating bias point.
The diffusion capacitance in the operating point is given by /Antognetti/:
CD is typically overlying the space charge capacitance.

When performing a 2-port measurement of a diode with a network analyzer, we can
calculate the total diode capacitance by converting the S-parameters to Y-parameters and
calculating
Cdiode= IMAG(-Y12) / (2 PI freq) .
The resulting CV curve is depicted below in fig.5:
When applying a NWA, compared to CV meter measurements, there are no restrictions

related to positive diode DC biasing and measurement instrument restrictions. Therefore,
442
we can easily forward bias the diode and study the transition from the space charge
capacitance to the diffusion capacitance. This gives the diffusion capacitance.
NOTE: in practice, especially for packaged devices, the diffusion capacitance is overlaid by the package
inductor! See further below!
Determining TT
As can be seen in fig. 6, CD can also be optimized in the S-parameters for medium DC
biases, below the influence of RS. In other words, related to S-parameters, TT shifts the
Sxx and Sxy traces (adds phase). The effect is dominant for medium and higher DC biases
below take-over of RS.
NOTE:
When RS begins dominating the diode DC trace, think of the 'inner' diode as a resistor with 1/gD in series
with a voltage source of e.g. 0,7V. Therefore, the capacitances CS and CD are shortened by this
decreasing diode resistor, and therefore, TT is shortened by this resistor!!
Finally, when taking also the series inductance into account, which is a typical first-order
model of the diode package, we get S-parameters like shown in fig.7:
The series inductor overlays the so far discussed S-parameters of the inner diode. It
basically adds phase to the inner diode S-parameters, and for high DC biasing (where RS
domiinates), the inductance affects the diode S21 trace considerable: S21 now turns
downwards, tending towards S21->0 for infinite frequency.
Related to modeling, LS can be obtained from optimization of high DC biases.
Model Limitations
In order to keep the models simple and usable and to have reasonable simulation times,
they might suffer from some limitations:
DC: diodes may show recombination effects at low forward bias voltages. This
shows-up as a lower slope on a half-logarithmic scale. In order to cover this effect,
the diode model is replaced by a subcircuit, consisting of a diode for the
recombination effect, another one in parallel for the upper voltage area and a
resistance in series with both diodes. (Both diodes have RS=1e-6 Ohm).
CV: any parasitic capacitance is not included in the diode model. Using again a sub-
circuit, a 2nd parasitic capacitance can easily be added.
RF: the parasitic series inductor is not included. Again, a subcircuit could be used for
modeling.
Details of the Berkeley SPICE Diode Modeling for

Experts
This section gives detailed information about the Berkeley Spice model including thermal
and noise modeling.
443
The complete SPICE parameter list
parameter description units default example

RS ohmic resistance Ohm 0 10
IS saturation current A 1.0E-14 1.0E-14
N emission coefficient 1 1
BV reverse breakdown voltage V infinite 40
IBV current at breakdown voltage A 1.E-3
CJO zero bias junction capacitance F 0 2.0E-12
VJ junction potential V 1 0.6
M grading coefficient 0.5 0.5
FC coeff.for forward-bias deplet.cap. 0.5 0.5
TT transit time sec 0 1E-10
EG activation energy eV 1.11 1.11 for Si
0.67 for Ge
XTI saturation current temp.exp. 3.0 3
KF flicker noise coefficient 0
AF flicker noise exponent 1
TNOM parameter measurement temp. 'C 27
GMIN min.SPICE conductance 1E-12
(a SPICE convergence parameter)
SPICE Diode Model Equations

DC model
forward and reverse leakage:
reverse breakthru current:
Note: when implementing the diode model into simulators, consider a linear continuation of the
exponential function, in order to improve the simulation convergence and to avoid numeric overflow
AC model
Junction capacitance
For user-defined models, charges have to be specified, .

Therefore the junction charge is:
444
for vd<FC*VJ :
and for vd>FC*VJ :
Diffusion capacitance:
and the corresponding charge (for model implementation):
Thermal model
Noise model (used only in AC analysis)
For an implementation example, see

demo_features\6_simulators\2_ADS\3_user_def_mdls\1_sdd\basic_sdds\1_diode_sdd.mdl
Publications
Diode modeling and modeling in general:
P.Antognetti, G.Massobrio, Semiconductor Device Modeling with SPICE,
McGraw-Hill, 1988, ISBN 0-07-002107-4
445
ADS Junction Diode Model
INSTANCE
ModelName [:Name] anode cathode <parameter=value> ... ; (device)
Parameters
Area Junction area factor.
Region DC operating region, 0=off, 1=on.
Temp (C) Device operating temperature.
Gd (Siemens) Small signal conductance.
Cd (F) Small signal capacitance.
Mode Nonlinear spectral model on/off.
Noise Noise generation on/off.
MODEL CARD
model ModelName Diode <Model Parameter=value> ...
model Parameters
Isr (A) recombination current parameter.
Nr Emission coefficient of recombination current.
Is (A) Saturation current.
Js (A) Saturation current.
N Emission coefficient.
Ikf (A) high injection knee current.
Rs (Ohms) Series resistance.
Fc Forward-bias depletion capacitance threshold.
Ibvl (A) low-level reverse breakdow knee current.
Nbvl low-level reverse breakdow ideality factor.
Bv (V) Reverse breakdown voltage.
Ibv (A) Current at breakdown voltage.
Nbv reverse breakdown ideality factor.
Cjo (F) Zero-bias junction capacitance.
Vj (V) Junction potential.
Pb (V) Junction potential.
M Grading coefficient.
Tt (s) Transit time.
Ffe flicker noise frequency exponent.
Eg (eV) Band gap.
Kf Flicker-noise coefficient.
Af Flicker-noise exponent.
Tnom (C) Parameter measurement temperature.
Xti Saturation current temperature exponent.
Pt Saturation current temperature exponent.
Imax (A) Explosion current.
wBv (V) Diode reverse breakdown voltage (warning).
wPmax (W) Maximum power dissipation (warning).
446
Step Recovery Diode
Note
This chapter has been contributed by Prof. Martin Sauter, University Bundeswehr, Munich, Germany and
Klaus-Willi Pieper, Infineon, Munich, Germany.
The problem of a diode, "reverse recovery is a common known problem resulting from
many circuits, especially in power electronics. It usually occurs, when a diode switches its
state from the forward conducting mode to reverse blocking mode. In this case, the stored
minority charge in the diode regions have to be removed from the diode anode and
cathode regions, which does not occur instantaneously. The whole process may from a
physical point of view be understood as the discharging of the diode diffusion capacitance.
An electrical equivalent circuit and the resulting waveforms are drawn in the Figure above.
After the voltage source changes its state from forward voltage UF to reverse voltage UR,
a discharge current is flowing, which is mainly limited by the series resistor Rs . This
current lasts as long as the diode is in the conducting state and the voltage drop at the
junction is about 0.7V. This switching phase is usually caused the storage phase.
447
After this time, a more or less exponential decay of the current is observed, although the
diode does not anymore operate in the forward region. This phase is usually called the
recovery phase. The recovery time now is the time between the switching point of the
voltage source and the time until the current has dropped to 10% of its maximum value.
A good comprehensive review of this effect (although in German) is found in the

Reference 2.
The physics of the reverse recovery process can in general be described by an

examination of the stored minority charge in the diode regions.
From quasistatic point of view, it can be assumed that the minority carrier charge is
related to the current by a constant diffusion time, which is in literature called the diode
transit time TT. The total diode current is given by the expression:
qT describes the stored minority charge in the diode regions and the variable iD is the
current which flows under DC conditions when a voltage V is applied to the diode.
If we regard a time dependent voltage u(t), usually a quasistatic assumption is made,

which states that the stored minority carrier charge is given by the equation (I). This
however is only valid when the stored charge is able to follow the rapidly changing voltage
u(t) applied at the diode terminals. A qualitative description of the resulting charge
distribution in the diode regions is given in the picture above.
Left: A quasistatic assumption is made, which assumes the minority carriers to follow the
diode bias without delay.
Right: A non-quasistatic assumption is made, which shows that the carriers near the
junction follow the bias faster than the carriers far away from the junction. The minorities
first vanish in the region which is close to the junction. In a later stage of the process, a
number of minorities is stored in the diode regions far away from the junction.
448
Here we describe how the measurements of the recovery time have been performed. We
have used a setup proposed by AVTECH Inc. using a double output pulse generator
AVTECH EB4B. This unit has two low-impedance (~2 Ohm) outputs. The output OUT1
generates a positive pulse, which is sent to the device passing the resistor of 47 Ohm. The
width T1 of this pulse can be varied between 100ns and 1us. The pulse rise time is about
1ns.
The output OUT2 generates a negative blocking pulse with a width T2 between 100 and
300ns, which is coupled to the device by a capacitor with typical values of several nF. The
time delay delta_T between the two pulses can be varied as well as the voltages of the
two pulses.
A fast Oscilloscope is placed at the output of the device. The input voltage of the
Oscilloscope can be used to calculate the current through the device. The frequency
characteristics of the oscilloscope had been incorporated into the simulation.
Attenuators have to be a added especially to the oscilloscope to measure voltages above 5
V rms.
The samples were prepared into a small CDSO-20 package.

The whole circuit was designed on a system board, which can be seen on the next page.
Data acquisition has been done by the used Tektronix oscilloscope, the ascii data have
been transferred afterwards to the UNIX system with ICCAP running.
Standard Diode Model

In the standard diode model, a rather simple equation is implemented. It calculates the
stored charge in the diode during forward conduction by one parameter (usually denoted
as transit time):
The total current in the diode is then calculated by:
449
Here, the junction capacitances are neglected for the reason of simplicity.
If we simulate the diode switching characteristics by this model, we obtain a time

dependent current which is rather Step-like with an abrupt change of its value from the
recovery peak to zero. The only possibility to change the waveform in the recovery range
offered by the standard diode model is by varying the junction capacitance. But this
parameter cannot be used as a fit parameter for the recovery waveform.
The following simulation shows the resulting transient waveforms for a rising parameter
Non Quasi-static Diode Model

An alternative approach for a model of this recovery phase was developed by Lauritzen
and Ma 1. Originally, it was developed for low-doped high-voltage pin-Diodes.
This model separates the charges in the diode region into two parts Qe and QM .
In the equations ID denotes the static (DC) current of the diode. Itot is the sum of the
static current and the current caused be the recovery.
Parameter Extraction
Let us have a look at the resulting transient waveform, if we perform variations of the
model parameters TM and Tau.
First we assume TM ~ Tau and perform a variation of Tau . The results look almost
identical with the waveforms obtained by the standard model, which is consistent with the
mathematical considerations above, where we derived the standard model as a special
case of the NQS model.
450
Variation of TM leads to a much more realistic description of the recovery behaviour by

adding an exponential tail to the step characteristic.
Model Implementation
If we take a look at the original equations defined by the Reference 1 , we can perform
the following operations with the second equation:
If we apply a Laplace Transform on this equation, we get:
The term in the square bracket on the left side now is equivalent to the transfer function
of a simple RC-circuit if the values of R and C fulfil the condition:
451
This gives an idea of how to implement the differential equations given by Lauritzen into a
PSPICE circuit by only using standard elements.
The diode current is measured by a current controlled voltage source. If the

transresistance of this element equals the value of Tau, the output voltage is equivalent to
the value of Qe
1. The following RC circuit performs the calculation of the transfer function mentioned
above. Here, the value of the resistor is set to 1, and the value of the capacitor is
given by the two time constants.
2. The third element, which is a voltage controlled current source, performs the
calculation of the additional current given in the third equation. The transconductance
of this element is equal to Tau^-1
As the resulting current source is in parallel to the original diode, this is an additional
contribution given by the equations.
The resulting schematic, which uses PSPICE ABM (Analog behavioural model) elements, is
drawn below.
Modelling Results
We apply our model and the parameter extraction strategy to a reverse polarity protection
diode, which is part of the Infineon Technologies SPT5 BCD-Process. This diode has a
breakdown voltage of about 60V.
This diode together with the substrate forms a low-gain pnp transistor, where the
substrate acts as collector and the diode acts as the emitter-base junction. The diode also
includes another doping region on the top, decreasing the number of stored minority
carriers in the anode.
For a proper modelling of the recovery effect, the circuit we showed before has to be
modified as drawn below. This has the advantage that the extracted substrate parasitics
can be modelled by the pnp transistor and the network, which models reverse recovery
behaviour has to be added later.
452
The results of the modelling of the dynamic characteristics are shown on the next page.
This picture shows the reverse recovery characteristic of the modelled diode. On the x-
axis, the time (in ns) is shown, on the y-axis the calculated current (Oscilloscope voltage
divided by 50 Ohm) is plotted.
The injection level has been chosen about 25mA on the diode, the reverse voltage was 5
and 10 V, respectively.
This picture shows the reverse recovery characteristic of the modelled diode with a higher
injection level (~100mA) of the forward current than in the picture on the page before.
Reverse voltage was also higher (15 V and 20V, respectively).
On the x-axis, time (in ns) is shown, on the y-axis the calculated current (Oscilloscope
voltage divided by 50 Ohm) is plotted.
Summary
The aim of this paper is to show how the well known model of Lauritzen and Ma can be
implemented very easy to model an integrated, very fast switching diode.
453
The way of implementation of the Lauritzen model is modular and can be added to
standard Diode and BJT models as an additional network using only controlled sources and
an RC-Circuit
The model is shown to predict the switching time of the measured diodes very well, the
simulation was verified at different levels of injection current and for different reverse
blocking voltage.
References
1. Lauritzen, P.O.; Ma, C.L.; A simple diode model with reverse recovery, IEEE
Transactions on Power Electronics, Volume 6, Issue 2, April 1991 Page(s):188 - 191
2. M. Reisch, Elektronische Bauelemente, Springer Verlag 1998, p. 378-382
3. AVTECH Inc., Manual of Pulse Generator Unit EB4-B
454
Transistors
Bipolar in General (iccapmhb)
BSIM (iccapmhb)
GaAs Transistor Models in General (iccapmhb)
General Transistor Modeling Strategies (iccapmhb)
Hicum Level 2 model (iccapmhb)
Learning the Angelov Model (iccapmhb)
MOS3 (iccapmhb)
MOS Transistors for Dummies (iccapmhb)
PSP (iccapmhb)
The Curtice Mesfet Model (iccapmhb)
The Gummel-Poon model (iccapmhb)
YParameter Modeling (iccapmhb)
VBIC (iccapmhb)
455
Bipolar in General
Contents
Advanced Bipolar Transistor Modeling (iccapmhb)
Bipolar Transistor Modeling (iccapmhb)
456
Advanced Bipolar Transistor Modeling
Why BJT Modeling Is Still Hot?

BJT is still dominated in RF and analog IC
SGP model is far not accurate for modern BJT's (L<0.6um) for low risk design
SiGe HBT for SOC?
Other advanced HBT process, such as GaAs HBT
What is a Compact Model?

A set of equations and related parameters implemented in circuit simulators
describe the static and dynamic behavior of the device
Parameter extraction is an issue
link people in foundry and design house
Compromise between physic and empirical
Compromise between accuracy and complexity
Multi disciplines / cross sciences
Why Standard Compact Model?

Standardized model formulations to enhance communication among IC and CAD
industries
Standardized interfaces to simplify and speed up model implementation and test for
CAD industries
Better compact model for the latest technologies to lead the edge design
Successful story is BSIM3 model from UC Berkeley
Which model will win as standard model?
VBIC
Mextram
HICUM
History of BJT Modeling

1954: Ebers-Moll model
Enhanced to EM1,EM2,EM3
1970: Gummel-Poon Model
widely accepted as Spice G-P model (SGPGP)
1986: Mextram model - Philips
1994: public domain
1987: HICUM model - Prof. Schroter
1999: public domain
1995: VBIC model
1995: public domain
1999: Bipolar model standardization
Most Exquisite TRAnsistor Model current v503.2, new releasev504 in June 2000, available
in ADS, IC-CAP. HIgh CUrrent bipolar compact transistor Model 1993 extended to HBT,
current v2, used in Rockwell to be available in ADS. HIgh CUrrent bipolar compact
transistor Model
1993 extended to HBT, current v2, used in Rockwell to be available in ADS. Vertical
Bipolar Inter-Company model current v1.1.5, v1.2 code has been just released available in
ADS, IC-CAP.
The Gummel-Poon model is not the same as the SPICE Gummel-Poon model which
everybody has access to in SPICE-like simulators. It wouyld be unfair to Hermmann
Gummel to create the perceiption that these models are the same. Also, the SGPM is not a
"new" model, but about 25 years old (it might not fit under the session title).
Example of Model Evolution
457
The Spice Gummel-Poon Model

Voltage-controlled Current Source Ic(vbe,vbc)
Ideal (n1) and non-ideal (n2) Base Currents in Forward and Reverse Biased Mode
Space Charge Capacitors
Diffusion Capacitors
Bias-dependent Base Resistor
Fixed Base and Emitter Resistors
SGP-Model: DC Formulation
Visualization of the most important DC Effects
SGP-Model: normalized Base Charge q b
458
The Dominant Internal Model Parameter: Normalized Base Charge q b
SGP-Model: Parasitic Resistors

Emitter Resistor RE constant
Collector Resistor RC constant
Bias-depending Base Resistor
SGP-Model: Capacitors
459
SG-P Model: Additional Phase Shift and Temperature

Additional Phase Shift at high frequencies: measured phase is bigger than simulated
Additional phase shift parameter PTF
Frequency Domain: added linearly
Transient Analysis: second order Bessel function
Temperature
Reflects constant ambient temperature
No self-heating
The Spice Gummel-Poon Model

Ideal (n1) and non-ideal (n2) Base Currents in Forward and Reverse Biased Mode
SGP: Parameter Extraction Strategy
460
SGP Extraction: Space Charge Capacitors
Linearizing the measured C(v) values
SGP Extraction: Parasitic Resistors: RE
461
SGP Extraction: DC biasing

Bias Considerationsapplying a Base current instead a Base voltage prevents from
measurements which include the thermal self-heating. Reason: the temperature-varying
vBE is not affecting the measurement, because we force the Base current.
G-P Extraction: Early-Voltages
G-P Extraction: Gummel-Poon Plot Parameters
SGP Extraction: TFF Parameters
462
SGP Extraction: Final Parameter Tuning

Check the modeling result for both, DC and S-parameters
Fine-tune by optimization, if required
SGP Extraction: Final Result - Example
463
Bipolar Transistor Modeling
History of Bipolar Transistor Modeling
1954: Ebers-Moll Model
Enhanced to EM1, EM2, EM3
1970: Gummel-Poon Model
Integral charge control relation
many improvements
widely accepted standard model:SPICE G-P Model
1986: MEXTRAM Model
Developed by Philips
1995: VBIC Model vertical bipolar inter-company model
Developed by a consortium ofover 10 US companies
The Gummel-Poon Model

Ideal (n = 1) and non-ideal (n = 2) Base Currents in Forward and Reverse Biased
Mode
GP-Model: normalized Base Charge q b
464
GP-Model: normalized Base Charge q b
GP-Model: DC Formulation
GP-Model: Parasitic Resistors

Emitter Resistor RE constant
Collector Resistor RC constant
Bias-depending Base Resistor
465
GP-Model: Capacitors
G-P Model: Additional Phase Shift and Temperature

Additional Phase Shift at high frequencies: measured phase is bigger than simulated
Additional phase shift parameter PTF
Frequency Domain: added linearly
Transient Analysis: second order Bessel function
Temperature
Reflects constant ambient temperature
No self-heating
GP-Model: Model Parameters

Using Default Parameter Values switches-off the Parameter Effect (do not confuse with
Typical Parameters).
466
G-P Extraction: Notes
For a better overview, we will use these simplifications:
no parasitic capacitors
no series inductors
no temperature modeling
Simulator: spice2
IC-CAP example file: gp_classic.mdl
G-P: Parameter Extraction Strategy

Parameters are dependent on each other:
extraction sequence is important

alternate approaches are possible
CV
Space Charge Capacitances
DC
Parasitic Resistors
Early Voltage
Gummel-Poon Plot Parameters
S-Parameters
RB Parameters
TFF Parameters
G-P Extraction: Space Charge Capacitors
G-P Extraction: Parasitic Resistors: RE
G-P Extraction: Parasitic Resistors: RBM

467
G-P Extraction: DC Biasing

Bias Considerations
Applying a Base current instead a Base voltage prevents from measurements which
include the thermal self-heating.
Reason: the temperature-varying vBE is not affecting the measurement, because we force
the Base current.
When measuring the Gummel-Plot with half the maximum output characteristics voltage
vCE, we will avoid the problem of the poor quasi-saturation modeling of the Gummel-Poon
model.
G-P Extraction: Early-Voltages

VAF, VARapplying a tangent to the output characteristics intersects the x-axis at VAF resp.
VAR
468
G-P Extraction: Gummel-Poon Plot Parameters
G-P Extraction: S-Parameter De-embedding

Since the S-parameters of a transistor are always affected by parasitics, de-embedding is
required. It is achieved in IC-CAP with a simple PEL program, suited to the specific needs
e.g. de-embedding using Y matrix subtraction only: usually ok for <10GHz-> or full OPEN-
SHORT de-embedding like below).
G-P Extraction: S-Parameter: Data Management

In order to reduce NWA calibration time and S-parameter measurement time
considerably, we will use the data managementfeature in IC-CAP:i.e.:
the NWA is calibrated only once for a full frequency sweep. Then, the S-parameters are
measured for all DC bias conditions, de-embedded once and finally stored to an IC-CAP
.mdm file. The parameter extraction setups for rBB (frequency swept, ib swept,
vc=constant) and TFF (frequency=constant, ib swept, vc swept) are then re-imported as a
subset of de-embedded data out of this .mdm file.
G-P Extraction: Data Management ASCII File

469
G-P Extraction: RB Parameters
G-P Extraction: TFF Parameters
470
G-P Extraction: Final Parameter Tuning

Check the modeling result for both, DC and S-parameters
Fine-tune by optimization, if required
471
BSIM
Contents
BSIM3.3v2 - Parameter Extraction Strategy
Fully Scalable BSIM4 Modeling From DC to RF
BSIM3.3v2 - Parameter Extraction Strategy
Since IC-CAP version 5.2, a special IC-CAP Toolkit is available for BSIM3 modeling.
Therefore, for a detailed technical documentation about the model and the parameter
extraction, see the IC-CAP manuals.
The following figures are intended to give an overview, which model parameters affect
which fitting problem. The sequence given here is also a possible model parameter
extraction strategy.
LARGE: Transconductance id(vg, vb)
VTH0: equally shifting the transconductance curves (threshold voltage)
LARGE:
NSUB: shifting the threshold voltage
472
LARGE:
NCH: 2nd order shifting of the threshold voltage
LARGE:
U0: slope of the transconductance curve (Mobility)
LARGE:
UA: change in slope of the transconductance curve (Reduction of the mobility U0).
LARGE:
UC: changing the slope of the mobility related to vB
LARGE:
UB: changing the slope for big values of vG
473
LARGE:
VOFF: y-intersect for a half-logarithmic transconductance curve
LARGE:
NFACTOR: slope of the half-logarithmically plotted transconductance curve
SHORT: Transconductance curve id(vg, vb)
NLX: an equal shift of the transconductance curve for the SHORT
SHORT:
RDSW: change in slope of the transconductance curve for the SHORT
474
SHORT:
LINT: higher order change in slope for higher Drain currents
NARROW: Transconductance curve id(vg, vb)
K3: shifting the transconductance curves
NARROW:
K3B: 1st order shifting of the transconductance curves
NARROW:
W0: higher order shifting
475
NARROW:
WINT: change in slope
SHORT: Output characteristics id(vd, vg)
VSAT, A0, KETA: amplification of the output characteristics level
SHORT: Output resistance rout(vd, vg)
PCLM: Fitting for low vD channel length modulation

PDIBLC1, PDIBLC2: Fitting for medium vD drain induced barrier lowering
PSCBE1, PSCBE2: Fitting for high vD substrate current induced body effect
Fully Scalable BSIM4 Modeling From DC to RF

Company Information
AdMOS offers user specific and flexible solutions for a wide variety of applications in the
area of modeling and simulation.
Engineering services:
Modeling service to generate design libraries for semiconductor devices (CMOS down
to 90nm, RF CMOS, etc.)
476
Simulation models for passive components like on-chip passive, PCB elements for
high speed circuits, connectors and packages.
For more info, please see www.admos.com

The principal services and products of Advanced Modeling Solutions are in the area of device
modeling and connector design.
Standardized and reusable measured data sets for MOS modeling

Data reuse for different models
The second part of the database strategy is to add additional information for each derived type of model
to the common set of measured data.
A user can define a certain extraction strategy (order and type of extraction functions) for a certain model,
e.g. BSIM4 which is used to determine the model parameters.
This extraction strategy is stored together with the final results (model parameters) and the boundaries of
the parameters in the project data base.
This makes all parameter extractions repeatable and self consistent.
Test devices for STI stress effects

A third dimension is added by BSIM4.3.0, with SA, SB (SD) as parameters. In this
diagram, the Temperature as a 4th dimension is neglected.
Until BSIM4.2.1, gate length (L) and gate width (W) have been the major device geometry
parameters. The layout of test devices was focused on these two instance parameters to
account for length- and width scaled effects.
Data concept
477
Data operation
Measurement concept
The measurement environment is in independent from any particular simulation model.
This slide describes the key features of the new BSIM3 Modeling Package. The following slides will show
each of these features in detail.
The resulting database of .mdm files can easily be generated by other programs than IC-
CAP, e.g. by high volume production measurement environment. All files are in ASCII
format.
478
General Parameter Extraction features in the BSIM4 Modeling

Package
Configurable Extraction Flow
This feature can be configured.
The extraction methodology in the BSIM3/4 Modeling Package is open and can be
configured very flexible to adopt the software to a certain CMOS process.
In many cases, model parameters are grouped together when they describe the same
physical effects. These extraction groups can be arranged in the first hierarchical level, the
extraction flow, in any order. It is also possible to invoke a certain group more than once
if necessary.
Inside an extraction group, the flow of the functions (second hierarchical level) to
determine a certain model parameter can be set by the user. For each model parameter,
either the physically based direct extraction method, the manual tuner or the optimizer
can be invoked to determine this parameter. In many cases, the tuner or the optimizer
are modifying different parameters in one run, which have been extracted separately.
Using this method, the complete extraction strategy can be easily modified and adopted to
the special need of a certain CMOS process.
Once a certain extraction strategy is established, the fully automatic run can be used. In
this mode, all extractions are called in the given sequence inside a macro and the
intermediate results are logged either in the failure log file or the output window log file.
This method is ideal to generate a large number of model parameter sets from different
measurements for a further use in a statistical modeling approach.
Extraction Method
As already mentioned in the last slide, the BSIM3/4 Modeling Packages support different
methods of the determination of model parameters.
The first one is the physically based direct parameter extraction. Here, the parameters are
479
calculated according to their meaning in the model equation directly from measured data.
These functions carefully check whether certain physical effects can be seen in the actual
data set and they isolate the valid range of measured data to determine a certain effect.
An example for such a parameter extraction for different width scaled devices is given in a
later slide.
The pre defined tuners and optimizers automatically select reasonable optimizing targets
and ranges. They contain a dedicated parameter check and can be customized by the user
as well.
Moreover, the BSIM3/4 Modeling Packages support the usage of the new plot optimizer
feature. In a special setup, certain diagrams from different devices at different
temperatures can be grouped together to build the base for applying the plot optimizer
feature. Once such an optimization is set up, it can be included in the extraction flow.
Example: Direct Extraction for BSIM4
This examples demonstrates the generation of such a condensed data array to extract the
width dependent bulk charge effect.
Here different devices with different gate widths and a constant large gate length are
taken to isolate this gate width dependent effect. From each device, the drain currents at
the maximum gate voltage and at different bulk voltages are extracted and all values are
arranged in a new diagram on the right hand.
Now, this diagram has the gate width as an x-axis while different curves accounts for
different bulk voltages Vb.
This diagram is the base for further extractions or optimizations as shown in the next
slide.
In this slide, the new diagram is taken as a base for either direct extractions or
optimizations.
On the left hand, the two equations demonstrate the principle of applying a linear
regression method to determine the both parameters B0, B1. This can be done by using
the pre-assumption that the gate length is constant while the gate width is variable. Then,
after a view transformations, which are not shown in this slide, B0 and B1 can be
determined from the slope and intersection of the applied tangent.
Moreover, the condensed diagram can be easily used to apply an optimizer or a tuner on
the gate width dependent data set (see the right hand example). While the measured data
is combined by data base operations, the simulated data can be directly generated. This is
achieved by defining the gate width as a parametric input with sweep order 1 (equal to
the x-axis of the diagram) and the bulk voltage as an input with sweep order 2. Using this
setup, any optimization runs very quickly invoking the built-in spice3e2 simulator or any
of the supported commercial simulators.
Example: Plot Optimizer
480
The plot optimizer allows an easy fine tuning of model parameters for different:
devices
setups
regions of operation at once
Binning
Binned models for BSIM3 and BSIM4 can be generated. They can be extended by
extrapolation to cover a wider range of device dimensions than available through
measurements.
It can be configured, whether a certain effect which depends on the device geometry,
e.g. the threshold voltage, is modeled by the built-in scalable model equations or by
the binning feature.
Taking only a few parameters as binning parameters can enhance the overall fitting
quality while preserving the mostly physical nature of the model.
Binned model sets are generated ready to use in the ADS, HSPICE and Spectre
format.
The BSIM3/4 Modeling Packages support the generation of binned models. They can
be extended by extrapolation to cover a wider range of device dimensions than
available through measurements. Without this extensions, many simulators would
not be able to simulate devices with W=10um or L=10um ! The condition for
selecting a certain binned area is:
Lmin <= L < Lmax.
That means, a device with L=10um cannot use the parameter set of bin #2 even but
must select the parameter set of the extended bin #9.
Moreover, it can be configured, whether a certain effect which depends on the device
geometry, e.g. the threshold voltage, is modeled by the built-in scalable model equations
481
or by the binning feature.
Taking only a few parameters as binning parameters can enhance the overall fitting
quality while preserving the mostly physical nature of the model.
Binned model sets are generated ready to use in the ADS, HSPICE and Spectre format.
Advanced BSIM4 RF models

Built-in BSIM4 RF Features
Missing BSIM4 RF Features
The BSIM4 RF model takes only into account the behavior of the semiconductor"
part of the multi-finger transistor. The influence of the metal layers (crosstalk, delay)
is not considered
Gate length variations (see one of the next slides) inside a multi finger device are not
taken into account
The substrate resistance network is not scalable.
Structure of Scalable BSIM4 RF Macro Model

A sub circuit contains the core BSIM4 model surrounded by external elements
representing crosstalk and delay behavior of metal layers.
The values the marked elements are calculated as functions of the device dimensions W, L
and NF rather than setting them to fixed numbers.
Scalable BSIM4 RF Macro Model
Extensions to BSIM4
Scalable substrate resistance model
Channel length reduction variation inside multi finger devices is included
Crosstalk and delay effects due to metal interconnection in multi finger device is
included.
482
Scalable transistor models cover a certain range of major device dimensions. In case of a
RF MOS transistor, these are the gate length and gate width of a single transistor finger
and the number of gate fingers. These models have a structure, such that a design
engineer can change the parameters to get an optimum transistor behavior for a certain
application.
The fully scalable model has in principle the same structure than the single transistor
model. However, it uses equations to determine the values for the substrate resistance
parameters and the modified channel length reduction. These equations are derived from
simple assumptions. Please see the following sketches and the model equations for more
details.
Example: Scalable Substrate Resistance

Standard BSIM4.3 model:
RBPS, RBPD, RBSB, RBDB are fixed numbers
Scalable RF model:
Implementation of a set of equations according to [1] and own developments to describe
a scalable substrate resistance behavior:
In the RF single device model, the values of the resistors RBPB, RBPD, RBPS, RBDB and
RBSB in the substrate resistance network parameters are set directly to fixed values.
In contrast, for the RF scalable model, the substrate resistance network parameters RBPB,
RBPD, RBPS, RBDB and RBSB are derived using 4 new model parameters:
RSHB sheet resistance of the substrate

DSBC distance between the source contact and the outer source area
DDBC distance between the drain contact and the outer drain area
DGG distance between two gate stripes
The equations to determine the values for the substrate resistance parameters contains in
addition the gate length, gate width of one finger and the number of fingers.
Measurement and De-embedding

The calibration plane of the NWA is located at the end of the probe tips.
The influence of the parasitic elements between the calibration plane of the NWA and the
transistor must be eliminated through de-embedding.
483
The goal of a S-parameter measurement of an integrated RF transistor is to determine the

frequency behavior of the transistor itself without taking into account the parasitic
influence of the cables, probes, pads and the interconnection to the transistor. The first
step to achieve this goal is to calibrate the network analyzer using a special probe
substrate. This will move the calibration plane to the end of the probe tips. However, the
parasitic influence of the pads and the interconnections to the transistor still remains. It
must be eliminated by using de-embedding methods. After all these procedures, the
remaining S-parameters describe the real behavior of the transistor as it will be used in a
design library.
The two diagrams on the right hand demonstrate the effect of the de-embedding. The
lupper one shows the input and output reflection S11 and S22 before and after the de-
embedding. Looking at S22, the difference in both, phase and magnitude can be seen very
clearly. A similar behavior can be observed in the transmission behavior. The raw
measured data of the forward transmission S21 clearly has an increased phase shift and a
decreased magnitude compared to the de-embedded data. As a summary, the raw
measured data underestimates the real behavior of the RF MOS transistor. It is absolutely
necessary to perform a correct de-embedding to get the correct transistor data as a base
for accurate modeling.
De-embedding and verification test structures
The pre requisite for a correct de-embedding is that certain test structures are available
on a wafer together with the device under test (DUT) itself. Depending on the selected de-
embedding method (see the next pages) an OPEN, SHORT or THROUGH dummy pad
structure must be available and must be measured. The principle layouts of these
structures are given above. However, in real layouts, the transistor itself can be much
smaller compared to the size of the dummy pads.
Together with the principal layouts, the ideal equivalent electrical circuits for these devices
are given. The described de-embedding methods are only valid, if the electrical behavior
of the dummy pad structures can be represented by these equivalent circuits.
De-embedding Procedures
OPEN
This method is limited to lower frequency applications (<3GHz) because it takes only into
account parasitic effects parallel to the DUT.
SHORT-OPEN
Should be applied for data measured at frequencies roughly > 3GHz.
484
This slide describe the two standard de-embedding methods which are implemented in the
BSIM3/4 Modeling Package.
A common and very often used de-embedding method is the de-embedding from an OPEN
device. The pre requisite for using this method is, that the behavior of the OPEN dummy
pad structure can be represented only by parallel elements to the DUT (see the equivalent
circuits). The transformed Y-parameters from the OPEN can simply be subtracted from the
DUT's Y-parameters. The advantage of this method is that only one device in addition to
the DUT is needed. However, this method is limited to lower frequencies roughly smaller
than 3-5 GHz. This frequency limit is not fixed and depends on the layout and size of the
whole test structure.A more advanced method to remove the parasitic influence from the
measured data is the de-embedding using an OPEN and SHORT dummy pad device. The
idea behind this method is, that the electrical behavior of the pads around the DUT can be
described by a combination of parallel and serial circuit elements. The implementation is
shown in the right part above. In a first step, the influence of the parallel parasitic
elements is removed from both, the SHORT and the device under test (DUT). In the
following step, the Y-parameters are converted to Z-parameters and in this
representation, the influence of the serial parasitic elements is removed by subtracting the
partially de-embedded Z-parameters of the SHORT from the partially de-embedded Z-
parameters of the DUT. After a further conversion, the S-parameters of the DUT are
available as a base for the device modeling.
Besides these both standard methods, the BSIM4 Modeling Package allows the
implementation of user defined de-embedding methods.
De-embedding verification with THROUGH

After a correct de-embedding of the parasitic components, the S-parameters of the
THROUGH should show the behavior of an ideal transmission line:
Delay time TD representing the distance between the two pads.

S11, S22 starting at characteristic impedance.
To ensure that the de-embedding is done correctly, the verification using a THROUGH
device is an ideal method. The verification is very important because if not all
parasitic effect are removed from the transistors measured data, the simulation
model must fit unknown parasitic effects and does not represent the pure transistor,
as it will be used in a design library.
The idea behind the verification is, that a THROUGH should behave like a
transmission line after the de-embedding of the parasitic pad influence. The diagrams
show the typical behavior of such a through before and after the de-embedding
procedure. The blue curve shows this ideal transmission line behavior which is given
by a phase shift in S21 due the delay time of the line and S11, S22 starting at the
characteristic impedance in the Smith chart.
DC model modifications
The first step in achieving a precise RF model is to modify the DC model to match both,
output resistance Rout together with the S-parameters at the lowest frequency. The
following effects have to be considered:
Output resistance: Has to be readjusted to match S22 and S21 at lowest frequency.
Channel length reduction DL :Due to lithographic effects, the channel length
485
reduction can show a significant variation inside a multi-finger transistor.
If the target simulator does not provide a multiplier for several gate fingers, the
following effects have to be adjusted:
drain source resistance Rds
channel width reduction W
After preparing the measured S-parameters, the real modeling task starts looking at the
DC model parameters. DC model parameters are normally derived from measurements of
single finger transistors of different gate lengths and gate widths. However, the typical RF
MOS transistors are implemented as multi finger transistors which may show a different
behavior than a singe finger transistor. The first step is, that the output resistance
parameters mostly have to be re-adjusted to match both, the output resistance Rout =
f(Vd ) and the starting point of the S-parameters S21 and S22. This behavior will be
shown in detail in the next slide. Another effect, which can often be observed within multi-
finger transistors is, that the channel length reduction DL is different compared to single
finger transistors. A research work of one of our customers showed, that DL can vary up
to 30% within a multi finger transistor. Without adjusting DL in the simulation model, an
accurate device behavior cannot be simulated. At least, depending on the target
simulator, the drain source resistance and the channel width reduction have to be re-
adjusted in the case that the target simulator does not have a multiplier for n numbers of
gate fingers.
Gate length variation in multi-finger transistors
This diagram shows the effect of gate length variation inside a multi finger device, which
can be observed in many CMOS processes. The lower sketch is a principle cross section
through a typical multi finger devices with 6 gate fingers. It could be shown through SEM
(Scanning Electron Microscope) pictures, that the length of the poly-Si gate is different for
the outer gate fingers than for the inner gate fingers. However, that means, that the
channel length reduction, which is normally derived from single finger transistors does not
correspond to the effective channel length in such multi finger transistors.
When doing modeling, it is naturally not always possible to make this SEM analysis and it
is mostly necessary to rely only on measured electrical data. The impact of this effect can
be seen very good in a shift of the threshold voltage Vth, which is through the short
channel effect heavily dependant on the effective gate length. In addition, it can be also
seen in a difference between measurement and simulation of the drain current at high
drain and gate voltages. However, if such an effect can be identified, it is necessary to
compensate and to readjust the parameter LINT in the BSIM simulation models. The
disadvantage of such a procedure is, that now the RF simulation model contains different
values for the model parameters than the model for the conventional single finger devices.
DC output resistance and S-parameters
This slide shows the correlation between the DC behavior and the S-parameters of the
DUT. The starting point of the S-parameters at lowest frequency (here 100MHz) are only
determined by the DC model parameters of the transistor model. The lower right diagram
shows the magnitude of S22 at different DC operating conditions at the lowest frequency.
The very good fit between measured data (red) and simulated curves (blue) is only
486
achieved by a perfect fit of the DC behavior, mainly the output drain current and the
output resistance Rout. Without correct starting points of the S-parameters at the lowest
frequency, no good overall fit can be achieved. This is again an indication, that only
physically based DC model parameters can be used for further RF modeling.
Extraction of external circuit elements
Example:
Extraction of the gate inductance from S-parameters measured at zero bias voltage.
After fitting the starting points of the S-parameters, the model parameter values of the external elements
are determined from S-parameter measurements over the full frequency range. In the example shown
above, the value of the gate inductance is determined from a measurement on the cold transistor, where
the gate and drain voltage is set to zero to switch of the transistor influence.
Sample Results
Frequency range: 0.1 .. 20GHz, L=0.24um,

Shown above are the final results of the model generation using the extended BSIM4 RF
model. The red curves are representing the measured behavior of the MOS transistor at
different DC operating conditions while the blue ones are the corresponding simulation
results. It can be seen, that the fit between measured and simulated is very accurate. This
could only be achieved by using physically based extractions which results in realistic
values for resistance, channel length reduction etc. It can also be seen that the
modifications of the extended BSIM3/4 RF model have been necessary to achieve this
simulation results.
Advanced usability features
487
Test & Measurement
Configurable Extractions
Features have been developed according to experience from AdMOS modeling service.
Define binned models
Simulator Support
The extraction module generates two types of outputs:
488
Outlook
Roadmap of BSIM Simulation Models
The diagram shows the roadmap of simulation models which are based on the BSIM3v3
model. Two principal branches can be identified which cover the mainstream of
"conventional" bulk CMOS devices and the upcoming Silicon On Insulator (SOI)
technology. All the models in both branches are based on the BSIM3v3.1 model.
BSIM3v3 is a quasi standard model in the semiconductor industry. However, to cover all
the physical effects of the steadily scaled MOS devices, UC Berkeley published the new
BSIM4 model in March 2000.
In the second branch, the new BSIMSOI 3.0 models now unites the fully depletion and
partially depletion mode of SOI transistors which was represented in the past by different
models.
Outlook
The BSIM3/4/SOI Modeling Packages will be kept up to date with the current model
versions released from UC Berkeley (see anouncement in the CMC minutes).
In a further step, AdMOS and Agilent Technologies are looking into the Next
Generation MOS models (BSIM5, PSP, HiSIM2, EKV3.0), which are currently in the
evaluation of the Compact Model Council. Our goal is it to provide supperior MOS
modeling support beyond the 90nm CMOS node.
489
GaAs Transistor Models in General
Different to Silicon Transistor models, GaAs models have a relatively small number of
model parameters. Another differentiator to silicon is that there have always been 'many'
models for GaAs transistors, compared to the pretty clear structure for MOS (BSIM, PSP
etc.) and for bipolar (Gummel-Poon, VBIC, Hicum etc.).
Also, in many cases, modeling engineers apply PI-schematic-based models for individual
DC bias points, and then apply scaling vs. DC bias to obtain the general model. (See the
chapter about Y-parameter modeling.)
This method is quite similar to the approach with the PSP MOS model.
Last not least, GaAs transistors exhibit in many cases a frequency trapping in the 10kHz
range. Therefore, some GaAs models exhibit model parameters for modeling the DC-gm,
and fully decoupled RF parameters for modeling the RF-gm.
490
General Transistor Modeling Strategies
The guidelines in this slide set present a general-purpose transistor modeling strategy,
independent of the selected model, but focusing on modeling single transistors.
In this slide set, we consider a bipolar transistor as an example. The same strategy can
also be applied to MESFETs, single MOSFETs etc.
Note
The proposed sequence of measurements, data verification and final device modeling can also be applied
to scaling transistor modeling like for MOS (BSIM, PSP etc). Of course, the modeling strategy for these
scaling models is then different from the here proposed one.
Measurements
from the demo_features directory
demo_features/3_MEAS_ORGANIZE_n_VERIFY_DATA\0_MASTER_FILES
load for example the file
NPN_MEAS_MASTERFILE_PELdep.mdl
and perform the predefined standard measurements.
Export the measured data into mdm files (Macro EXPORT_DATA)
Note
besides the conventional transistor measurements, you find also in this MEAS_MASTERFILE Setups for the
measurement of the probe contact resistances, the NWA cal.verification and the de-embedding
verification.
Screenshot of NPN_MEAS_MASTERFILE_PELdep.mdl
Data Consistency Checks DC

After the measurements have been performed, it is absolutely important to check the
consistency of the data:Concerning the DC data, you should check the following:
does a transfer curve, converted to a pseudo-output curve, match the original output
curve?
do the sweep limits of the transfer and the output characteristics match each other, or are
the transfer sweeps in another voltage range than the output sweeps?
are the DC contact resistances known?
From the demo_features directory, load the file
demo_features\3_MEAS_ORGANIZE_n_VERIFY_DATA\0_MASTER_FILES\
1__CHECK_DATA_CONSISTENCY\NPN_MEAS_CHECKTOOLS_PELdep.mdl
import the corresponding subset of the previously exported meas.data and follow the
READMEs in the model file.
Data Consistency Checks S-Parameters

1. DC Bias Conditions of the S-parameters:
Are the DC bias currents of the S-parameter measurements identical to those of the
DC-only measurements?
491
Are the DC voltage sweep ranges of the S-parameters identical to the DC-only
measurements?
Are the DC contact resistances known?
1. Continue applying the file NPN_MEAS_CHECKTOOLS_PELdep.mdl
NWA calibration verification:

demo_features\3_MEAS_ORGANIZE_n_VERIFY_DATA\0_MASTER_FILES\
2__NWA_CAL_VERIFICATION\CAL_VERIFY_MASTERFILE_PELdep.mdl
De-embedding Verification:
demo_features\3_MEAS_ORGANIZE_n_VERIFY_DATA\0_MASTER_FILES\3_DEEMB_VERIFICATION\
CHECK_OPEN_SHORT_DEEMB_with_a_THRU_MASTERFILE_PELdep.mdl
Wrap-up: The applied checktools and verification

model files:
Enter the DC contact resistances and also the measured ohmic losses of the NWA's S-
parameter Testset into your MASTER_EXTRACTION.mdl file
for example:
demo_features\1_BASIC_MDLG_EXAMPLES\21_Gummel_Poon_demo\GP_CLASSIC_NPN_1998.mdl
Note
We will refer to this model file during the rest of this course.
After all the data have been checked, we can finally

start the modeling
usually, the CV modeling is performed first, since in many models (especially the bipolar
models except Gummel-Poon) the capacitance parameters affect the DC modeling.
Additionally, the CV model parameters are not overlaid by any other model parameter,
what makes them the ideal parameter to start our modeling work.
492
Hicum2 is an exception to all other models, here you need to start with modeling the
transit time ft.
Example: CV Modeling Gummel-Poon
In a second modeling step, the parasitic ohmic resistances are extracted from special
measurements. Either from DC measurements or from hot S-parameters (strong bias
current into the Base or Gate (MSFETS), and half the current suck out of the Collector or
Drain. However, no simulation or optimization is performed at this step, since the other
DC parameters are not yet known.
Note
You can also skip this step and extract the ohmic parameters from the transfer curves (Gummel-Poon
plot) and the output characteristics for low vCE or vDS.
493
Note
Extracting the parasitic resistors and inductors from Spar measurements and the
NPN_MEAS_CHECKTOOLS file.
Now we are ready to perform the DC modeling. This DC modeling is the fundamental step
for all transistor modeling, since the DC curves determine with 100% the starting point of
the S-parameters at lowest frequency (as will be seen later). Therefore, an as good as
possible DC modeling fit is mandatory for a good model.
Note
For bipolar models, we first need to fit the slopes of the foutput and routput curves (Early voltage
parameters), but still ignore the fitting of the curves themselves. This is necessary since the Early voltage
parameter affects the Gummel-Poon plot iC(vBE). The final curve fitting will be possible after the transfer
curves (Gummel-Poon plot) are fitted. For all other types of transistors, we usually skip this extra step and
commence directly with the Transfer Curve modeling.
The bipolar intermediate step: fitting the Early Voltage parameter for foutput and routput,
494
but ignoring the matching of the curves. Only the slopes are considered for the moment.
This is the intermediate result of the Early modeling for bipolar transistors. As you can
see, only the slopes are fitted (Early Voltage Parameters), but the curves themselves do
not fit yet. As mentioned, for all other other transistor types, this step is skipped.
We are now back with a common, transistor-type independent modeling strategy. Usually,
we bias a transfer curve Setup such that the Drain or Collector voltage is half the max.
value of the output characteristics. This means that once we have fitted the transfer curve
(Gummel-Poon curve for bipolar), we have automatically also fitted the output
characteristics at that selected Drain/Collector voltage. All what remains is to adjust in the
output curve is the slope (Early effect) and the quasi saturation range.
Transfer Curve(s) Modeling

For bipolar transistors, modeling the transfer curve (Gummel-Poon Plot) means to model
the input characteristics iB(vBE), and the two transfer characteristics iC(vBE) and
beta(vBE).
For FETs with non-zero Gate current (e.g. Junction-FETs etc, we model the input
characteristics iG(vGS) with vDS=0
495
and the transfer characteristics iD(vGS) and gm(vGS).
For FETs with zero Gate current, it's only iD(vGS) and gm(vGS).
Example Bipolar Transistor (Gummel-Poon): iC(vBE) modeling
Output Characteristics Modeling

After the transfer characteristics has been modeled, we finally come to the output
characteristics modeling. Since you can think of the transfer characteristics as a cut thru
the output characteristics for a given vCE or vDS, the fitting we have applied so far to the
transfer characteristics must match the output curves at exactly that vCE or vDS.
With the Transfer Curve(s) fitted, the Output curves match perfectly at the Collector/Drain
voltage used for the transfer curves.
Typically, only a fine-tuning of the Early voltage and the Collector (Drain) resistance is
required for obtaining the final fit.
496
S-parameter Fitting
We are now ready to commence with the S-parameter modeling

In the first step, we switch off all extracted or non-zero capacitances (i.e. we divide all
extracted capacitance parameters CJx by 1E6, what means we do not loose their
extracted values, but only switch-off their effect).
The same with transit time parameters.
497
If the data consistency check was o.k., if the de-embedding was perfect and if the DC
modeling was perfect too, then the simulated S-parameters should all be located on the x-
axis, and exactly at the extrapolations (freq -> 0Hz) of the measured S-parameters
(whicih are now representing an ideal, frequency-independent transistor).
If this is not the case, then consider modeling the RB parameters of the bipolar transistor
(cannot be modeled correctly in DC, only in S-parameters), consider the bias TEE losses,
consider the DC contact losses etc.
In any case, consider a DC parameter.

Finally, switch on again the capacitance and the transit time parameters and copy from
the DEPOTS.mdl file the GUI. CopyPaste_ITEMS/S2Y_Mdlg_GUI into your
MASTER_EXTRACTION model file.
Apply the GUI for separating individual curves (CV curves, ft curves etc) out of the S-
parameter. From the up-popping plots, select PlotOptimizer and perform a parameter
tuning or optimization on the selected plots.
Note
Hit the HELP button in the GUI for detailed explanations and help.
498
Conclusions
The presented general transistor modeling strategy should assist you in finding the golden
modeling path' for your special selected transistor model.
499
Overview of the Hicum Level 2 model
Forewords
Compact models are only a simplified representation of real life:
There's always a bias, frequency, temperature or geometry range for which a given model
is not accurate enough!
One should keep this in mind when doing modeling in order to achieve the best possible
compromise.
The best compromise should match the following requirements:
describe the complete process behavior (not just dumb fitting)

avoid unphysical behavior
focus on bias, frequency, temperature and geometry ranges that are relevant to
circuit design
What is compact modeling?

Many significations behind the word modeling:
Note
Behavioral languages are more and more used for compact models definition (see CMC web site)
Need for accurate models
HiCUM can help circuit design

Better RF modeling:
1. Better ICK modeling

2. Better description of abrupt fT falloff
3. Better maximum fT voltage dependence modeling
This is important for many applications:
Selecting Circuit bias

for PAs where current can go beyond peak fT
500
Output conductance modeling (good derivatives are required for large signal simulations)
Large Signal behavior useful for PA design
LNA design -> associated to a scalable model, designers can choose optimal transistor
Bandgap references -> Very difficult to obtain correct trends. Not only requires "fitting"
but physical results: non physical XTI gives wrong results!
501
HiCUM is also better for Scaling and Statistics

The model physics based definition allows a better geometry scaling
Note that there are no scaling equations in the model!

But it's easier to describe model parameters variations with geometry
Scaling is usually achieved via an external program (able to deliver static libraries) or
using simulator macro language
In the same way, it's easier to describe statistical variations with HiCUM
Corner models (be careful of definition)

True Monte-Carlo (process)
Matching
Additionally HiCUM can be used to realize predictive models
Obtain realistic trends for next process generation

Start designs early then switch to finalized models close to tape out
Statistics can help designers

Designers are sometimes skeptical about Monte-Carlo Simulations: they are right!
If one uses model parameters directly as Monte-Carlo simulation input : correlation is lost
(left figure)
HiCUM help delivering TRUE statistical models which give the right trend (on the right)
HiCUM Core Model

HiCUM Equivalent Circuit
Transfer current
Infinite ICH leads to no effect,
502
Reducing ICH increases IC at high current densities (only for 2D and 3D effects)
The curve depends also on Capacitances, weighting factors and transit time
Note
C10/QP0 = IS saturation current
Base Current
BE base current
No Beta unlike SGPM. Identical behaviour for internal and peripheral currents
Note
If no volume recombination occur, MBEI=1 (except a very small neutral emitter modulation)
BC base current
No Reverse Beta unlike SGPM
Identical behavior for internal and external current
Note
Usually, due to GSG test structures (grounded emitter) reverse characteristics are difficult to obtain
VE=VB=0 and VC negative can be used (CS junction and parasitic PNP are on.)
Depletion charges and capacitances

503
The well known capacitance formulation is:
With C j0 the zero bias capacitance, Vd the junction potential and z the grading coefficient.
This formulation is not satisfying from a numerical point of view:
In SGPM the function is made linear above FC
The capacitance is still increasing, but negligible compared to the diffusion capacitance
In Hicum, the capacitance value at forward bias plays a role:

Used to compute qjEi and qjCi used in the GICCR (Early effect)
A different behaviour has to be used !
BE Capacitance
The BE capacitance formulation is based on weighting functions e1 and e2:
The capacitance is split in two parts using e 1 and e 2
Effect of the ALJEI parameter on the capacitance:
3 Weighting functions:
The capacitance is divided in 3 parts (High reverse bias, reverse and low forward bias,
high forward bias) With weighting functions: The punch through function (large reverse)
depends on parameter VPTCI. Here VPTCI = 2 V.
504
Effect of the VPTCI parameter on the capacitance:

This effect is more pronounced for devices with low collector doping profiles (No SIC). Plot
of CjCi for VPT= 30, 6, 3 and 2 V.
Transit Time
The transit frequency fT is a consequence of stored charges which are described via the
transit time:
The transit time model in HiCUM is separated into 3 parts:
At low current density: the transit time (TF0) has a voltage dependence (VBC):
At medium and high current densities: the critical current ICK has a voltage
dependence (VCE)
At high current densities: the transit time increase has a current dependence (IC)
and a voltage dependence via ICK.
Below is presented the transit time increase Delta_tf versus collector current
(normalized with respect to the critical current ICK) and the influence of the HiCUM
model parameters THCS and ALHC on the curve.
Transit Frequency
505
Transit Frequency and Critical current

Ft Drops when IC reaches the Critical current ICK.
fT(IC, VCE)
Critical current ICK

ICK versus VCE plot:
VPT= 1000
VLIM = 100
VPT and VLIM have no effect
The slope at VCE = 0 is 1/RCI0
Plots for RCI0=100, 200 and 300 Ohms

Critical current ICK
VPT= 1000
RCI0 = 300
VLIM = 2, 3 and 4 V
The saturation value is:

Ilim= VLIM / RCI0 =6.66 mA, 10 mA and 13.33 mA
506

RCI0 = 300
VLIM = 3 V
VPT= 10, 20 and 50 V
The slope at high VCE is: Ilim / VPT
Scalable Modeling
Scalable models are mandatory for Bipolar devices to:
Reduce parameter extraction and measurement effort

Work on reduced transistor sets
Reduce test chip area
Provide more transistor configurations to designers
Allow modeling of non measured devices
Help designers to optimize circuits wrt device sizes
Scaling equations must be continuous and differentiable
Improve statistical modeling
Model geometry induced statistical effects
Note
Scaling methodologies also help to understand and capture physical effects and their
geometry dependence which could simply not be done with single transistor approach
Method #1: MOS like, eg bsim3

Scaling equation included in model code.
In Bipolar scaling equations are too complex: would need a new model every year
Models need to apply to a wide variety of bipolar processes plethora of scaling
equations in model code
Model code never stable for most advanced technologies
Burden for model developers, less support, less research
Pressure quick fix of deficiencies with fitting parameters = bad model for a long time
507
Method #2: SMACH Extraction Toolkit

HiCUM with simulator macro language
Scaling equation included in macro language
Scaling equations and model core decoupled

Very flexible solution
Complex equations difficult to handle with simulator macro language
Need efficient tracking of scaling equation and rigorous update of related extraction
methods
No scaling rules IP protection
Solution used in SMACH: Very efficient when flexibility is the priority
Method #3: TRADICA /HMT EXTRACTION TOOLKIT

TRADICA approach
Scaling equation in an external program
Advanced scaling equations can be used

Model library has very simple syntax (simulator support)
Scaling equations and model core decoupled
Models can also be produced dynamically
Scaling rules IP protected
508
Method #2 and #3 can be used to apply "external" scaling to HiCUM, but no scaling
equations are in the model code ...
So, why HiCUM is considered as a SCALABLE MODEL?
Every HiCUM model parameter can be expressed wrt physics based quantities and
geometrical quantities relying on physics based approach
Note this is the key of successful modelling activities. This is also the most powerful
path to capture proper statistical variations
The ultimate goal is to model the process, not a single die!
Scalable Modeling: the TRADICA/HMT example

TRADICA produces scalable model parameters (according to scaling rules and geometry
independent model parameters)
Parameters are mapped in single device model cards for any configuration: width, length,
number of emitter stripes, number of base contacts, collector contacts.
Can be done statically or dynamically

Geometry independent parameters are extracted from a subset
Virtually W and L continuous models
TRADICA Input File Example:
Process and geometry data collection

1st step: determine geometry parameters from SEM cross section
509
HMT Window manual entry
Process specific parameters extraction using HMT/TRADICA:
Select available devices

Import data (automatic procedure)
Navigate through data if necessary
Run extraction
Go to the next extraction step
Note
Process specific parameters should be extracted from multi-geometry data in order to preserve
consistency,
one should avoid extracting model parameters on individual transistors and to fit scaling rules to model
parameters
The latter method (although proposed as standard method for PSP mos model) often provides usually poor
or/and unphysical results.
Extraction Example: Junction Capacitance
510
Example results: f T (30GHz Process)
Comparable results are routinely obtained for processes up to 200GHz.
Statistical Modeling: The SMACH Example

SMACH User Interface
Strategy for Statistical Modeling

Which parameters should be used as independent input variables?
In SMACH Technological Parameters (TPs) are used:
TPs are Base doping, Emitter width, Base thickness, Collector doping, etc.
One can find relationships between model parameters and TP
TPs are uncorrelated
Reduces measurement and extraction effort
But TPs statistical variations are typically unknown!
SMACH allows determining TPs (unknown) statistical variations from (known) PCM
measured variations
In order to do this, PCM are re-simulated using the scalable and statistical model
Optimization is used: target is RMS error between measured and simulated PCMs,
parameters are TPs statistical variations
511
This is done automatically within the SMACH statistical PLUGIN
Definition
Example of (ADS based) HiCUM scalable & statistical macro model
Statistical Modeling Procedure

Need to define a list of used PCMs:
bias conditions
Transistor size
Get back a proposed list of TPs related to these PCMs
Based on the previous step a list of PCMs Gaussian distributions appear

Measured and Simulated PCMs are compared in these plots
TPs deviations are the variables to be optimized
Once finished, TPs are known!
One can run full Monte-Carlo analysis to check correlation between PCMs
512
Check Correlation between PCMs
The following type of multi-plot is automatically generated:
check correlation between PCMs

compare measured and re-simulated PCMs based on full Monte-Carlo simulation
(here 100 samples but can go up to several thousands)
Statistical Modeling Procedure

The Statistical modeling procedure has the following advantages:
It's fast: the whole procedure takes about 30 minutes

It's reliable: Measured PCMs are "guaranted" to be reproduced by the model
it's physics based: model parameters and Technological Parameters (TPs) are linked
via physics based equations
Only PCM measurements are required (usually available since routinely done all along
the foundry process manufacturing)
The procedure is automatically adapted depending on which PCMs are available
Since the model is a scalable model, Statistical variations are also "natively"
geometry dependent
513
Learning the Angelov Model
The following sequence of plots gives an overview about the influence of the model
parameters and their fitting properties. It represents also a possible sequence for the
model parameter extraction.
IC-CAP File: Learning_the_Angelov_Model.mdl
The default parameter values for the plots below are:
Idsmod = 0 Igmod = 0 Capmod = 1 Selft = 1 Noimod = 1

Ipk0 = 0.1
Vpks = -0.2
P1 = 2.0
P2 = 0.0
P3 = 0.0
Alphar = 0.1
Alphas = 1.5
Lambda = 0.01
Lambda1 = 0.0
The parameter values specified above are the recommended 'starting values' of I.Angelov
Bias conditions:
vGS: -5V .. 1V
vDS: 0V .. 10V
IPK0 Current for maximum transconductance
VPKS Gate voltage for maximum transconductance
P1 Polynomial coefficient for channel current
514
LAMBDA Channel length modulation parameter
ALPHAR Saturation parameter
ALPHAS Saturation parameter
515
After all, optimize the above parameters altogether, on idvg and idvd simultaneously, and
include also the ohmic resistor parameters RG, RD and RS in the optimization.
516
UCB MOS Level-3 MODEL
The main application of MOS transistors is in integrated circuits and digital circuits. For
this purpose, they are used within the same circuit with different layout sizes. Therefore,
they have to be modeled including their geometry information, what is essentially the
Gate length and width. This means, MOS models and also the MOS3 model are scalable
and contain always the model parameters length L and width W.
Cross-section of a MOS transistor and the different geometry devices needed for modeling
For the parameter modeling, these measurements have to be performed:
transfer (transconductance) curve iD(vG,vB) for a constant and low value of vD,
output characteristics iD(vD, vG)
and for BSIM3 also the output characteristics RD(vD,vG)
This output characteristics is of special importance for MOS modeling in analog

applications. Hereby, the output resistance of a transistor stage is degrading the input
characteristics of the following one by composing a lowpass filter consisting of Rout and
Cin.
For detailed information about the model equations, see /Vladimirescou/ . Table 1 puts the
most important MOS3 model parameters together.
517
Parameter Description Unit Default Example
process parameters
TOX oxide thickness meter 100n 100n
extracted from the linear region
VTO extrapolated zero bias threshold voltage V 0 1
NSUB substrate doping concentration 1/cm 0 4.00E+15
UO surface mobility at low gate voltage cm/Vs 600 700
THETA mobility reduction parameter 1/V 0 0.1
extracted from the subthreshold region
NFS effective fast surface state density 1/cm 0 1.00E+10
ETA static feedback - 0 1
extracted from the saturation region
VMAX maximum drift velocity of carriers meter/s 0 5.00E+04
KAPPA saturation field factor - 0.2 0.5
short channel effect
XJ metalurgical junction depth meter 0 1u
LD lateral diffusion coefficient (eff.channel meter 0 0.8u
length)
narrow channel effect
DELTA width effect on threshold voltage - 0 1
WD channel width reduction (eff.channel meter 0 0
width)
parasitic elements
RS source ohmic resistance Ohm 0 1
RD drain ohmic resistance Ohm 0 1
gate capacitance
CGDO G-D overlap capacitance per meter F/meter 0 40p
channel width
CGSO G-S overlap capacitance per meter F/meter 0 40p
channel width
junction capacitances
large area, small perimeter (a square
test device)
CJ zero-bias bulk capac. per square meter F/m 0 2.00E-04
of junct.area
MJ bulk junction grading coefficient - 0.5 0.5
PB bulk junction potential V 0.8 0.87
large perimeter, small area (a meander
test device)
CJSW zero-bias sidewall capac. per meter of F/m 0 1n
perimeter
MJSW perimeter capacitance grading - 0.33 0.33
coefficient
Literature:
HP94452A UCB MOSFET Model and Parameter Extraction Manual
A.Vladimirescou, S.Liu, The Simulation of MOS Integrated Circuits Using Spice2,

Memorandum UCB/ERL M80/7, University of California, Berkeley
The following sequence of plots gives an overview about the influence of the model
parameters and their fitting properties. It represents also a possible sequence for the
model parameter extraction.
518
519
520
521
MOS Transistors for Dummies
or: The MOS Drain current equation on one page (after B.Hfflinger, IMS, Stuttgart)
We start with neglecting the influence of V DS on the electric field between Gate and
Source,i.e. V DS << V GS , and consider the charge in the Gate capacitor.
Now, let's extend this idea to also include the effect of V DS on V GS vs. the Gate length L
i.e. V DS > V GS.
This equation is valid until I ds gets into saturation.
522
Compare this result with the MOS Level 1, where is:
Let's now interpret these basic formulas a bit and learn some basics about gm and fT.
We commence with:
Derivation gives:
Inserting the charge velocity v = *E DS ~ 107cm/sec, gives
from what we learn that wide transistors exhibit more gm:
Considering now the cutoff or transit frequency of MOS transistors:
Like stated above, it is TAU = L / v with v ~ 107cm/sec, what gives
from what we learn that short transistors exhibit more fT:
This means that for high-frequency MOS transistors and high gain gm, we need short and
wide transistors. Practically, this is a finger-structured transistor layout !
As an example, assuming L = 0,1um, we get f T ~ 100GHz.
MOS Modeling for Dummies

In the MOS model Level 1, I DS is including the lateral underdiffusion LD:
or, simplified by neglecting the output characteristics slope in saturation, i.e. LAMBDA,
The following sketch depicts two I DS curves of two transistors with same W, but different
lengths L1 and L2.
523
It is obvious that the threshold voltage, V TH, can be estimated from the x-intersection of
this plot.
In order to extract the model parameters KP and LD, we consider the slope gm of I DS :
Reading the two slopes m1 and m2 from the above plot, and from the starting point
Solving this set of equation gives:
524
PSP
Contents
Background of the PSP Model Extraction Package (iccapmhb)
Global PSP Model (iccapmhb)
PSP Model Extraction Package (iccapmhb)
PSP Parameter Extraction Strategy (iccapmhb)
525
Background of the PSP Model
Extraction Package
Introduction to PSP
The PSP model is a compact MOSFET model intended for digital, analogue, and RF
design,
It has been jointly developed by Philips Research (now NXP) and The Pennsylvania
State University (the PSP co-author moved to Arizona State University).
The roots of PSP lie in both MOS Model 11 (Philips Research) and the SP model (Penn
State University).
In December 2005, the Compact Model Council (CMC) has elected PSP as the new
industrial standard model for compact MOSFET modeling.
Physical principles in PSP

PSP is a surface-potential based MOS Model, containing all relevant physical effects
(mobility reduction, velocity saturation, DIBL, gate current, lateral doping gradient
effects, STI stress, etc.) to model present-day and upcoming bulk CMOS technologies
in the nanometer regime. The source/drain junction model, c.q. the JUNCAP2 model,
is fully integrated in PSP.
PSP not only gives an accurate description of currents, charges, and their first order
derivatives (i.e. transconductance, conductance and capacitances), but also of the
higher order derivatives, resulting in an accurate description of electrical distortion
behavior. The latter is especially important for analog and RF circuit design.
The model furthermore gives an accurate description of the noise behavior of
MOSFETs. Finally, PSP has an option for simulation of non-quasi-static (NQS) effects.
Internal model structure of PSP

BSIM4 Simulator
Structure of the BSIM4 model
Calculation of the electrical behavior of the MOSFET in a complex set of equations which
contain both, device dimensions (L, W, NF, ..) and model parameters (TOXE, DVT0,
VTH0,..)
PSP Simulator
Structure of the PSP model
In a first step, a scaling of parameters is performed. From a so called Global Model which
contains both, device dimensions and model parameters, an effective parameter is
calculated which is called Local parameter.
The effective transistor properties are calculated from Local parameters only.
Types of parameter scaling
526
Parameter extraction principle for scalable models

Taking into account the structure of the PSP model with different types of scaling, we can
identify the following kinds of parameter extractions:
Parameters, which do not scale and which are identical in the local and in the global
model like the very fundamental parameters like TOXO.
Parameters not affecting the transistor" part in the PSP model (junction capacitance
or the diode).
Parameters which scale with the transistor dimensions (L, W). These parameters are
mostly describing the transistor" behavior and influence the core of the PSP model.
The following pages will highlight the extraction methodology for those devices in
detail.
Local extraction
The goal is to extract a set of local parameters from one device only without taking into
account the device dimensions:
VFB, BETN, NEFF, ...
A set of Local model parameters, which is valid for one dedicated transistor
Repeat this procedure for different devices with different Dimension
527
Local Parameter vs. Dimension

The local parameter values are plotted vs. device dimensions
Determine the scaling parameters from this curve according to the built in scaling
equation:
Global Scaling
With a more complex scaling equation, parameters are plotted vs.
L @wide W,
W @ wide L,
and L,W simultaneously
528
Example for a typical scaling equation:
Binning in PSP
The PSP model provides in addition a binning feature.
Instead of the global scaling, a scaling of local parameters inside defined bin
boundaries can be applied.
However, the difference to former MOS models with binning features is, that PSP
provides 3 types of scaling equations which are shown on the following page.
This approach allows to make the binning a little bit more physical compared to the
approach in the BSIMxx models.
Type 1:
Type 2:
Type 3:
Parameter Explanation Equivalent BSIM4 Expression

LE effective channel length L- Leff
WE effective channel width W Weff
LEN normalized channel length = 1e-6 -
WEN normalized channel width = 1e-6 -
Example extraction flow

Extraction Steps
The extraction procedure is clearly defined by the model definition itself.
Due to the different scaling equations, the stepping sequence (1..10) shown in the figure
below must be followed.
A few typical steps are described below:
Extraction of Local parameter sets of a similar group of devices.

Adjustment of Global model parameters to a group of extracted Local parameters.
1. Long-Wide
529
Determination of a basic set of Local parameters (NEFF, BETN,..)
Extraction of some Local parameters, which will be fixed for all other devices (VP,
THEMU, ..). These parameters are immediately transformed into appropriate global
parameters (VPO, THEMUO,..)
1. Long-Width Dependence
Determination of a basic set of Local parameters (NEFF, BETN,..)

Extraction of some Local parameters, which will describe narrow channel effects only
(MUE,CS)
1. Scaling of Width Dependence

In this step, some global parameters, which are only width dependent are already
determined from the appropriate local parameters.
In this case these are:
MUEO, MUEW from MUE from steps 1,2
CSO, CSW from CS from steps 1,2
Impact on further extraction steps

When the parameter set of the highlighted device is extracted, some local parameters
must be set and calculated from the global parameters before the extraction starts.
These are exactly those parameters, which have been extracted only in steps 1 and 2 and
which are describing narrow channel effects.
In addition, there is no way to simply "transfer" these parameters from a certain other
device, because such a device with an appropriate gate width does not exist.
Summary
We introduced the internal structure of the PSP model and highlighted the approach
530
in separating the model equations into a so called local and a global model.
Starting from this model structure the principle of the two step parameter extraction
method going from local to global models is demonstrated.
Finally, the example for the extraction flow shows how to do such a parameter
extraction step by step in detail taking into account the complex relation between
different parameter groups.
As a conclusion, this approach is implemented in the brand new PSP Extraction
Package in IC-CAP from Agilent Technologies.
PSP Extraction Flow
Requirements from PSP modeling teams

High level description of necessary functionality of a simultaneousadjustment of both,
global and local model parameters:
Show more than one group of global parameters simultaneously (e.g. DPHIB and
NEFF)
Display in addition scaling behavior of the global model for several devices (e.g. Vtlin
vs. L, Idlin vs. L, etc.)
Select one device and invoke the local parameter tuning/optimization for this device.
Show the change of the parameters in all the following diagram groups: - local device
behavior (I-V curves)- "measured" data of global parameters (Parameters vs. L, W)
Invoke the optimization/tuning of global parameters. Show the effect of changing a
parameter in the "simulated" part of the global parameters group and in the selected
diagrams for electrical scaling (e.g. Vtlin, Idsat, ..).
Plot Configuration
Select the parameters to display the scaling as well as some electrical scaling diagrams
(e.g. Vtlin(L) etc.) to be displayed.
Adjust Scaling
Select the parameter to be adjusted by tuning or optimization.
The electrical scaling diagrams can be updated simultaneously or manually.
Due to the use of ADS as a simulator, this is done in real time.
531
Switch to local extraction

Select a device, which is far off the scaling. Using the right mouse button, a menu can be
invoked which allows the switching to the Local Extraction mode.
Adjust local parameter

In addition, the I-V curves of a local device are shown
All available tuners / optimizers for this device can be selected.
After selecting such an optimizer / tuner, the appropriate diagrams will appear in the plot
area.
Simultaneous update of I-V curves of a local device and the local parameter in the scaling
diagrams while tuning a local parameter.
532
Documentation
The integrated documentation in IC-CAP describes the extraction steps in detail.
The different kinds of extracted local parameters depending on the dimensions of the
device as well as the following scaling rules are outlined in a easy to follow flow.
Technical Details
Simulators supported: ADS*), Spectre (others like HSPICE will be added depending
on the availability of PSP1020 in those simulators.
Supports the generation of scalable PSP1020 models as well as of the binned
PSP1021 model.
533
Global PSP Model
The IC-CAP Model File structure with its organization in DUTs and according Setups (Measurement,
Simulation and Extraction/Optimization environment) makes it an ideal tool for handling the PSP model.
The local model is hosted in the Circuit on DUT level, while the global model is located in the Circuit on
Model File level.
Scaling e.g. vs. L, W etc, is done by using the LSYNC sweep in any Setup.
534
PSP Model Extraction Package
Key Features
The IC-CAP PSP Modeling Package features modules for both, measurements and
extractions, of DC, CV and S-parameter behavior.
It has been developed based on the recommendations of the model developers (NXP,
Philips Research and Penn State University) as well as from modeling teams already
using the PSP model.
The extraction procedure was successfully tested on devices with a minimum feature
size down to the 45nm technology node.
PSP Extraction Flow
Basic Capacitance
The oxide capacitance as well as the overlap capacitance effects are derived from the well
known test structures.
For this purpose, the global model of PSP is invoked and the fitting is done on multiple
devices simultaneously.
This extraction must be performed prior to the generation of local models of single
transistor devices
Local Model Extraction

In general, the extraction flow control of the PSP Modeling Package can combine both,
local and global parameter extraction.
This is necessary due to the dependency of local model parameters on basic parameters
from Cox and Coverlap extractions.
535
Initial conditions can be specified for each local parameter extraction.

Complete start parameter sets as well as single parameters can be taken from
previous results.
Extractions of local parameters are defined according to the recommendations of the

model developers.
All
extractions
optimizations
tuners
can easily be customized with respect to the selection of parameters, optimization
algorithms, optimization region etc.
Scalable Parameters
Plot local parameters vs. gate length or gate width to extract and adjust global model
parameters.
The IC-CAP PSP package extraction flow provides all groups for global model parameter
536
fitting according to the categories of the different device dimensions.
Verification and Fine tuning of Global Model
It is very likely, that it is necessary to fine tune a global model after extracting the global
model parameters from local models.
This functionality is provided by the combined data display and the Plot Optimizer inside
the PSP Modeling Package.
The user can select a huge variety of characteristic diagrams and can apply own,
graphically defined optimizers/tuners to those diagrams.
Once such an optimizer is defined, it can be reused for later extractions.
Extract STI Parameters

The STI (Shallow Trench Isolation) stress parameters can only be extracted after having a
well fitted global model available.
This is done as a final step in the extraction flow.
Extensions to the PSP Model Extraction Package from

the Modeling Bench
537
Requirements from PSP modeling teams

High level description of necessary functionality of a simultaneous adjustment of both,
global and local model parameters:
Show more than one group of global parameters simultaneously (e.g. DPHIB and
NEFF)
Display in addition scaling behavior of the global model for several devices (e.g. Vtlin
vs. L, Idlin vs. L, etc.)
Select one device and invoke the local parameter tuning/optimization for this device.
Show the change of the parameters in all the following diagram groups: - local device
behavior (I-V curves)- "measured" data of global parameters (Parameters vs. L, W)
Invoke the optimization/tuning of global parameters. Show the effect of changing a
parameter in the "simulated" part of the global parameters group and in the selected
diagrams for electrical scaling (e.g. Vtlin, Idsat, ..).
Plot Configuration
Select the parameters to display the scaling as well as some electrical scaling diagrams
(e.g. Vtlin(L) etc.)
Adjust Scaling
Select the parameter to be adjusted by tuning or optimization.
The electrical scaling diagrams can be updated simultaneously or manually.
Due to the fast link to ADS as a simulator, this is done in real time.
538
Switch to local extraction

Select a device, which is far off the scaling.
Using the right mouse button, a menu can be invoked which allows the switching to the
Local Extraction mode.
Adjust local parameter

In addition, the I-V curves of a local device are shown.
All available tuners / optimizers for this device can be selected.
After selecting an optimizer / tuner, the appropriate diagrams will appear in the plot area.
Simultaneous update of I-V curves of a local device and the local parameter in the scaling
diagrams, while tuning a local parameter.
539
Documentation
The integrated documentation in IC-CAP describes the extraction steps in detail.
The different kinds of extracted local parameters depending on the dimensions of the
device as well as the following scaling rules are clearly outlined and can be easily followed.
PSP Modeling
Don't worry about the size and readability of a single plot. Using 2 or more FlatScreens,
the PSP Modeling Package can display all required information at once.
Technical Details
Simulators supported: ADS*), Spectre (others like HSPICE will be added depending
on the availability of PSP1020 in those simulators).
Supports the generation of scalable PSP1020 models as well as of the binned
PSP1021 model.
The PSP Modeling Package is delivered starting December 7, 2006 through an update
to IC-CAP2006 (available on the Agilent EEsof Knowledge Center)
For more information, please visit the Agilent website:
http://eesof.tm.agilent.com/products/iccap_psp.html
540
PSP Parameter Extraction Strategy
This information below gives an overview about the impact of the PS model parameters
and a proposed parameter extraction sequence.
CV Modeling
541
Modeling of junction capacitance
VBIRxxx: Built-in voltage at reference temp

PBOTxxx: Grading coefficient
CJORxx: Zero-bias capacitance per unit-of-area
Diode
542
Individual Transistor DC Modeling

id_vg
BETN: (0..70m..inf) Channel aspect ratio times zero-field mobility

MUE: (0..0.5..inf) Mobility reduction coefficient at nominal reference temperature TR
XCOR: (0..0..inf) Non-universality factor
gm
543
THEMU: (0..1.5..inf) Mobility reduction exponent at nominal reference temperature TR
id_vg Subthreshold
NEFF: (1e20..5E23..1e26) Effective substrate doping

DPHIB: (-inf..0..inf) Offset parameter for Surface Potential PHIB
CT: (0..0..inf) Interface states factor
id_vd
544
THESAT: (0..1..inf) Velocity saturation parameter at nominal reference temperature TR
AX: (2..3..inf) Linear/saturation transition factor
gds_vd
ALP: (0..10m..inf) channel-length-modulation (CLM) pre-factor

ALP1: (0..0..inf) CLM enhancement factor above threshold
ALP2: (0..0..inf) CLM enhancement factor below threshold
VP: (1e-10..50m..inf) CLM logarithm dependence factor
AX: (2..3..inf) Linear/saturation transition factor
PSP Parameter Extraction Flow for a single transistor (Large/Wide):
1. id_vg: NEFF, BETN, THEMU, MUE

2. id_vg substhreshold: NEFF, DPHIB, CT
3. id_vg: XCOR, CS, BETN, THEMU, MUE
4. fine-tuning: All above parameters on all id_vg
5. id_vd: THESAT, AX
6. gds: ALP, ALP1, ALP2, VP, AX
7. fine-tuning: All above parameters on id_vd, gds
Local <-> Global Parameters
For each geometry, local params are extracted

The geometry scaling of the local params is plotted, and the global params are fitted
to this curve.
Simulate devices with the final global model.
Finally, each device is fine tuned to simultaneously fit both, - the device behavior
(local params) as well as - the scaling equation from the global model.
While the modeling of the individual transistor (parameter mini-set) seems easy, the modeling of all
different geometries (mini-sets) is much more complex because of model-specific dependencies between
the miniset parameters.
Additionally, the global PSP model (parameter maxi-set) has to handle measurement outliers and mini-set
parameter outliers.
545
The Curtice Mesfet Model
An important model for GaAs-MESFETs is the Curtice model.
Walter R.Curtice developed it around 1980 with the RCA Labs in Princeton, New Jersey,
USA.
Fig.1 depicts the equivalent schematic, consisting of the Curtice model itself, as well as
additional enhancements for high frequency modeling using Lx and Rx, x = G,S,D.
Curtice model
The following table compiles the model parameters:
Common parameters for all Curtice model levels:

Parameter Description default unit
RD, RG, RS external Drain, Gate and Source resistance 1m Ohm
VBR reverse GD resp. GS breakdown voltage 0 V
RIN series resist.to CGS (models HF input impedance) 1m Ohm
CGSO Gate-Drain junction capacitance with zero bias 0 F
CGDO Gate-Drain junction capacitance with zero bias 0 F
VBI built-in voltage (same as PB in Statz model) 1 V
CGS interelectrode Gate-Source bias-independent cap 300f F
CGD interelectrode Gate-Drain bias-independent cap 30f F
CDS interelectrode Drain-Source bias-independent cap 300f F
LD, LG, LS external inductances 0 H
TAU internal time delay from drain to source 10p sec
A5 dependency of TAU versus Vds 1sec /V
Diode model parameters:

IS diode saturation current 10f A
N diode emission coefficient 1 V
XTI diode saturation current temp.coefficient 0 Ohm
EG diode energy gap 1.11 eV
Piecewise linear model parameters:

R1 approx. D-G breakdown resistance 0 Ohm
R1 resistance relating breakdown volt.to channel currents 0 Ohm
RF eff.value of forw.bias G-S resistance 0 Ohm
MODEL switch between level 1 and level 2 model 1 or 2
Level 1 parameters:
ALPHA coeff. of Vds in tanh function for quadr.model 0.8 1/V
BETA transcond.param. (same as BETA in Statz) 3m A/V
LAMBDA channel length modulation param for quadr.model 40m 1/V
VTO quadr.model gate threshold volt.(same as VTO in Statz) -6 V
Level 2 parameters:
546
GAMMA coeff. of Vds in tanh function for quadr.model .5 1/V
BETA transcond.param. (same as BETA in Statz) 100u A/V
A0 0-order coeff. of v1 in Ids cubic equation 83.6m A
A1 1-order coeff. of v1 in Ids cubic equation 15.28m A/V
A2 2-order coeff. of v1 in Ids cubic equation -1.293m A/V
A3 3-order coeff. of v1 in Ids cubic equation -233u A/V
RDSO Drain-Source leakage path resistance 1m Ohm
VDSO value of Vds at which A0 .. A3 are determined 6 V
VDSDC the Vds bias at which RDSO, CDG and CGS were measured 0.8 V
A special of this model is the fact that it behaves differently, depending on parameter
values, which act as model switches. This means, the input characteristic can be modeled
by using either diodes or simply a piece-wise linear function. Also, for the transfer curve,
id = f(vg), a so-called 'quadratic' (Level_1) or 'cubic' (Level_2) modeling can be selected.
The parameter BETA has a double meaning (see table above). Also, for AC modeling, the
user can select between different modeling features. Fig. 2 lists these modeling specials.
Overview about the different modeling possibilities of the Curtice model
DC Modeling
The two modeling possibilities for the input characteristics are given in fig.3.
Parameter determination is done following the common diode modeling techniques.
547
Modeling with diodes: this is true if RF=0, R1=0 and R2=0
Possibilities to model the DC semiconductor input characteristics
Modeling with piece-wise linear functions: this is true if RF, R1 and R2 <> 0
The transfer curve can either be modeled using Level_1 (quadratic model) or using
Level_2 (cubic model). Fig. 4 puts both possibilities together.
Level_1 (quadratic): Parameter MODEL = 1
The determination of the parameters for Level_1 is similarly to the Statz model: first the
modeling of the transfer curve with BETA , and then the output characteristics using the
parameters ALPHA (low vDS) and LAMBDA (saturated output characteristics, 'Early
effect'). For Level_2, the parameters A0 .. A3 can be obtained by deriving the transfer
curve several times versus V1, GAMMA fits the output characteristics for low vDS and
BETA models the slope in the saturation region.
Dynamic performance
Electronic charge is modeled by a constant charge between Drain and Source as well as a
delay time TAU:
Qds = CDS * Vds + TAU * Ids
TAU stands for the internal time delay (carrier lifetime). This delay time is modeled with a
constant value, if TAU is specified, i.e. non-zero. However, if TAU=0 and A5 is specified,
SPICE calculates the delay TAU using:
TAU = A5 * Vds
Also, with the modeling of the Drain-Gate or Source-Gate capacitances, the Curtice model
offers two possibilities. Either, when specifying CGS and CGD, what means they are non-
zero, the charges are modeled linearly with
Qgs = CGS * Vgc where Vgc is the part of Vgs across CGS.
Qgd = CGD * Vgd
Qds = CDS * Vds
Otherwise, if CGS or CGD are set to zero each, the known space charge formulation is
used:
with
548
For positive bias voltages above Vgx > FC*VBI is:
When interested only in the modeling around the operating point, the linear model of the
capacitances and thus the determination of the parameters CGS, CGD and CDS is
sufficient. This can be achieved by a S-parameter measurement at the operating point in
question. Otherwise, the modeling is done using CV measurements of the capacitors CGS
and CGD following the common CV parameter extraction rules (see chapter diode or
bipolar transistor).
To determine the external parasitic inductor values Lx and Rx, the transistor is measured
bias-overdriven (id = -ig/2). This means, the transistor is fully saturated and behaves like
a good SHORT. Fig. 5a visualizes the measurement considerations and fig. 5b the
parameter extraction.
Measurement technique to bias-overdrive the GaAs transistor. This allows the measurement of the series inductances
and resistances
how to determine the parasitic component values in bias-overdriven measurement mode
Finally, the remaining capacitors of the Curtice model are determined. This is done by de-
embedding the measured S-parameters from the now known series components (Z matrix
de-embedding). Afterwards, the result is converted into Y matrix parameters. For the
linear modeling of the capacitors with Qij = Cij * Vij , fig.6 gives the details.
549
TO conclude, fig. 7 shows the measurement and simulation result.
The modeling result: measurement (lines) and simulation (dotted)
Publications for the Curtice model

W.R.Curtice, GaAs MESFET Modeling and Nonlinear CAD, IEE Trans.on Microwave Theory
and Techniques, vol.36, no.2, Feb.1988
W.R.Curtice, A MESFET Model for use in the Design of GaAs Integrated Circuits, IEEE
Trans.Microwave Theory and Techniques, vol. MTT-28, pp.448-456, May 1980.
550
The Gummel-Poon model
Contents
AC Small Signal Modeling (iccapmhb)
Appendices (iccapmhb)
Gummel-Poon Bipolar Model (iccapmhb)
Gummel-Poon CV Modeling (iccapmhb)
Introduction to Gummel-Poon (iccapmhb)
Limitations of Gummen Poon Model (iccapmhb)
Modeling the Base Resistor (iccapmhb)
Modeling the Parameter XCJC (iccapmhb)
Modeling the Resistors (iccapmhb)
Modeling the Temperature Effects (iccapmhb)
Non Linear DC Modeling (iccapmhb)
Preface (iccapmhb)
Publications (iccapmhb)
Transit Time Modeling (iccapmhb)
551
AC Small Signal Modeling
Overview
We are now ready to consider the basics of modeling for frequencies higher than 100MHz.
It is assumed that the measurements have been made on the pure semiconductor device
without being affected by packaging parasitics, bond pads or other parasitics. If this is not
possible, de-embedding techniques have to be applied. This means to find the proper
semiconductor behavior out of the distorted measurement by de-embedding.
It should also be mentioned that the probe pins have to have an excellent frequency
performance within the transistor measurement frequency range. Once again the key to
meaningful AC measurements and thus modeling is a good network analyzer calibration
with excellent standards and a correctly defined calibration kit data in the network
analyzer.
If you need additional support for de-embedding, calibration and for better understanding
S-parameters, please refer to the additional toolkits for IC-CAP. Please contact the author.
Let us go first for the AC small signal equivalent schematic. It can be derived from fig.2b
of the introduction chapter as a linearization at each bias point of the transistor.
The following equations give the values of the internal elements in fig.AC-1.
They represent the linearized DC- and CV-equations at the DC operating point.
gB'E'
From equ.(B) in the introductory chapter we get from the derivative of iB versus vB'E':
where the second term can most often be neglected for operating points of iC above 1 uA.
g0 The output conductance is:
CB'E' or CPI
including the delay time effect modeled by TFF is given in equ.(P) and (R3) of the
introductory chapter for the particular operating point voltages.
As a first order estimation, CB'E' simplifies to
while
CB'C' or CMU (MU or stands for 'mutual'
simplifies because of vBC < 0 at forward operation to:
gm
The transconductance gm finally is using equ. (H)
552
Appendices
Linear Curve Fitting: Regression Analysis
IC-CAP File:
$ICCAP_ROOT/examples/demo_features/4extraction/basic_PEL_routines/1fit_lin.mdl
Let's assume we made 'N' measurements y i at the stimulating points x i. I.e. we obtained
the array (x i , y i). Subsequently, these measured values were plotted.
A curve Y ( x ) shall be fitted to this array of measured data points using least square
curve fitting technique. Referring to an individual measurement point, the fitting error is:
The fitting will be done by varying the coefficients of the fitting curve of equation above.
The minimum of the total error E depends on the values of these coefficients. This means,
we have to differentiate E partially versus the curve coefficients and to set the results to
zero. We obtain a system of equations, solve it, and get the values of the coefficients for a
best curve fit. This is known as regression analysis.
Note
Provided we have got an array of N measured data points of the form (x i , y i).
A linear curve with the equation:
y(x)=mx+b

Ei = [ m xi + b ] - yi
Using the least means square method yields:
We obtain (from previous equations) after a re-arrangement:
and from previous equations:
Multiplying by -N and xi and adding these two equations allows the elimination of the
coefficient 'b', and we can separate the slope 'm':
553
or
and from previous equations for the y-intersect 'b':
with 'm' according to previous equations:
With previous equations, we determined the values of the two coefficients of the linear
curve which fits best into the 'cloud' of measured data.
Finally, a curve fitting quality factor r 2 is defined. Its value ranges from (0 < r 2 < 1). The
closer it is to 1, the better is the fit of the linear curve.
Verifying the quality of extraction routines

Let's start with the first problem: assumed, we would know the parameters to-be-
extracted in advance, it would be easy to check the validity of the extraction routines!
Using IC-CAP, it is simple to perform such a check. The trick is to 'synthesize' quasi-
measured data out of a set of parameters and to apply then the extraction routines to
these data. This can be done as follows:
1. Define a measurement setup in IC-CAP, for which the extraction routines shall be
tested. Example: an output characteristic for an Early-voltage extraction.
2. Select a 'typical' set of parameters (no default values like 'zero' or 'infinite', but
instead real realistic values!)
3. Change the 'Output' data type to 'S' (simulated only). The array behind that output is
no longer one-dimensional, i.e. no measurement data any more, but only simulation
data.
4. Simulate this setups using these parameter values.
5. Change the 'Output' data type back to type 'B'.
IC-CAP doubles now the data field to measurement and simulation data. Thus the
simulated data of step 4. is now converted to measured data!
6. Reset the model parameters by clicking 'Reset to Defaults' and simulate the setup
using the default parameters.
7. Apply the extraction routine-under-test and check the quality of the extracted
parameters.
Provided we get the parameter values back within a good tolerance, we can be sure that
the extraction works correctly. If we now apply the extraction to real-world measured
data, we should obtain the right parameters. This is true if the measured data have the
same shape like the model equations! If not, we might have to choose another model or
go for subcircuit modeling. And this leads us to the second part of this paper:
Direct visual parameter extraction of BF, ISE and NE

or: Checking if the model can fit the measured data at all
Using IC-CAP for the extraction of model parameters offers a lot of flexibility in terms of
creating user-defined models and implementing the corresponding extraction routines.
But when developing a new extraction strategy, we may run into two major problems:
do the routines extract the parameters correctly?
and is the model able to fit the measured device at all?
This appendix proposes a method that allows us to
verify the quality of the extraction routines

check the fitting of our model to the measured data
and to perform the parameter extraction simultaneously.
This model-fit-check method is also called 'Direct Visual Parameter Extraction'
The point is to transform the measured data to a domain where the parameter itself can
be displayed against the measurement stimulus. As an example, we know that the x-
intersect of lines fitted to an output characteristic of a bipolar transistor should hit always
the same point, the Early voltage. If we apply an IC-CAP PEL (parameter extraction
language) program to calculate the x-intersect of a line that is fitted to two adjacent
measured points, and if we display the result of this operation versus the collector voltage
(first order sweep), we will obtain a plot of the 'equivalent' Early-voltages of adjacent
measurement points.
The advantage of using this method is that we can see clearly, if the model is able to fit
the measured data at all. We only have to check if there is a flat region in the transformed
data domain or not. If it is there, we can extract the parameter very simply by calculating
the mean value of the flat region. And we know at the same time, in which range the
parameter is dominant and can therefore be used for fine-tuning with the optimizer. If
there is no flat range, the model cannot fit the measured data. We could vary the
parameter as much as we like and would not achieve a fit of the simulated to the
measured data!
554
For the application of this method, we start with some basic equations that refer to figure
below which defines some of the basic equations for the direct visual parameter
extraction:
Assumed we have:
y = m*x + y0
where m: slope & y0: y-intersect
Then it is:
x0 = -b/m
and:
Some more useful formulas to calculate:
These equations are implemented to the model files of directory "visu_n_extr".
The following plots give some examples on how to apply this idea to the parameter
extraction of a bipolar transistor using the Gummel-Poon model. It should be mentioned
that this method can be applied to all the parameters of this model, as well as to other
models like Statz, Curtice, BSIM etc.
555
RBM
Next, the y-intersects of two adjacent data points in this transformed data plot are drawn
versus 1/iB again. From the curve above, we might expect a flat range, but in reality we
get a curve like the left one below: In our example, RBM ranges from 5 to 12 Ohm. In this
case, we might display the visualized data versus iB, and interpret the result similar to
figures in the rBB chapter in order to estimate not only RBM but also IRB and RB (see plot
on the right).
The figure above represents the visual extraction of RBM and to estimate IRB and RB
possibly.
VAF
Determine VAF out of the x-intersect of a line through two adjacent measurement points:
X=vc
Y=ic.M
i=1
WHILE i < SIZE(Y)-1
VAF[i]=ABS(X[i+1]*Y[i-1]-X[i-1]*Y[i+1])//(Y[i+1]-Y[i-1])
i = i + 1
END WHILE
556
Early voltage is calculated from the x-intersect of a line through two adjacent data points
in the above figures.
Let's study also the application of the "visual" method for the Gummel-Poon plot:
interpretation of the Gummel-Poon plot for parameter extraction:
NF can be obtained from the inverse derivative of the iC-Gummel-plot:
IS
!Determine IS out of the y-intersect of a line through two adjacent measurement points:
X=vb
Y=log10(ic.M)
i=1
WHILE i < SIZE(Y)-1
ISE[i]=(X[i+1]*Y[i-1]-X[i-1]*Y[i+1])//(X[i+1]-X[i-1])
i = i + 1
END WHILE
IS from the y-intersect of a line through two adjacent measurement points:
557
NE
The following figure shows NE as obtained from the inverse derivative of the iB-Gummel-
plot
ISE
Determine ISE out of the y-intersect of a line through two adjacent measurement points:
X=vb
Y=log10(ib.M)
i=1
WHILE i < SIZE(Y)-1
ISE[i]=(X[i+1]*Y[i-1]-X[i-1]*Y[i+1])//(X[i+1]-X[i-1])
i = i + 1
END WHILE
As can be expected from an inspection of fig_6, the measured data does not show a big
recombination effect on the ib(vbe) curve. This means that the parameters ISE and NE will
not contribute a lot to the curve fitting and may be difficult to extract. This is exactly the
meaning of the transformation results in above figures.
IKF
In order to be able to visualize the effect of IKF (see fig.6 in the introduction chapter), we
have to 'strip-off' the effect of RE on the Gummel-Poon iC curve:
! this extraction assumes that RE is already extracted properly. It eliminates the effect of RE on
the Gummel-Plot, such that IKF can be extracted from the 'knee' of the 'stripped-off' ic-Gummel-
plot!
!strip-off the effect of RE on the Gummel-plot
!iCmeas = IS*exp(vBEint/(vt*NF) = IS*exp((vBEext-(iC+iB)*RE-iB*RB)/(vt*NF))
! i.e. multiplying iCmeas by exp((vBEext-(iC+iB)*RE-iB*RB)/(vt*NF))
! will give iC without the influence of RE and RB !!
X =ABS(vb)
Yic=ABS(ic.M)
Yib=ABS(ib.M)
! calculate the stripped-off Gummel Plot
558
tmp = Yic*(exp((MAIN.RE*(Yic+Yib)+MAIN.RBM*Yib)//(MAIN.NF*VT)))
! calculate the ideal ic Gummel curve

! as a reference
tmp1= MAIN.IS*(exp(X//(MAIN.NF*VT)))
RETURN tmp+j*tmp1
Calculating IKF
X = ABS(vb)
Y = IMAG(strip_off_RE)*(REAL(strip_off_RE))^-1
Y1 = REAL(strip_off_RE)
i=SIZE(Y)-1
index = 0
WHILE i > 0
IF Y[i] < SQRT(2) THEN
index = i
i = 0
END IF
i = i - 1
END WHILE
PRINT index
! calculuate IKF out of iC(vBE) at that index
MAIN.IKF = Y1[index]
Calculating IKF from the ratio ideal_iC_curve / stripped-off iC:
BF
The idea is to strip-off the base charge effect (NqB) from the beta plot and then apply the
often cited "max. beta" extraction for BF:
!Calculation of = iC/iB
vbe = vb-ve
vbc = vb-vc
beta = ABS(ic) / ABS(ib)
! with some simplification is:
! beta ~ BF/nqb
! nqb = q1//2*(1+SQRT(1+4*q2))
! with q1 ~ (1-vbe//VAR-vbc/VAF))^-1
! q2 ~ IS//IKF*exp(vbe//(NF*VT))
! VT : temp.voltage (a model variable)
q1 = (1-vbe//MAIN.VAR-vbc//MAIN.VAF)^-1
q2 = MAIN.IS//MAIN.IKF*exp(vbe//(MAIN.NF*VT))
559
nqb=q1//2*(1+SQRT(1+4*q2))
BF=ic//ib*nqb
RETURN BF
This method of direct visual parameter extraction can be applied to all Gummel-Poon
parameters. For more examples see the IC-CAP files in directory "visu_n_extr".
Let us conclude with an example of a PEL program used for this method. It is the VAF
transform. The program performs either the data transformation (SWITCH=1) or performs
the parameter extraction (SWITCH=-1):
X=vc ! link to stimulus data

Y=ic.M ! link to measured data
Y = SMOOTH3(Y) ! smooth measured data
SWITCH = -SWITCH ! a model variable
LINPUT "data transform (1) or extraction(-1) ?",SWITCH,dummy
SWITCH = dummy
PRINT "calculating the x_intersect ..."
tmp = Y ! a lousy array declaration
i=1
WHILE i < SIZE(Y)-1
tmp[i]=ABS(X[i+1]*Y[i-1]-X[i-1]*Y[i+1])//(Y[i+1]-Y[i-1])
i = i + 1
END WHILE
tmp[0] = tmp[1] ! watch-out for the array bounds
tmp[SIZE(tmp)-1] = tmp[SIZE(tmp)-2]
IF SWITCH == -1 THEN ! sum-up all parameters within the box
i = 0
N = 0
result = 0
WHILE i < SIZE(X)
IF X[i] > X_LOW AND X[i] < X_HIGH THEN
result = result + tmp.M[i]
N = N+1
END IF
i = i+1
END WHILE
MAIN.VAF = result//N ! export parameter value
PRINT "MAIN.VAF = ",MAIN.VAF
ELSE
MENU_FUNC("Visu_va","Display Plot")
LINPUT "click a box and re-run this transform to extract VAF",dummy
RETURN tmp
END IF
Note
Included in your "bipolar toolkit" is a directory called "visu_n_extract" that contains
IC-CAP model files with suggestions on direct visual extraction for most of the Gummel-Poon parameters.
Literature:
HP-EESOF, Characterization Solutions Journal, Spring 1996.
Calculation of h21 of the Gummel-Poon Model

We start from the small signal equival.schematic of the following figure, using some small
simplifications (Collector is AC-wise shorted, since h21 = iC / iB for RL = 0):
The equations:
Input:
vB = iB rBB' + vB' _ iB = vB / rBB' - vB' / rBB'
Output:
0 = (iC - iB') / pCB'C + vB'
560
Internal 1:
vB' = (iC - iB' + iB) rB'E
using rB'E = rB'E // CB'E
Internal 2:
iB' = vB' / rB'E
This gives the matrix:
Solved for iC:
Solving for the 2nd column of the nominator and the 2nd column of the denominator:
Now solving for the 1st column of the nominator and the 3rd row of the denominator:
gives for h21:
Finally re-substituting
1 / rB'E = 1 / rB'E + pCB'E
yields:
or
what is depicted below:
561
Now we are ready to calculate the transit frequency fT:
from previous equations, follows for h21 = 1 :
or:
Simplicification
In order to keep things a little simpler for parameter extraction, previous equation is
modified a bit, neglecting the zero (at high frequencies) against the pole(low frequencies):
Calculating again the transit frequency for this simplified h21 yields:
since = iC / iB', we get the pretty simple form
Or solved for TFF:
what is the well-known formula for the TFF parameter extraction.
Modeling the output characteristics quasi-saturation

effect with Gummel-Poon
562
The Gummel-Poon model is known for its poor fitting capabilities in the quasi-saturation
range. However, an output characteristics of a bipolar transistor can be split into two
overlying Collector currents: a first one (MAIN transistor) responsible for the output slope
in saturation, and another one, added, for modeling the quasi-saturation.
In this case, the trick lies in the fact to use the same model parameter set for both
transistors. However, the two transistors are in parallel with a series Collector resistor for
the quasisaturation range. See the next slide for the schematic:
In this example: increasing RC_quasisat affects the slope in quasisaturation, while

Q2.area determines the transition to the non-saturated range of iC.
See the previous slide to better understand the effect of these 2 parameters.
Note
Of course, during the Gummel-Poon plot modeling,, this scenario has to be considered as well.
Therefore, it is smart to set the area of the quasisaturation transistor to a variable , e.g. 'area', ranging
from 10m to 1, and to set the area of the MAIN transistor to (1-area).
This way, the Gummel-Poon plot modeling is not affected outside the quasisaturation range !
563
Gummel-Poon Bipolar Model
Structure of this Manual
Introduction
Operating Modes of the Bipolar Transistor
The Equivalent Schematic and the Formulas of the SPICE Gummel-Poon Model
A Listing of the Gummel-Poon Parameters
A Quick Tutorial on the Gummel-Poon Parameter Extractions
Proposed Extraction Strategy
CV Modeling
Extraction of CJE, VJE, MJE, as well as CJC, VJC, MJC
Parasitic Resistor Mmodeling
Extraction of RE
Extraction of RC
Extraction of RBM
Nonlinear DC Modeling
Extraction of VAR and VAF
Extraction of IS and NF
Extraction of BF , ISE and NE
Extraction of IKF
Reverse Parameters NR, BR, ISC, NC and IKR
AC Small Signal Modeling, Parameter Extraction
Extraction of RB, IRB and RBM
Extraction of TF, ITF, and XTF
Extraction of VTF
Extraction of PTF
Extraction of TR
Modeling of XCJC
Temperature Effects
Model Limitations
Appendices
Linear Curve Fitting: Regression Analysis
About the Modeling Dilemma
Verifying the Quality of Extraction Routines
Direct Visual Parameter Extraction of BF , ISE and NE
Calculation of h21 of the Gummel-Poon Model
Publications
This product has been developed to meet the local demands of European IC-CAP users for
more technical background information on extraction techniques and for the availability of
extraction source code.
Published for the first time in 1990, it has been updated since then several times.
It is part of a series of supplementary modeling toolkits for the IC-CAP users. These
products feature source code and detailed technical description of the extraction routines.
Please contact the author for further information.
The author would like to thank the many users for valuable inputs, and is hoping for
fruitful discussions also in the future.
About This Manual

This manual is intended to explain the basics of modeling a bipolar transistor using the
Gummel-Poon model as it is implemented in the simulation program SPICE of the
University of California Berkeley (UCB) /see publication list/.
It is part of the Gummel-Poon Bipolar Model Parameter Extraction Toolkit.
This toolkit includes the IC-CAP model file

GP_CLASSIC_NPN.mdl the MASTER model file
which is described in this manual
and featuring the data management features of IC-CAP 5.x, i.e. separating measurements
from extractions:
NPN_MEAS_MASTER.mdl a master file for measurement

GP_EXTRACT_NPN.mdl a master file for modeling
As well as many other IC-CAP model files covering topics like:
Model parameter extraction using the tuner feature

Direct visual parameter extractions
Alternate modeling methods for DC- CV- and RF-parameters.
Bipolar transistor modeling including the parasitic transistor.
564
After you have become familiar with the modeling procedure itself, i.e. file GP_CLASSIC_NPN.mdl, you are
encouraged to split the modeling into 2 parts: separate measurements and separate extraction strategy.
In this case, all measurements are performed using the file NPN_MEAS_MASTER.mdl. Then, the data are
exported into IC-CAP mdm files (ASCII files) and imported into the master extraction file
GP_EXTRACT_NPN.mdl for extraction. This method allows to improve continuously the extraction strategy
file, independent of the measurement data.
NOTE:
This manual and the underlying IC-CAP model file GP_CLASSIC_NPN.mdl are intended to explain the
basics of the Gummel-Poon modeling. Therefore, it covers the classical Gummel-Poon model without
enhancements for also modeling the parasitic transistor.
However, as stated above, such model files are included in the file sets of this toolkit. Please see the
README macros in these IC-CAP model files for more details.
You are also invited to get in contact with the author for assistance with such modeling problems.
The IC-CAP model file "GP_CLASSIC_NPN.mdl" features

The extractions are written using PEL (parameter extraction language) and are open to
the user. They can be easily modified to meet specific user needs.
Subcircuit model description, open for user enhancements (HF modeling, parasitic pnp
etc.).
All transistor pins are connected to SMUs for flexible measurements
The transistor output characteristic and S-parameter measurements use a Base current
stimulus rather than a Base-Emitter voltage in order to avoid 1st order thermal effects
being visible. However, self-heating might be present and affect the Gummel plots in the
ohmic range.
See also the file GP_MEAS_MASTER.mdl
Organization of the chapters in this manual

There are 5 main chapters, which explain how to determine the model parameters from
CV (capacitance versus voltage), then parasitic ohmic resistors, and DC, to finally high
frequency measurements using network analyzers.
More chapters cover side aspects of bipolar modeling.
The individual chapters follow always this scheme:
Explanation of the parameter-dependent measurement setup

Explanation of the mathematical basics for the parameter extraction
Explanation of the parameter extraction
Explanation about how to use the IC-CAP file.
565
Gummel-Poon CV Modeling
Since the CV parameters are like for most bipolar models, independent of the other model
parameters, they are usually extracted first. We follow this idea and begin with the CV
modeling, followed then by the parasitic resistor modeling and the non-linear DC curves.
Finally, the S-parameter measurements are modeled.
The Gummel-Poon Capacitor Equations

Provided that: vBE = vB'E' and vBC = vB'C', the capacitors in the Gummel-Poon model
given in the introduction chapter with equations (O)... (S) are:
with equation (H) from the introduction chapter.
again with equation (H1) from the introduction chapter and additionally with,
and the ideal Collector current if from (C).

CSBi models the space charge and CDBi the diffusion capacitance between Base and
Emitter or Base and Collector respectively. vBE and vBC are the stimulating voltages.
Note
For CBE, and its diffusion capacitor part CDBE, the exact definition is (private communication with D.Celi,
ST Crolles, France)
where TFac is the small signal transit time which is different from TFF. However, in this
toolkit, we keep the simplified equation of CV-2 and a corresponding raw parameter
extraction, and apply a quick optimizer run to get the true TF parameter.
Modeling the Space Charge Capacitors

Extraction of CJE, VJE, MJE, (CJC, VJC, MJC and CJS, VJS, MJS is the same)
CJE B-E zero-bias deplet.capacitance

VJE B-E built-in potential
MJE B-E junction exponential factor
These parameters model the Base-Emitter and the Base-Collector space charge
capacitance, i.e. the first term in (CV-1) and (CV-2). The second terms with the TFF and
TR parameters will be modeled later by S-parameter mesurements. For the measurement
of the Base-Emitter capacitance, the Collector is left open while the Emitter is open during
the measurement of the Base-Collector capacitance. In both cases, the modeling formula
is the same. Therefore this chapter covers only the modeling of the Base-Emitter
capacitance.
566
Measuring and modeling the Base-Emitter

capacitance.
Note on the Influence of the Remaining Capacitances of the Open
Pin:
as one of the transistor pins is left open, the measurements of CSBC and CSBE are always
an overlay of the other capacitances. CSBi (i = E,C ) and, when measuring packaged
devices, the small parasitic package capacitors CBC , CBE.and CCE.
The total measured capacitance is therefore CSBi in parallel with the parasitic ones. This
means that the measurement results are always too big. When using a capacitance meter
like the Agilent4284, that eliminates by its measurement principle parasitic capacitances
to ground, this effect can be avoided by applying an AC short to the open transistor pin
versus ground (big capacitor).
The Equation
The behavior of the space charge capacitor is given by equations (CV-4a) and (CV-4b):
Determination of the CV Parameters

For simplicity, we only use the measurement data from the negative bias. The logarithmic
conversion of (CV-4a) yields:
This equation can be interpreted as a linear function according to the ideas of linear
regression analysis:
567
Linear regression means to fit a line to given measurement points. Therfore, the three
main equations of a linear regression are b=f(xi,yi) and m=f(xi,yi), together with a fitting
quality factor r=f(xi,yi,m,b). For a good fit, r~0.9...0.9999. See also the appendix.
How to Proceed
the measured values of CSBC are logarithmically converted according to (CV-6a).
Following (CV-6d), the stimuli data of the forcing voltage vBE are nonlinearily converted
too. This is done using a starting value for the unknown parameter VJE (e.g. 0,2V). These
two arrays are now introduced into the regression equations (see appendix) as
corresponding yi- resp. xi-values. A linear curve is fitted to this transformed 'cloud' of
stimulating and measured data. Thus we get the y-intersect b(VJE) and the slope m(VJE)
for the actual value of VJE. In the next step, this procedure is repeated with an
incremented VJE, and we get another pair of m(VJE) and b(VJEC). But now the regression
coefficient r 2 will be different from the earlier one. i.e. depending on the actual value of
VJE, the regression line fits better or worse the transformed data 'cloud'. Once the best
regression coefficient is found, the iteration loop is exited and we finally get VJE_opt as
well as the corresponding b(VJE_opt) and m(VJE_opt).
Thus we get from (CV-6c):
and from (CV-6b):
Validity of this extraction: The parameter extraction for the space charge capacitor is valid
only for stimulus voltages vBE below FC * VJE, FC_default = 0,5.
What to do in IC-CAP
Since this is our first parameter extraction step, we first reset all parameter values to
default, see IC-CAP Window: 'Model Parameters'
Otherwise, we might end up with a mix of parameter values obtained during our last
transistor modeling and today. Therefore, execute Macro 'INITIALIZE' for this task.
Then, for our intended manual parameter extraction, make sure to set the Model Variables
MACRO_CONTROLLED=0
HELP_ENABLED=1
Extract the CV Parameters

open setup "/gp_classic_npn/cv/cbe_bhi"
( means modeling of CBE, with Base contact at high voltage pin),
import the data from NPN_MEAS_MASTER_demodata_PELdep.mdl
(located in demo_features\3_MEAS_ORGANIZE_n_VERIFY_DATA\0_MASTER_FILES)
click a box into plot "cbe" (capacitance vs. voltage) to select the measurement data used
later for extractions. Click 'Copy to Variables' under 'Options' in that plot. This will cause
IC-CAP to save the box corners in the 'cbe_bhi' Setup Variables
X_LOW, X_HIGH, Y_LOW, Y_HIGH
perform transform "br_CJE_VJE_MJE" (box regression). This transform applies a data
transformation and regression analysis to the data inside the box.
Then simulate with the extracted parameter values, using simulation or the substitute
transform calc_cv. Do the same for the capacitor CBC in setup 'cbc_bhi'.
{info:try also macro 'extract_n_opt_CV', or execute Transform 'OPEN_MDLG_GUI'
Some comments on CV-modeling

In practice there is always an overlay of this capacitance with some parasitic ones, e.g.
package or pad capacitances. If they are not known and therefore cannot be de-
embedded (calculated out of the measured data), the extracted CV parameter values may
have no physical meaning. This may happen especially to VJC and MJC.
If there are resolution problems with fF-capacitances and CV meters, a network analyzer
can be used instead of the CV meter as well. In this case, the Base is biased and Emitter
and Collector are grounded. The measured S-parameters are deembedded, converted to Y
parameters and the CV traces can be calculated out of their imaginary parts.
568
Introduction
Contents
Operating modes of the bipolar transistor
The Gummel-Poon equivalent schematic
The Gummel-Poon model equations
List of the SPICE Gummel-Poon parameters
A quick tutorial on the Gummel-Poon parameters
Proposed global extraction and optimization strategy
This manual describes the modeling of a bipolar transistor using the Gummel-Poon model
as implemented in the simulator SPICE. It should be mentioned that the Gummel-Poon
model itself covers only the internal part of a real-transistor. Therefore, on-wafer
parasitics like a parasitic pnp transistor are not covered. Also, packaging parasitics and
other non-ideal effects are not part of the model. However, they can be added by using a
sub-circuit rather than just the stand-alone model.
Please check the example files included in the file directory of this toolkit for examples.
Parasitic effects is specially important for network analyzer (NWA) measurements. The
modeling procedures presented in this manual refer to already de-embedded
measurements.
For on-wafer measurements, test probes that allow NWA calibrations down to the chip
(like Cascade or Picoprobe probes) are commonly used. De-embedding means here to
eliminate on-wafer parasitics, which are due to the test pads (OPEN dummy) and the lines
from the test pads to the transistor itself (SHORT dummy). This is done by subtracting the
Y matrix of the OPEN from the total measurement, followed if required by the subtraction
of the Z matrix of the SHORT. It should be mentioned that in this case the SHORT itself
has to be de-embedded first from the OPEN parasitics!
For packaged devices, we need to use a test fixture. In this case, the NWA has to be
calibrated down to the ends of its cables using the calibration standards (SOLT) of the
actual connector type. As a next step, the test fixture has to be modeled (OPEN, SHORT,
THRU). Finally, the DUT (device under test) is inserted into the test fixture and measured.
The now known test fixture parasitics can be de-embedded and the extraction techniques
of this manual can be applied to the down-stripped inner device. A file including such a
procedure is included in the toolkit filesets. See the example
more_files/packaged_xtor_in_testfixture.mdl
Please contact the author if you wish more info or help on de-embedding.
Operating Modes of the Bipolar Transistor

There are four operating modes of a bipolar transistor as illustrated in figure 1. The
saturation region, for example, the region vCE<0.3V in the DC output characteristics, is
described by the ohmic resistors. The DC and AC extraction procedures that are proposed
in this manual cover mainly the forward region. Since the model is symmetrical, the
reverse parameters can be extracted following the same ideas, but applied to the reverse
measurements.
Note
In the saturation range, BE and BC layers are 'overcharged'.
Operating Modes of a Bipolar npn Transistor
The Gummel-Poon Equivalent Schematic

The figure, "Gummel-Poon large signal schematic of the bipolar transistor" shows the
large signal schematic of the Gummel-Poon model. It represents the physical transistor: a
569
current-controlled output current sink, and two diode structures including their capacitors.
This structure represents pretty much the physical situation of a bipolar transistor, see
"physical situation for a bipolar transistor, neglecting the parasitic pnp transistor".
physical situation for a bipolar transistor, neglecting the parasitic pnp transistor
Gummel-Poon large signal schematic of the bipolar transistor
From figure "Gummel-Poon large signal schematic of the bipolar transistor" the small
signal schematic for high frequency simulations can be derived. This means, for a given
operating point, the DC currents are calculated and the model is linearized in this point
(fig.AC small signal schematic of the bipolar transistor). Such a schematic is used later for
SPICE S-parameter simulations.
It should be noted that the schematic after figure "AC small signal schematic of the bipolar
transistor" is a pure linear model. It cannot be used to predict non-linear high-frequency
behavior of the transistor. For this case, RF simulators like ADS perform high-frequency
simulations using the large signal model (harmonic balance simulations).
AC small signal schematic of the bipolar transistor
570
Note
XCJC effect neglected.
In order to make the presentations of the schematics complete, fig."subcircuit schematic

when including the parasitic pnp" depicts the subciruit used for modeling a npn transistor
including the parasitic pnp. As said above, IC-CAP files for this type of modeling are
included in the file sets of this toolkit. However, the description of this manual does not
cover this. See the macros in the model files instead.
Sub-circuit schematic when including the parasitic pnp
The Gummel-Poon Model Equations

For the reader's convenience all the Gummel-Poon equations are presented at a glance. In
order to make them better understandable, we assume no voltage drops at RB, RE and
RC, i.e. vB'E'=vBE and vB'C'=vBC.
TEMPERATURE VOLTAGE
Base Current
571
Collector Current
Base Resistor
Space Charge and Diffusion Capacitors
572
List of the SPICE Gummel-Poon Parameters
A quick Tutorial on the Gummel-Poon Parameters

Although it is recommended to go through the individual extraction steps of the
corresponding sections of this manual, this chapter puts together the graphical
equivalents of the parameter extraction techniques.
573
574
Proposed Global Extraction and Optimization

Strategy
First reset the model parameters to default (Window Model Parameters). This will firstly
get rid of old parameter values which belong to the last modeling and not to our actual,
current one, and secondly, the default parameters are those which reduce the complexity
of a model completely. For example, VAF=1000 means: no Early effect, IKF=1000 no
knee current, all resistors Rx=1m means no ohmic effects and so on. During the
extraction process, we get more and more parameter values, and thus, the model
becomes more and more complex and accurate.
CV
Extract the CV parameters CJx, MJx and VJx for the BE- and BC-capacitance, optional
also for CCS.
optimize the CV parameters.
Ohmic Parasitics
Extract the parasitic resistors RE, RBM and RC from flyback-measurements.
or:
Extract them from overdriven S-parameter measurements: high current at Base, half
the current out of the Collector, Emitter grounded, frequency swept.
575
DC
Extract VAR and VAF from the output characteristics.
Extract IS, NF, ISE, NE and BF from the forward Gummel-Poon plots.
Optimize the Gummel-Poon plot for IS, NF, ISE and NE, well below ohmic effects
show up.
Extract IKF from the -curve.
Optimize RE in the upper region of the Gummel-Poon plots (iB and iC).
Optimize BF and IKF in the -curve at high bias.
Fine-optimize VAR, BR, VAF and BF in the output characteristics setup.
Fine-tune all DC parameters in all DC setups.
S-parameters:
De-embed the measurement data.

Extract the base resistor parameters RB, IRB and RBM from S11 measurements with
swept frequency and base current as a secondary sweep.
Transform S- to H-parameters and get a frequency f-20dB from the -20dB/decade of
h21.
Measure again S-parameters, but now with the constant frequency f-20dB and swept
iB and swept vCE and extract TF, XTF and ITF, as well as VTF.
Optimize S-parameter fitting of TF, XTF, ITF (lowest vCE).
Optimize the S-parameter fitting of VTF (all vCE).
Go back to the rBB' setup and optimize the S-parameter fitting of the RB, IRB and
RBM.
Then, again in the rBB' setup, optimize the TF, XTF, ITF and VTF parameters.
Finally,
Re-simulate all setups and check the fitting quality in the verify setups.
If required, perform optimizer fine-tuning.
Macro 'extract_n_opt_ALL' in IC-CAP file gp_classic_npn.mdl contains an example for such

a modeling strategy based on the measurement data included in the file.
This strategy may vary a bit depending on the actual data. In this case, simply modify the
macro to meet your local requirements.
Note
A smart way of defining or verifying the most appropriate extraction strategy is to synthesize quasi-
measured data from simulation results, and to check the extraction routines on these data. This means to
simulate all setups using a given parameter set, to transform these simulated data into measured ones
and to try to get the (known) parameters back again.
In this way you are sure that your extraction strategy works well for a perfect Gummel-Poon transistor. If
you have afterwards problems during the curve fitting, you might consider that your physical device under
test may not be so well represented by the Gummel-Poon model!
To 'synthesize' such pseudo-measured data in IC-CAP, make sure the parameter values in the IC-CAP
parameter list are all set to typical values that you will expect later for your parameter extraction, perform
a simulation for every setup in your model file, change the setup output data type to 'S', hit <RETURN>,
change it back to 'B' again and hit again <RETURN>. Now you have identical data in both measured and
simulated arrays. Then reset the parameter values to default and try your extraction strategy.
Last but not the least, before performing your measurements, i.e. before defining the
measurement ranges, contact tour design engineer colleagues and ask them about the
specific operating range.
As a general rule, modeling should be done in those regions where the transistor will be
operated later.
576
Limitations of the Gummel-Poon Model
Ohmic Effects
The Collector and Emitter resistance parameters are constant and not functions of current
or voltage. They have no temperature coefficients.
Forward DC Modeling
The parameter IKF models the begin of the decrease in beta. As a limitation of the model
the slope of above the knee current IKF has a fixed value of "-1" on a log-log scale.
However, this is most often overlaid by RE.
The modeling of the saturated region in the output characteristics (VCE < 0.5V) lacks of
specific parameters. Therefore the model cannot cover modern transistors in this range
(quasi-saturation).
No reverse breakdown effects are included in both Base-Collector and Base-Emitter diode.
Reverse DC Modeling
The reverse DC modeling suffers from a separate parameter IS. Thus NR sometimes has
to be mis-used for better fitting the reverse iE versus vBC plot.
Like in the forward region, the slope of above the knee current IKR has a fixed value of
"-1" and also the output characteristics saturation range is modeled inflexible.
AC Modeling
The TF modeling, especially versus vCE, is not physical and often not accurate
The TR parameter is not a function of current or voltage like TFF.
Temperature Modeling
The TNOM value of VJE, VJC and VJS must be greater than 0,4V to insure convergence for
temperature analysis up to 200'C.
Applications in Integrated Circuits

There is no parasitic transistor included in the model.
Conclusion
Disregarding these limitations, the Gummel-Poon model is a good compromise between
accurate modeling and a limited amount of parameters. It is still very useful especially
when enhancing it with external parasitics like inductors, parasitic diodes or lateral pnp
transistors.
577
Modeling the Base Resistor rBB'
Extraction of RB, IRB and RBM
RB zero bias Base resistance
IRB curr. at medium Base resistance
RBM min.Base resistance at hi current
It is assumed that XCJC = 1, and that is the DC current amplification.
This chapter explains how to model the Base resistor from S11 data.
It is organized like this:
Derivation of the small signal schematic for the parameter extraction

Short introduction into the basics of the Smith chart
Discussing the expected frequency dependence of rBE, considering rBB' constant
Enhancing the schematic for the bias-dependent omic resistor rBB'
As an approximation to keep the equations simpler, we further assume: vBE = vB'E' and
vBC = vB'C'. Simulations and optimizer runs after the parameter estimation will eliminate
this simplification.
The measurement setup for the rBB' characterization is given below in the following figure
(RBB-1):
To begin with, we refer to fig.AC-1 from the previous chapter. We simplify it to cover
mainly the imput impedance. This leads to the schematic of fig.RBB-2. This figure explains
the two cases: frequency -> 0 Hz and frequency -> Hz
In order to evaluate the schematic and the device parameters of fig.RBB-2, we have to
consider the measured S11 data. This is best done by displaying them in a Smith chart.
As a reminder, a Smith chart transforms the right side of the complex resistor plane R into
the area of a circle of radius '1' using the transform
Therefore, we can use S11 instead of H11 for the RIN modeling as well and our
578
measurement result should look like figure RBB-4 .
Note: In order to get familiar with the problem, we consider first the hypothetical case that rBB' is no
function of bias. In other words, the Base resistor is considered as a constant, ohmic resistor RBB'.
Ideally, RIN should look like a circle.

The starting point at DC is RBB'+1/gB'E'+RE(1+). For higher frequencies, CB'E' will act
more and more like a short and eliminate the influence of resistor 1/gB'E'. For infinite
frequencies, RIN should hit the x-axis at RBB' + RE (effects of CB'C' and RC omitted!).
Now the Base-Emitter capacitance has completely shorted 1/gB'E' and thus the
transconductance gm became 0 as well. This means that the transistor has no beta any
more.
As RE is known from DC measurements, the value of RBB' can be estimated quite

accurately.
This method is advantageous because the estimation of the Base resistor is affected only
by the parameter RE. Moreover, there is mostly RBB' >> RE , so that the influence of a
uncertain value of RE is minimized using this method.
So far we considered rBB' to be simply ohmic, i.e.constant. In reality, rBB' is modeled

more complexly. One separate resistor from the outer Base contact to the inner Base
contact (ohmic RBM) and a bias-dependent part from the inner Base contact to inside the
inner Base. This means, the higher iB, the more iC is extending its flow area closer to the
internal Base contact due to current crowding. This means, we expect a lower Base
resistor for higher bias. The sketch below depicts that. The Gummel-Poon model combines
both resistors into a single, bias dependent one.
Now, overlying this DC bias dependency with the frequency dependence from above, we
end up with S11 curves like sketched in fig.RBB-6.
Real-world measurement curves will look like these curves at low frequencies only. This is
579
due to the overlay of more second order effects. In order to separate rBB' with the
proposed method, we must fit circles to the S11 curves at low frequencies and then
calculate the x-intersect from an extrapolation of the circle for infinite frequency, which is
then assumed to be equal to rBB'+RE. This is shown in fig.RBB-7.
Notes on some limitations of this extraction strategy

Because of the influence of 1/gB'E' on the range where the circle must be fitted to, iB
should be as high as possible to not dominate the r BB' - effect by 1/gB'E' . Also, to keep
the rBB' influence dominant over RE (1+), iB should also be as high as possible, so that
iC > IKF, and therefore , is as small as possible.
Unfortunately, the rBB' -measurement could now be dominated by thermal effects.
Moreover, this range of iB typically is also not the operating one. This contradicts general
rule to always concentrate on meaningful measurements close to the operating range for
good parameter extractions.
Finally, the trace of Fig.RBB-7 is often overlaid by the parameters TF, ITF, XTF and VTF,
which will be extracted next. Therefore, an optimization (of the S11 parameters) of this
setup should only be applied after the fitting of these transit time parameters.
The extraction strategy

Circles must be fitted to the low-frequency sections of interest. They are centered to the
x-axis. The suitable circle formula is:
This again can be considered as a linear form (!) with
This means: The measured data xi and yi are introduced into equ.(RBB-3a). Next the ylin
are plotted versus the xlin and a straight line regression is applied. From the slope
m, using (RBB-3c), we get:
and from the y-intersect b using (RBB-3b):
Finally the left circle intersection with the x-axis (for the frequency -> infinite) for our
rBB'-extraction is:
After all these pre-considerations, we are now able to generate the trace of RBB out of the
measured S-parameters. This means we are now ready to consider the formula for RBB in
the Gummel-Poon model.
Equation
The nonlinear Base resistor is described in the Gummel-Poon model as:
Fig.RBB-8 shows the plot of this equation:
580
This means: after we got rBB' from the measurement, we now have to fit the model curve
from fig.RBB-8 to the measured data of fig.RBB-7 (after subtraction of RE).
Procedure
When iB -> 0 then z -> 0 and therefore
We get from (53) solved for RB:

RB = rBB' [ iB -> 0 ] , zero bias Base resistance
When iB -> infinite then z -> PI/2 and
what gives from equation (RBB-4) solved for RBM:

RBM = rBB' [ iB -> infinite ] , min.Base resistance at hi current
Finally when iB = IRB then z = 1,21. Thus
That's why
IRB is the current where the Base resistor is half its max.value
RBM + (RB - RBM) / 2
After importing the de-embedded data, the modeling steps are:
Open setup "/gp_classic_npn/nwa_extr/rbb",

Import the data from the .mdm file,
Simulate the setup with the so far determined DC and CV parameters
Execute transform 'initialize'
Perform transform "calc_RBB" to convert the measured S-parameters to rBB'
(set model variable MACRO_CONTROLLED=0 in order to obtain the
Check of the upper frequency limit for the rBB extraction)
Check the plot "rbb_vbe" (rbb versus vbe)
Perform transform "e_RB_IRB_RBM"
Simulate with the extracted parameters.
Optimize after you are finished with the TFF parameter extractions.
(see proposed optimization sequence given there).
See also transform READ_ME
NOTE: As you might experience, it can be quite complex to obtain a reasonable S11 plot from which a rBB'
curve like that one in fig.RBB-8 can be derived. If despite all of these efforts the transformed measured
data do not match the curve of fig.RBB-8, set RB=RBM and model the Base resistor bias independent.
Note: avoid thermal self-heating (esp.when measuring packaged devices). This will show up if the fitting
of the forward Gummel plot of iB for high vBE becomes worse when the fitting of the S11 plot is improving
during fine-tuning of RE and RB. If this occurs, reduce the bias for both the rBB and forward Gummel
setup. If you need these high bias values, consider using pulsed measurements (DC bias pulse width
around 1us).
581
Modeling the Parameter XCJC
This parameter distributes the CCB junction capacitance between the inner and the outer
Base contact. Its default value is XCJC=1, meaning the CCB capacitance is tied completely
to the inner Base. For XCJC=0, the capacitance is between the outer Base and the
Collector.
This parameter is difficult to determine from CV measurements. However, if the geometry

of the device is known, it can be calculated pretty easily after:
XCJC = 1 - AE/AB,
where AE is the Emitter area, and AB the total area of the Base, including the Emitter
area.
Usually, this parameter is fine-tuned after the other HF parameters have been
determined.
The following S-parameter figures show the effect of XCJC for device modeling.
Generally speaking, if S12 becomes 'big' for high frequency, it is either Re or XCJC!
Note
fmax may also be used to model the effect of XCJC, as depicted in the figure below.Although there is no
effect of XCJC on ft, fmax is heavily affected.
A final remark on the S-Parameter Extraction and

Optimization Strategy
When you run into problems when fitting both, the transformed rBB' and TF curves, you
should try to optimize what you have measured, i.e. S-parameters rather than the rBB' or
the ft and TF plots. This is the real world and the fitting there might be more important
than the fitting of the rBB' plot with all its limitations (extrapolated S11 at infinite
frequency) or the fT plot (again extrapolated from H21 from S-to-H parameter
transformation).
582
Of course, the best modeling result is a good fit in all domains, the S-parameters and the
transformed rBB' and TFF curves.
583
Modeling the Resistors
Contents
Extraction of RE
Extraction of RC
Extraction of RBM from DC measurements
An alternate method to calculate the ohmic parasitic resistors from s-parameter

measurements.
The methods given below are considered as standard extractions. But the parameter
values are pretty often merely a 'first guess'. Also, the other model parameters are still
not yet knwon. Therefore, no simulation or optimization is performed in the setups of DUT
prdc in file gp_classic_npn.mdl
Instead, these parasitic resistor parameters are fine tuned in the setups dc/fgummel and
dc/rgummel. In the Gummel plots, they are tuned in order to fit the ohmic regions: RE in
the forward Gummel plot (iC and iB vs vBE) and RC in the reverse plot (iE and iB vs vBC).
Modeling the Emitter Resistor

Extraction of RE
Extracting the parameters:

The ohmic emitter resistor is physically located between the internal Emitter E' and the
external Emitter pin E. When we apply a Base current and have the Emitter pin grounded,
we get a voltage at the open Collector that is proportional to the Base current through this
Emitter resistor. If we derivate vCE with respect to iB, we get the equivalent RE for each
operating point. The value of RE is then the mean value of the flat range in this plot.
What To Do in IC-CAP:
Import the data of setup rb_re
Run transform visu_RE and select 'data transform'
this will derive the measured data and display the
calculated effective RE against the stimulus iB.
Click a box around the most constant range of measured data and click 'Copy to
Variables'
Re-execute transform visu_RE to extract the RE value (select 'extract parameter'
for this operation mode).
Do not simulate or optimize this setup, since
The other DC model parameters are not known yet

The Gummel-Poon model cannot represent 'unconventional' measurement conditions
like the actual flyback method. The values of the ohmic parasitics will be fine-tuned
later in the Gummel-Poon plots
Extraction of RBM from DC measurements

RBM min.Base resistance at high current
584
There are several methods to determine the Base resistor: either the constant part of it
(RBM) from pure DC measurements, or the non-linear RBB' = f(RB,IRB,RBM) from a s11
plot or from noise figure measurements.
Applying these three methods to the same transistor 'will generate typically three different
values' for the Base resistor.
An interesting method to determine RBM is to use the RE-flyback method, with

additionally measuring vBE /T.Zimmer/. This method is applied now.
The theoretical values of the measured voltages are:
Subtracting these equations and dividing by iB yields:
i.e. a regression analysis applied to these transformed measured data will give the y-
intersect RBM.
In a final step, we then apply a loop to these data, in which a line is fitted to two adjacent
points, and the y-intersect is calculated. The incremental y-intersect is then displayed
against the stimulus iB.
When RB becomes measurable DC-wise (the 'ohmic' range in the Gummel-Poon plot), its value is typically
already lowered to the value of RBM. This means, parameter RB (the higher Base resistor value for lower
Base bias), cannot be determined by this method. Therefore, we simply set RB=RBM.
NOTE: See also the appendix chapter 'direct visual parameter extraction'.
The measurement of setup rb_re is re-used
Run transform visu_RBM and select 'data transform'
this will calculate the local Base resistor for each bias point, as described above,
and display the RBM value against the stimulus iB.
Variables'
Re-execute transform visu_RBM to extract the RBM value
Again, do not simulate or optimize this setup, since the other DC model parameters are
not known yet.
If a sensitivity analysis for a Gummel-Plot shows a reasonable impact of the Base resistor to the forward
and reverse Base current, an optimizer run on these two curves simultaneously might make sense to
obtain a guess on the actual value of RBM.
However, this is usually not the case.
Modeling the Collector Resistor

For the extraction of RC, the same flyback method like for RE is applied. The only
difference is that the Collector pin is grounded, and the Emitter pin is left open and its
voltage is measured.
Import the data of setup rc
Run transform visu_RC and select 'data transform'
this will derivate the measured data and display the
calculated effective RC against the stimulus iB.
585
Variables'
Re-execute transform visu_RC to extract the RC value.
Again, do not simulate or optimize this setup
Note: Try also and study macro 'extract_resistors".
Notes
As mentioned above, these 'classical' extractions of the ohmic model parameters are used
to get a good estimation about the parameter values. The values will be fine-tuned later in
the setups fgummel and rgummel.
For details on alternate DC modeling methods of the parasitic resistors, see also the
publications of /Berkner/ and /MacSweeny/.
If there is a parasitic pnp transistor present, this method will not give accurate RC values.
See the corresponding model file of this toolkit.
An alternate method to calculate the ohmic parasitic

resistors from s-parameter measurements
Since the fitting of the S-parameters is the goal of a good transistor modeling, it makes
sense to think about extracting the ohmic parameters from S-parameter measurements
also. The following figures sketch a reliable way to do that. The basic idea is to overdrive
the transistor and to reduce its effect to simple diode characteristics ('hot' measurement).
With the known value of the Base current, the remaining resistor values can be calculated
easily.
586
The resistor values are finally displayed versus frequency and their values are obtained as
a simple mean value. If there is a frequency drift, take the mean value from the lowest
frequency. The plot below gives an example:
587
Modeling the Temperature Effects
The parameters given below are modified when the selected simulation temperature TEMP
is different from the extraction temperature TNOM. (Temperatures in 'K).
Used auxiliary variables:
Temperature dependent modeling parameters:
588
Non Linear DC Modeling
Contents
Extraction of BF , ISE and NE
Extraction of IKF
Extraction of the remaining reverse parameters NR, BR, ISC, NC and IKR
Three measurements are required in order to extract the DC parameters:
An output plot including both, forward and reverse operation,

And two so-called Gummel plots, one for forward and another for reverse mode.
These three plots have a certain context between each other. Neglecting this context can
easily lead to one of the famous, so-called 'infinite modeling loops'.
This can be explained as follows:

Let's consider the forward Gummel plot. It is based on a measurement of i B and i C
simultaneously, versus v BE and is typically plotted half-logarithmically. Most often, the
applied Collector-Base voltage is set to v BC = 0V. The reason for this is that it simplifies
the modeling equations (H)...(L) drastically. However, this approach can easily lead to the
'infinite loop' mentioned above.
IC-CAP does not need the simplified equations. The optimizer in IC-CAP always uses a true
simulator like SPICE in the background that includes the complete Gummel-Poon
equations. Therefore, while extracting the DC parameters or other parameters, there is no
reason for having v BC = 0 for the Gummel plot.
We can take a smarter approach. We first measure the forward output characteristic and
extract VAR and VAF. Then, we leave this setup for the moment, and measure the forward
Gummel plot. Differently from the commonly used method mentioned above, we apply a v
CE that is not zero, but between 2V and half the value of the maximum v CE of the output
plot. We do this for the following reason. Once the Gummel plot is fitted for this special
voltage, the following output plot simulation already hits the measured curves exactly in
the middle of the output characteristic. A final fine-tuning is then easily achieved by
adjusting VAF and BF. Otherwise, if we use v BC = 0 for the Gummel plot, it can easily
happen that if the Gummel plot itself is nicely fitted, the output characteristics doesn't
match and so on. Because, in this case, if the Gummel plot fits, this means that the output
characteristic fits in the saturation range (v CE ~ 0.2 ... 0.9V) and not int the desired
linear range (v CE ~ 0.5 to v CEmax).
An illustration of this idea is presented below in fig.DC-1. First, the output characteristic is
measured and VAR and VAF are extracted. Then, considering a cut through this plot for a
fixed v CE (4V in the example), and using this value of v CE when measuring the Gummel-
Poon plot, we have data points that refer directly to our previous output characteristics
measurement with the corresponding v BE . The relationship between i C and i B leads to
the beta plot, also plotted against linear v BE instead of the usual logarithmic i C (which is
the same for i C below the ohmic effects in the Gummel-Poon plot) and again highlighting
the corresponding output data points by buttons. Therefore, if beta fits, so does the
output characteristic, which we were starting from.
Therefore, if we extract the DC forward parameters from a Gummel-Poon measurement

that is biased like this, all measurements fit together.
589
For the output characteristic, by forcing i B instead of v BE , we also prevent from
measuring thermal self-heating effects, which are not included in the standard Gummel-
Poon model. However, we should also measure the same Collector currents values with a
corresponding v BE as well. This is sketched below in fig.DC-2.
Note:Such a check is implemented in file data_mgmt/BIP_MEAS_MASTER.mdl
Note: i C is scaled identically for both plots.
Because, if the output characteristic drifts off when forcing v BE , we should be careful
when measuring the Gummel plot, because it could be affected by self-heating as well.
The ohmic effects are in this case overlaid by the thermal self-heating, and we will either
get wrong model parameters for RE and IKF, or no good fitting at all.
It is recommended in this case to apply a v CE as low as possible for the Gummel plot
(below the thermal runaway), but well above the saturation region of the output plot.
Note: See also IC-CAP file bip_output_char_i_or_v.mdl in the more_files directory
Modeling the Output Characteristic

VAR reverse Early voltage
VAF forward Early voltage
Modern fast bipolar transistors exhibit small values of VAR. Due to the simplifications of
the G-P space charge model implementation in SPICE, this may affect the other model
parameters. This typically happens for VAR<5.
Therefore, the extraction of the nonlinear DC parameters is best started with the
extraction of VAR, followed by VAF. As will be shown in this chapter, VAR/VAF can be
determined with only little overlay of the other (actually still unknown) parameters.
After the Early voltages are extracted, and before optimizing the fit of this setup, we need
to go ahead and extract the remaining DC forward model parameters from the Gummel
plot. Only then, with the correct BF etc., the simulation of the output characteristic can fit
the measured data.
Therefore, we come back to this setup and fine-tune the VAR/VAF values by optimization
later.
Now, let's discuss the theoretical background of the Early voltage extraction. For an easier
understanding, we consider VAF.
The equation:
Provided that: vB'E'=vBE and vB'C'=vBC , the Gummel-Poon model describes iC by
with the Base charge equation
for the modeling of non-idealities like the Base-width modulation:
and the hi-level injection effect:
590
In order to handle this complex formula, we have to start with some simplifications:
We consider only the forward active region. Here, the Base Collector voltage is vBC < 0V;
therefore the terms
in equ.(DC-4) and (DC-4) may be neglected.
Thus (DC-4) becomes:
Equ.(DC-6) in (DC-5) yields:
Fig.DC-3 showed iC versus vCE. Therefore, (DC-8) has to be re-arranged using
(DC-9) in (DC-8):
with typically vBE << VAR and vBE << VAF we get:
or
Thus we got iC = f(vCE,iB) as shown in fig.DC-3, with vBE = f(iB).
Extracting the parameter:

We consider all assumptions from above valid. This means that we should be sure that the
output characteristics measurement has been taken in the linear range of the Gummel-
Poon plot, i.e. with a maximum iC well below the ohmic effects. Then, the Collector
current of the output characteristics measurement (equation DC-11), becomes zero for
vCE = -VAF.
How to proceed:
VAF is the x-axis intersect of the tangent fitted to the linear region of the output
characteristics.
NOTE: As you will find out with your own measurements, VAF is rather a function of the bias current than
a constant. The standard deviation of the values of VAF found by applying tangents to all slopes in the
output plot is most often very big. Depending on the type of transistor, sigma(VAF) can range up to
VAF/2. The reason is that the assumptions in equations (DC-5)..(DC-11) are pretty straight forward.
Therefore an estimation of VAF by using only 1 tangent may be sufficient, when an optimizer run is
performed later (after the extraction of the remaining DC forward parameters). Please note again that the
IC_CAP optimizer calls the simulator which includes the full set of model equations and therefore finds the
correct final value of VAF.
An alternate method could also be to determine VAF out of the delta of two Gummel plot
curves iC(vBE) for two different Collector-Emitter bias voltages. See equation (DC-18) of
the next chapter.
What to do in IC-CAP:
open setup "/gp_classic_npn/dc/routput",
import the measured data
in plot 'ie_vec', specify a box for parameter extraction
perform transform "br_VAR" (extract VAR)
simulate with the extracted value of VAR.
Then,
open setup "/gp_classic_npn/dc/foutput",
import the measured data
in plot 'ic_vce', specify a box for parameter extraction
591
perform transform "be_VAF" (extract VAF)
simulate with the extracted value of VAF.
Do not be confused about the simulation result, and that the curves do not match.
Because all other DC parameters are still set to default, it is only important that the slopes
of simulated and measured curves match!
We will have a much better fitting after the extraction of the other DC forward
parameters.
Have also a look into "/gp_classic_npn/dc/routput/READ_ME".
Modeling the Collector Current

IS transport saturation current
NF forward current emission coefficient
These 2 parameters, together with the already known Early voltages, are the only ones
that are dominant in the measurement setup given in the figure DC-4 below. NF
determines the slope and IS the y-intersect of the half-logarithmically plotted iC(vBE).
The equation:
Provided that vB'E'=vBE and vB'C'=vBC , we start again with the iC formula (H) ... (L)
from the introduction chapter:
In order to determine IS from (DC-12) for small vBE, i.e. no ohmic and no IKF effects, we
get for forward biasing
vBE ~ 0,7V << |VAR|

vBC < 0 V
and |vBC| << |VAF|.
Therefore (DC-12) simplifies to:
Let's have a closer look to equ. (DC-13). Firstly, the formula reminds to apply the
following series approach for small values of x :
what means for our case:
NqB from (DC-17) is split into two parts: q1S represents a lowering of the Collector
current for increasing Early voltages (DC-14). This can be seen in the iC Gummel plot as a
curve shift to lower Collector currents. On the other hand, the other coefficient q2S begins
to contribute for high Collector currents above IKF in forward operation resp. IKR in
reverse (DC-15a), and reduces the Collector current as well.
592
For the modeling of IS and NF, we consider the lower and medium current ranges well
below the Effect of IKF or the influence of the ohmic Resistor RE . Therefore, (DC-16)
simplifies to:
For bigger values of the Early voltages, the terms |vBx| < |VAx| can be neglected and we
obtain:
NOTE: Compare (DC-19) with the measurement result given in the figure DC-4 above.

Following the curve fitting techniques given in the chapter on regression analysis in the
appendix, a transformation can be applied to the measured data in order to obtain a linear
context between the measured values of iC and the stimulating values of vBE in (DC-19):
A log10 conversion of (DC-19) gives:
This can be considered as a linear form:
How to proceed:
We select a sub-range of the measured data, where the half-logarithmicly plotted data
represent a straight line. Then the logarithmically converted iCi of this sub-range are
interpreted as y- and the linear vBEi values as x-data for the regression formula. Applying
these formulas, we obtain y-intersect 'b' and the slope 'm' of the straight fitted line.
From comparing (DC-20a) with (DC-20b) we know how to re-substitute the parameters
out of 'b' and 'm':
and
Validity of the extraction:

vBE between 0,2V no noise and 0,7V [no high current effects]
open setup "/gp_classic_npn/dc/fgummel",
import the data,
click a box into plot "ibic_vbe" around a linear range for the IS/NF extraction
click 'Copy to Variables' (check how the box bounds are exported into the
setup variables X_LOW, X_HIGH, Y_LOW, Y_HIGH)
perform transform "br_IS_NF" (box regression IS, NF), which refers to X_LOW etc.
simulate with the extracted parameter values.
optimize with transform "bo_IS_NF"
Also have a look into "/gp_classic_npn/dc/fgummel/READ_ME".
HINT:
Transforming the measured data such that the model parameter can be displayed directly
against the stimulating voltage or current is another smart way to determine model
parameters. In the case of NF this would mean to start with
to convert it logarithmically in order to obtain
593
This is the mathematical representation of the half-logarithmic Gummel plot for iC. The
parameter NF is proportional to the slope and we have therefore to differentiate ln(iC)
with respect to v BE and obtain:
Solved for NF gives
Therefore, if we display the calculated NF (what is the 'effective NF' for every measured
data point) versus v BE, we get
This allows us to check, if the model is able to fit the measured data at all (if there is a
constantly flat range) and then to easily extract the parameter as the mean value of that
flat range.
In directory "visu_n_extr" of this toolkit you will find more IC-CAP model files that follow
the idea of direct visual extraction. See also appendix A for more infos.
Comments on ic_measured above IKF:

Referring to many graphical parameter extraction methods and also to fig.DC-4 above,
some more background information is given for the model curve of ic above IKF.
From (DC-17) we get neglecting the Early-effect:
or solved for NqB:
Let us consider the two cases:
Interpreting the result:

From (DC-22b) we learn that the ic curve has half the slope for currents above IKF (see
fig.DC-4). In practice, however, there is always an overlay with the ohmic resistor RE, and
therefore (DC-22b) is not so well suitable for extracting IKF.
However, the overlaying parameter RE is affecting basically both, the i C and the i B curve
in the same way. This means that the effect of RE cancels out for the beta curve =i C/i B.
On the other hand, parameter IKF affects only i C. Therefore, IKF is commonly extracted
from the beta curve of the transistor
Modeling the Base Current

Extraction of BF, ISE and NE
594
BF ideal forward maximum beta
ISE B-E leakage saturation current
NE B-E leakage emission coefficient
In the literature, the three parameters of this chapter are most often introduced with their
corresponding influence on the different ranges of the iB curve in DC-6.
In practice, there is most sometimes an overlay of the influences. This is especially true
for BF in the beta plot (overlaid from IKF and NE). Also, modern transistors have pretty
low recombination effects for the B-E diode: the 'famous knee' (see finger pointer in
fig.DC-6) is not visible. Therefore we will not follow the graphical extraction method, but
develop another method instead. We will derive a formula for the 3 parameters directly
from measured data that has been taken from the range around the 'knee'.
The equation:
Provided that vB'E'=vBE and vB'C'=vBC , then
We assume once again that:
vBE ~ 0,7V << |VAR|

and vBC < 0 V
This simplifies equ.(DC-23) to:
Introducing (DC-19) -i.e. the Collector current iC with neglected high current effects into
(DC-24) yields the pretty simple form:
We will use both iC = f(vBE) and iB = f(vBE) from the simultaneously measured currents
of the Gummel-Poon measurement of fig.DC-4.
We now have iB as a function of vBE as desired.

This equation (DC-25) is one of the few cases during the bipolar modeling, where a non-
linear transform applied to the measured data doesn't give a straight line. (At least, the
author had not sufficient intuition!). Therefore the partial derivations of the curve fitting
error in (DC-25) versus BF, ISE and NE have to be calculated and then set to zero. The
solution of this system of equations finally gives these 3 parameters.
As iB ranges from pico- to milli-Ampere, we will have to minimize the relative error
between measured and fitted curve. Thus we get from (DC-25) after dividing by iB:
Equation (DC-26) is only approximately true for the real measured data iBi, iCi and vBEi.
Therefore it is expanded by the individual error Ereli for every data point of index i:
Using least means square techniques we now have:
It can be shown that the parameters BF and ISE can be separated out of the partial
derivations with respect to BF and ISE with a reasonable effort. This is unfortunately not
595
possible for NE . This parameter has to be iterated - similar to VJ of the space charge
capacitor - until the sum of individual errors according to (DC-29) is minimized.
Step by step:
The partial derivation of (DC-29) versus BF is:
and versus ISE:
(DC-30) is expanded by
and (DC-31) by
These two new equations are added and their sum is solved for ISE:
Now we can also separate BF from (DC-30):
Validity of the extraction:

vBE above measurement resolution and below high current effects.
How to proceed:
A subset of the measured data iBi and iCi , i.e. the range around the 'KNEE' (see fig.DC-6)
are selected and introduced into equations (DC-32) and (DC-33). Next a suitable starting
value for NE is selected ( e.g. NE = 1) and the error according to (DC-29) is calculated. NE
is then incremented until this error becomes a minimum. The triplet of NE, BF and ISE of
this minimized error is the final parameter extraction result.
NOTE: the complexity of (DC-32) and (DC-33) illustrates that transforming measured data to a linear
context and applying linear regression techniques is often a much smarter approach for parameter
extraction.
in setup "/gp_classic_npn/dc/fgummel",
click a box into plot "ibic_vbe" around the 'knee' at low vb,
click "Copy to Variables",
perform transform "br_ISE_NE_BF" (box regression ISE, NE, BF),
simulate with the extracted parameter values.
perform transform "bo_ISE_NE_BF" (box optimization ISE, NE, BF),
If there is no 'knee' with your measured transistor, the Base current recombination effect
does not occur. In this case, switch off the Base current recombination effect in the G-P
model. This can be done by setting ISE to a very small value (ISE=1E-30) and the slope
parameter NE to a flat slope (NE=2).
Have also a look into "/gp_classic_npn/dc/fgummel/READ_ME".
Note:
For low values of VAR, the Collector current formula of (DC-19) inserted into (DC-25) is not quite correct.
It would lead to a too low extracted value of BF, due to the shift of the iC Gummel plot. Equation (DC-18),
without the assumption of big Early voltages, is better in this case. Therefore, correct the measured
Collector current values to
VAR<<VAF, so this correction is sufficient before inserting them into equations (DC-32) and (DC-33).
Modeling the Current Amplification at High Current

Extraction of IKF
IKF forward beta high current roll-off
Referring to fig. DC-4, IKF models the Webster push-out effect. This effect describes a
decrease of the proportionality of log(iC) versus vBE. Unfortunately, as already
mentioned, this effect is also overlaid by RE. However, while RE affects mainly both, iB
and iC, IKF only affects iC. Therefore, it can be best extracted out of =iC/iB.
From a modeling standpoint, the beta plot should not be considered isolated from its
origin, the iB and iC curve. Therefore, we display it always versus the same stimulus vBE,
together with iB and iC. This helps a lot in better understanding the influence of the
parameters ISE and NE on the increase of beta, and of IKF for the decrease.
596
It becomes also clear why for some transistors, BF seems not to affect the maximum of
the beta trace at all: there is no parallel region between log(iC) and log(iB) in the Gummel
plot, or referring to the beta plot, IKF reduces already beta, before it can reach the value
of the BF for increasing bias. This is shown in fig. DC-8.
Note: log() is plotted versus vBE . The conventional plot using log() versus log(iC) is identical to it,
provided we neglect high-current effects. But this way of plotting log() is more straightforward, because
it displays a measurement result versus a stimulus and not another measurement result.
NOTE: for best results, estimate or extract RE first.
Hint
Avoid thermal self-heating effects. For Collector currents >10...50mA, thermal self-
heating has to be taken into account. This becomes visible if the beta-plot for a forward
and reverse vBE sweep looks different at high vBE. To avoid this, DC pulsed
measurements with pulse widths about 1us should be used in this case (the HP4142 offers
only pulse widths 100us).
Before we start with the extraction of IKF, we are now ready to understand the following
schematized Gummel plots. They characterize at a glance the different effects for the Base
and Collector current in the Gummel-Poon model.
597
The equation:
The current amplification is defined as:
Provided that: vB'E'=vBE and vB'C'=vBC and further
We introduce (DC-24) for iB and(DC-16) for iC into (DC-34):
We further introduce an approximation for NqB, see equ.DC-18. This gives:
Divided by
we get:

As we want to consider again relative errors, we proceed as in the chapter of the
determination of the iB parameters:
This formula is again more ore less true for the measured data iCi and iBi with the
stimulating voltage vBE. Thus we have to introduce again an individual error Ereli
for each measured data point of index i:
Finally the total error for all measured data ( 1 .. N ) is (least means square):
598
How to proceed:
In order to keep things simple, (DC-39) is solved for a best IKF by iteration. Thus IKF is
set to a starting value, e.g. 10A, and then divided by 2 in every iteration, until the total
error given in (DC-39) is minimized. Fine-tuning is then done by the optimizer.
open setup "/gp_classic_npn/dc/fgummel" and plot "ibic_vbe"
(beta is the right-axis data), then
perform transform "e_IKF" (extract IKF) and
check the simulation result.
run transform 'o_BF_IKF_RE' for fine-tuning the parameters of this setup.
Have also a look into "/gp_classic_npn/dc/fgummel/READ_ME"

and also in some alternate methods on the IKF-extraction in the 'direct visual parameter
extraction' model file.
Extraction of the remaining Reverse Parameters

Extraction of NR, ISC, NC, BR, IKR
The reverse modeling can be performed like the forward modeling. Simply exchange
Emitter and Collector.
open transform README in setup "/gp_classic_npn/dc/rgummel" and follow the
modeling sequence given there
Last not least, macro 'extract_n_opt_DC' includes a suitable automated modeling strategy
for both DC forward and reverse.
Included in this example is also the interesting and pretty often recognizable effect, that
the reverse Early voltage is affecting the forward modeling, due to its low value.
The strategy used in this macro covers that effect by looping a bit between forward and
reverse extraction and optimization.
This sequence may be different for your actual transistor. Just correct the macro if
required.
Note on reverse Gummel-Poon modeling

If your reverse beta curve and reverse Gummel plot look like below,
(a steeper slope of log(iB) versus vBC for vBC < vx, something that is not included in the
Gummel-Poon model), you might consider replacing the Gummel-Poon recombination
modeling (parameters ISC and NC) by a external diode with its parameters IS, N and RS.
For more details, refer to file rgummel_special.mdl under directory 'more_files' in this
toolkit file collection.
See also the chapter on the limitations of the Gummel-Poon model at the end of this
manual.
599
Preface
This product has been developed to meet the local demands of European IC-CAP users for
more technical background information on extraction techniques and for the availability of
extraction source code. Published for the first time in 1990, it has been updated since then
several times. It is part of a series of supplementary modeling tool-kits for the IC-CAP
users. These products feature source code and detailed technical description of the
extraction routines. Please contact the author for further information.
About Gummel Poon User Information

This information is intended to explain the basics of modeling a bipolar transistor using the
Gummel-Poon model as it is implemented in the simulation program SPICE of the
University of California Berkeley (UCB) /see publication list/. It is part of the Gummel-
Poon Bipolar Model Parameter Extraction Toolkit.
This toolkit includes the IC-CAP model file

GP_CLASSIC_NPN.mdl, the MASTER model file which is described in this manual and
featuring the data management features of IC-CAP 5.x, i.e. separating measurements
from extractions:
NPN_MEAS_MASTER.mdl a master file for measurement

GP_EXTRACT_NPN.mdl a master file for modeling
as well as manyother IC-CAP model files covering topics like:
model parameter extraction using the tuner feature

direct visual parameter extractions
alternate modeling methods for DC- CV- and RF-parameters.
bipolar transistor modeling including the parasitic transistor.
Note
After you have become familiar with the modeling procedure itself, i.e. file GP_CLASSIC_NPN.mdl, you are
encouraged to split the modeling into 2 parts: separate measurements and separate extraction strategy.
In this case, all measurements are performed using the file NPN_MEAS_MASTER.mdl. Then, the data are
exported into IC-CAP mdm files (ASCII files) and imported into the master extraction file
GP_EXTRACT_NPN.mdl for extraction. This method allows to improve continuously the extraction strategy
file, independent of the measurement data.
IMPORTANT NOTE:
This manual and the underlying IC-CAP model file GP_CLASSIC_NPN.mdl are intended to
explain the basics of the Gummel-Poon modeling. Therefore, it covers the classical
Gummel-Poon model without enhancements for also modeling the parasitic transistor.
However, as stated above, such model files are included in the file sets of this toolkit.
Please see the README macros in these IC-CAP model files for more details. You are also
invited to get in contact with the author for assistance with such modeling problems.
The IC-CAP model file "GP_CLASSIC_NPN.mdl"

features
The extractions are written using PEL (parameter extraction language) and are open to
the user. They can be easily modified to meet specific user needs. Sub-circuit model
description, open for user enhancements (HF modeling, parasitic pnp etc.). All transistor
pins are connected to SMUs for flexible measurements
The transistor output characteristic and S-parameter measurements use a Base current
stimulus rather than a Base-Emitter voltage in order to avoid 1st order thermal effects
being visible. However, self-heating might be present and affect the Gummel plots in the
ohmic range. See also the file GP_MEAS_MASTER.mdl.
Organization of Information
There are 5 main chapters, which explain how to determine the model parameters from
CV (capacitance versus voltage), then parasitic ohmic resistors, and DC, to finally high
frequency measurements using network analyzers.
More chapters cover side aspects of bipolar modeling.
The individual chapters follow always this scheme:
1. explanation of the parameter-dependent measurement setup

2. explanation of the mathematical basics for the parameter extraction
3. explanation of the parameter extraction
4. explanation about how to use the IC-CAP file.
600
Publications
P.Antognetti, G.Massobrio: Semiconductor Device Modeling with SPICE, McGraw-Hill,
1988, ISBN 0-07-002107-4
F.van der Wiele, W.L.Engl, P.G.Jespers: Process and Device Modeling for Integrated
Circuit Design, NATO Advanced Study Institute, Louvain-la-Neuve, Belgium, 1977
D.A.Hodges, H.G.Jackson: Analysis and Design of Digital Integrated Circuits,

McGraw-Hill, New York, 1983, ISBN 0-07-029153-5
P.R.Gray, R.G.Meyer: Analysis and Design of Analog Integrated Circuits, John

Wiley&Sons, New York, 1984, ISBN 0-471-87493-0
I.Getreu: Modeling the Bipolar Transistor,Tektronix Publication No. 062-2841-00,

1976
H.K.Gummel, H.C.Poon: An Integral Charge Control Model of Bipolar Transistors, Bell

Syst.Techn.J., 49 1970 , p.827
Simplification of DC characterization and Analysis of Semiconductor Devices, HP

application note 383-1
EEsof, Xtract Supplement Manual, March 1994, HP-EEsof.
Application Note AN 1202-4: Advanced Bipolar Transistor Modeling Techniques,

IC_CAP marketing group
T.Zimmer: Contribution la modlisation des transistors haute frquence, Thse

l'Universit de Bordeaux I, 17.7.1992.
Zimmer, Meresse, Cazenave, Dom, 'Simple Determination of BJT Extrinsic Base

Resistance', Electron.Letters, 10.10.91, vol.27, no.21, p.1895
Paul Schmitz, Vector Measurements of High Frequency Networks, Hewlett-Packard,

Publication HP5958-0387, April 1989
P. van Wijnen, On the Characterization and Optimization of High-Speed Silicon

Bipolar Transistors, PhD Thesis University of Delft, 1992, publshed 1995 by Cascade
Microtech, Inc., Beaverton, Oregon
F.X.Sinnesbichler, Grosignalmodellierung von SiGe-Heterobipolartransistoren fr den

Entwurf von Millimeterwellenschaltungen, Dissertation (PhD Thesis) Universitt
Mnchen, VDI-Verlag, Dsseldorf, ISBN 3-18-332709-0
B.Ardouin, Contribution la modlisation et la caractrisation en hautes frquences

des transistors bipolaires a heterojunction Si/SiGe, PhD Thesis University Bordeaux,
cole doctorale de sciences physiques et de l'ingenieur, December 2001
Paras.resistor modeling
J.Berkner: A Survey of CD-Methods for Determining the Series Resistances of Bipolar
Transistors Including the new I Sub Method, SMI GmbH, Frankfurt/Oder, published
in the proceedings of the European IC-CAP User's Group Meeting 1994, October 10-
11, Colmar, France
DC and paras.resistor modeling:

D.MacSweeny et.al.: Modeling and Parameter Extraction for Lateral Bipolar Devices
Using IC-CAP Sub-Circuits, Proceedings of the 1996 Europ.IC-CAP User Meeting in
The Haague, Netherlands
T.Ning, D.Tang: Method for Determining the Emitter and Base Series Resistance of
Bipolar Transistors, IEEE Transactions on Electron Devices, vol.ED-31, pp.409-412,
April 1984
CV and TF modeling:
J.Berkner: Parasitre Effekte bei der SPICE-Modellparameterbestimmung fr
integrierte Bipolartransistoren, Halbleiter Elektronik Frankfurt/Oder GmbH, published
in the proceedings of the European IC-CAP User's Group Meeting 1993, June 22-23,
Esslingen, Germany
TF modeling:
Kendall et al.: Direct Extraction of Bipolar SPICE Transit Time Parameters Without
Optimization, published in the proceedings of the US IC-CAP User's Group Meeting
1993.
For providing valuable feedback on the extraction methods presented in this document,
the author would like to especially acknowledge:
Mr. Jrg Berkner, Infineon, Munich, Germany
Mr.Dermot MacSweeny from the NMRC in Cork, Ireland,
601
Transit Time Modeling
For a better overview, the Gummel-Poon basic capacitor equations are repeated first from
the CV chapter:
Provided that: vBE = vB'E' and vBC = vB'C', the capacitors in the Gummel-Poon model
given in the introduction chapter with equations (O) ... (S) are:
And
again from the introduction chapter and additionally with:
and the ideal Collector current If from C

CSBi models the space charge and CDBi the diffusion capacitance between Base and
Emitter or Base and Collector respectively. vBE and vBC are the stimulating voltages.
For CBE, and its diffusion capacitor part CDBE, the exact definition is (private
communication with D.Celi, ST Crolles, France):
where TF ac is the small signal transit time which is different from TFF. However, in this
toolkit, we keep the simplified equation of CV-2 and a corresponding raw parameter
extraction, and apply a quick optimizer run to get the true TF parameter.
Modeling the Diffusion Capacitor CDBC

Extraction of TF, ITF and XTF
TF - ideal forward transit time
XTF - coefficient for bias dependence of TF
ITF - high-current parameter for effect on TF
For forward active operation of the transistor, the AC behavior is modeled by CBC and
CBE. In this operating mode, the already modeled CV-capacitance CSBC dominates over
the diffusion capacitance CDBC , while in another equation, the more important term is
CDBE. This chapter covers the modeling of CDBE.
CDBE is described by the bias-dependent transit time T FF.
TFF is modeled with the formula:
with the ideal forward Base current i f

TF is the ideal forward transit time modeling the 'excess charge'. The parameters XTF, and
ITF cover the operating point dependence from the DC bias iC ~ if , while VTF describes
the dependence from vCB ~ vCE.
Pre-considerations concerning the measurement

Like in the previous chapter, the parameter estimation is again performed using a
602
simplified model, whereas the parameter fine-tuning is finally done during an optimizer
run using the full set of SPICE model equations.
Referring to appendix B, it can be shown that the transistor's h21(f)-parameter behaves
frequency wise like a low-pass filter with the transfer function
or inverted
where f T1-pole is a function of the bias current i C and the bias voltage v CE
In many publications, like e.g. /Sinnesbichler p.106/, it is mentioned that the transit time
after the above equation is:
In this case, the TFF used for modeling is RC*CBE smaller than the value converted from f
T.
In some other publications, this formula is extended to
or after /B.Ardouin, p.198/
In practice, however, with the goal of a direct extraction of the TFF parameters followed
by a post-optimization, the additional terms can be neglected and the simple equation
(TFF-1) is sufficiently correct.
Now back to the parameter extraction

We will first consider the extraction of the parameters TF, ITF and XTF. VTF will be
covered later. This means, a special measurement for fT(iC) and later for fT(iC, vCE) is
needed. As we assume a 1-pole low-pass for h21, the gain-bandwidth product is a
constant. Therefore it is sufficient to measure a h21(iC, vCE) at a fixed frequency higher
than the the -3dB frequency.
In other words, this fixed frequency should be from a -20dB/decade range of MAG[
h21(freq, iC, vCE) ] . This measurement frequency can be found when transforming the
measured rBB' S-parameters to H-parameters (using the TwoPort function). From the dB-
plot of ABS(h21(f)) versus log(frequency) we determine a frequency where the slope fits a
-20dB/decade roll-off.
Note
If your MAG[ h21 ] does not follow the -20dB/decade law, there is probably a so-called over-
deembedding. This means more parasitics are subtracted than present in reality. The opposite de-
embedding problem, under-deembedding, does not affect the slope, but it can show up like a second -
20dB/decade slope shifted in frequency.
This frequency is now used as a fixed frequency f-20dB for the S-parameter
measurements of this setup. The underlying DC bias values are a swept iC and a constant
and small value of vCE (to neglect the VTF effects). Then, these S-parameters are
converted into H-parameters and we get for the constant gain-bandwidth product of this
assumed one-pole low-pass filter:
or
f T1-pole after equ.TFF-2 is valid for all DC bias conditions, i.e. for the whole bias-
dependent array of h21. This new array fT1-pole is then introduced into (TFF-1), what
gives the bias-dependent array of TFF to be fitted.
Figure below shows log|h21| as a two-dimensional function of the Collector current i C and
the frequency freq. It shows the transit frequency with and without simplification
(Appendix B). The dependence of v CE is neglected for simplification.
603
Usually, f T is plotted against i C . This is the typical diagram published in many data
sheets. the following figure shows such a curve, also including the dependence of f T from
v CE.
Note
For a correct modeling, check the fT curve at low iC for so-called self-biasing! This effect occurs if the RF
signal power at the Base is in the range of the DC bias power. Under this condition and considering the
non-linear diode characteristic at the Base of the transistor, the rectified AC signal will contribute to the
DC bias! A flat trace of the fT curve at low Collector current is an indicator for that effect.
For more detailed examples about how the RF power level might affect the fT curve, see
literature P.v.Wijnen, chapters 3 and 4, and the IC-CAP examples on non-linear RF
modeling, available from the author.
Pre-considerations concerning the model equation:

As cited at the beginning of this section, we start with:
with the ideal Collector current i f v BC as well as v BE are the DC bias voltages at the
operating point.
In this equation, i f is the ideal Collector current. If we consider currents below IKF ,we
can set if = ic . After the extraction of the parameters of this section, we will use a final
optimization on the S-parameter curves, which will eliminate this small error.
Therefore the curve-to-be-fitted is:
If we choose vCB ~ 0, we can further simplify and get finally:
Validity of (TFF-4): iC<IKF and linear forward operating point with vCB~0 or as small as
possible.
Performing the measurement
604
First the network analyzer is set to a constant frequency on the -20dB/decade roll-off of
|h21(f)|. The used test frequency had been estimated from the S_to_H parameters of the
RBB measurement. More details were given above.
Next the transit frequency for a 1-pole low-pass model is calculated as given in (TFF-2):
Then we calculate TFF = 1/(2 PI fT1-pole) as the bias-depended total transit time.
Extracting the parameters

The following figure highlights the theoretical trace of TFF of equation. This figure shows
the theoretical curve in addition to the measured one to make things more clear. The real
measured curve is overlaid by the space charge capacitor effects for low collector
currents. This can be seen also from equation in the chapter on CV modeling. For a
detailed description of this effect, see /Berkner 1993/
Due to these overlay and measurement problems, it had been found that a pretty simple
and straight-forward extraction technique can be applied that gives nevertheless quite
reasonable results. This method is explained below. There exist some more complex
strategies, but the extraction results may be not much better. As sketched in the above
figure, this is mainly because it is not possible to force such a high Collector current that
the trace can be obtained from measurements without being overlied and distorted by
thermal self-heating effects.
How to proceed
TF
TF is extracted as the minimum value of TFF.
Note
A prerequisite is vBC=0, i.e. select a vCE ~ 1V for the extraction, and no Collector voltage in quasi-
saturation!
XTF
The behavior of TFF is given in the above figure4. It is difficult to measure for a higher
Collector current due to thermal limitations. So XTF is estimated from the trace of TF at
max. applicable Collector bias current under the assumption that it would be TFF at infinite
current:
MAX(TFF) = TF (1 + XTF)
or
605
ITF
Referring to the same measurement restrictions as above, a good first-order estimation of
ITF is related to the max. Collector current measured:
ITF = MAX(iC_meas) / 2
Again, since the end of the TFF trace is often not measurable, correct this estimation by
ITF=5*ITF(equ.TFF-6).
This ITF extraction method follows the idea of the Base resistor IRB parameter extraction!
Note
The extraction of the parameter TF, ITF and XTF is performed simultaneously with parameter VTF in a
single extraction routine.
Extraction of VTF
VTF Voltage describing the vBC dependence of TF
Finally, we consider also the vCE sweep.
VTF can be obtained from for a fixed value of iC:
Or
This gives:
And finally:
After the extraction of these four parameters for CDBC , we will next run an optimization
to improve the fitting of the fT plot. However, it is very important, that we do not forget to
also optimize the S-parameter fittings for all bias conditions after that (setup rbb).
Note
Notes:
From the pre-considerations given above, the plots RBB vs. iB and TFF vs. iC represent curves which had
been extrapolated from the S-parameters. So, the S-parameter measurement is "the real world" and the
fitting should be optimized in this domain.
Again, all the AC extraction methods need absolute clean measurements and elimination
of parasitics by de-embedding techniques. Otherwise no curve fitting might be possible or
the parameters obtained might make no physical sense.
1. open setup "/gp_classic_npn/nwa_extr/tf_vbe_vce",
2. select a -20dB/decade frequency from the plot 'mag_h21vsf' of setup 'rbb' (middle
606
mouse click)and enter it to input 'freq'
3. import the de-embedded data of this -20dB/decade frequency into setup 'tf_vbe_vce'
from the .mdm file
4. perform transform "calc_TFF"
5. check the plots "ft_vbe" and "TFF"
6. perform transform "e_TF_ITF_XTF_VTF"
7. simulate with the extracted parameters
How to optimize the S-parameter Setup in IC-CAP

For the data supplied with the file 'gp_classic_npn.mdl', this S-parameter sequence is
best:
1. In setup "/gp_classic_npn/nwa/tf_vbe_vce", run the optimizer transforms

"to_TF_ITF_XTF" to optimize for the smallest vCB bias., then, execute "o_VTF" to
opotimize the fitting for the biggest vCB bias finally, call "o_TF_ITF_XTF_VTF to
optimize all bias points"
1. Go back to setup "rbb" and optimize the S-parameter fitting of the RB, IRB and RBM
setup by running "o_RB_IRB_RBM"
2. In the same setup, execute finally "o_TFF", what optimizes all TFF parameters at all
bias conditions of setup "rbb".
See also macro "extract_n_opt_NWA"
Note
Depending on the device type, it has been observed that the fT-fitting affects also the rBB-fitting.
Extraction of PTF
PTF excess phase at frequency 1/(2PI * TF)
Implemented into the model as a 2nd order all-pass Bessel-function, this parameter can
be used to add some extra phase to the RF simulation curves. It can be obtained when
plotting the phase of h21 of the TFF measurement from above (fig.TFF-5)versus vBE
/T.Zimmer/. As a limitation of the method, the measurable range is again only covering a
small part of the desired curve. Therefore it is advisable to use an optimizer run in order
to get the value of PTF.
Extraction of TR
TR ideal reverse transit time
The reverse transit time of the Gummel-Poon model is modeled by one parameter only,
TR.
TR can be obtained from an h21 measurement versus frequency at a typical reverse

operating point. An optimizer run on the S-parameters of this setup is used to obtain the
parameter value from a typical starting value, e.g. TF * 100.
607
Note
After /Sinnesbichler/, TR can be optimized in the S22 and S12 plot of the reverse biased
S-parameters. In this publication, the reverse biasing of the S-parameter is for example:
1st order DC sweep: vCE = -1 ... -5V
2nd order DC sweep: vBE = -0,7 ... -1V
Note
TR can also be obtained from pulse measurements using an oscilloscope. In many cases, TR is the
dominant parameter in such a setup, often more important than TFF and the junction capacitances.
Therefore, for the modeling of transistors in digital applications, the TR modeling is a must.
608
YParameter Modeling
Contents
Modeling of Transistors in the Operating Point (iccapmhb)
S2Y-Parameter Transistor Modeling GUI (iccapmhb)
Testing S-Parameters for Small Signal Transistors (iccapmhb)
609
Modeling of Transistors in the
Operating Point
The Modeling of Transistors in The Operating Point
FIRST EXAMPLE: GaAs TRANSISTOR
If a transistor is operated at a fixed bias point only, it is not necessary to model it also for
large signal behavior. In this case, a small signal model is used instead. This chapter gives
a examples on how to do this.
It is assumed that a packaged GaAs transistor shall be modeled. The package contributes
with series components like inductors and resistors. This will be modeled in a first step.
The 'inner' small signal transistor is modeled afterwards.
In order to separate the serial components of the package, the transistor is over-biased
and behaves like a good SHORT (so-called cold measurement). This is achieved by biasing
a strong current iG into the Gate and drawing half of its value out of the Drain with the
Source connected to Ground. The assumed underlying equivalent schematic is sketched in
the following figure:
Schematic of the serial components of the transistor package under the condition of a shorted inner transistor
(Figure 1)
The following figures show the measurement (and simulation for later reference) of the
transistor under this bias condition (COLD measurement).
Measurement and simulation result of the packaged transistor in COLD measurement mode (Figure 2)
We transform the measured S-parameters to Z, since the assumed schematic of Figure 1

is a TEE structure.
Solved for the components of the schematic, we get:
After having plotted these parameters versus the frequency (fig.3 and fig.4), we can
easily check the applicability of our assumed underlying schematic. If the parameters have
no major frequency dependence, then the assumption is correct. In our case, this is true,
and we calculate the mean parameter values out of these plots. A simulation with these
parameters is also illustrated in figures 3,4 and 2. The slight difference stems from the
fact that the measured data obviously represents not quite simply a TEE schematic.
!opr_pt3_resistor values.gif !
Resistor values of the schematic of fig.1 and the simulation result using the extracted mean parameter values (Figure
3)
610
Inductor values of the schematic of fig.1 and the simulation result using the extracted mean parameter values
(Figure 4)
In the second step, we bias the transistor in the operating point. The measured S
parameters (fig.6) are transformed to Z, and the previously obtained inductors and
resistors are de-embedded. The result is depicted in fig.5.
Measurement (and simulation results for later reference) of the transistor biased in operating point (Figure 5)
Then, the result is converted to Y parameters, since we now assume the 'inner transistor'
schematic of Figure 6, which represents a PI structure:
Schematic of the Inner Transistor (Figure 6)
The S to Y converted parameters are now interpreted following the assumptions of fig.6a:
and solved for the component values as given in fig.6b:
Cgs and Rgs: Cgs = -1 / (2*PI*freq*IMAG(1 / Ygs))

Rgs = REAL(1 / Ygs)
Cds and Rds: Cds = IMAG(Yds) / (2*PI*freq)

Rds = 1 / REAL(Yds)
611
Trace of Rgs and Rds versus frequency (Figure 7a)
Trace of Cgs and Cds versus frequency (Figure 7b)
Cgd and Rgd: Cgd = -1 / (2*PI*freq*IMAG(1 / Ygd))

Rgd = REAL(1 / Ygd)
Trace of Rgd versus frequency (Figure 8a)
Trace of Cgd versus frequency (Figure 8b)
Ygm: Ygm = Y21 - Y12
612
Trace of MAG(gm) versus frequency (Figure 9a)
Trace of PH(gm) versus frequency (Figure 9b)
After this modeling, we can check the quality of the developed equivalent schematic. The
simulation result, which matches properly the measured data, is depicted in fig. 5.
Modeling with Parameter Tuning

Parameter tuning is another method to determine the values of the circuit components of
small signal schematics. Again, a GaAs transistor is assumed. The underlying schematic is
given in fig.1.
Small signal GaAs transistor schematic
The S-parameter modeling can be done in the following order:
1. All components of the equivalent schematic are 'switched-off'. I.e. all resistors set to
1mOhm, inductors and capacitors to 1 Atofarad resp. -henry (10-18).
2. Modeling of S21 using Rout for f->0.
gm is known from DC-measurements and should not be changed. Therefore, R3 is
used for the fitting.
(Attention, if id>5mA, self-heating of the transistor chip may have occurred. In such
a case, gm_DC will be different from gm_AC !!)
3. Modeling of S11:
R1 is the ohmic value for infinite frequency (may be overlaid by R3)
C1 fits the phase shift
Note: if Sxx shows a notch, we should consider that the power flows for low frequencies through
C12 to the port 2 of the NWA and passes through C1 or C2 only for higher frequencies. This is the
case for C12 > C2 resp. C1.
In such a case, we should remember that the opposite NWA port represents 50 Ohm and that these
50 Ohm will show up in the S11-curves!
Modeling of S22:
Rout is the ohmic value for low frequencies
R2 is the ohmic value for infinite frequencies (may be overlaid by R3)
C2 turns the phase
1. Fine-tuning of S21:
First, the influence of L3 should be checked
Then, fine-tuning C12 and R3, we improve the upper frequency fitting.
2. Finally, L1 and L2 are tuned.
!opr_pt_small_signal1.gif!
Solved for the conductance of the Y schematic branches, we obtain
613
ZIN
For ZINlike given above, we assume the following schematic:
what leads to the component values:
Cgs and Rgs: Cgs = -1 / (2*PI*freq*IMAG(1 / Ygs))
ZTRANS
Cgd and Rgd: Cgd = -1 / (2*PI*freq*IMAG(1 / Ygd))

Rgd = REAL(1 / Ygd)
ZOUT
614
Neglecting Cout for the moment, we can separate for

Rout, Cds and Rds, Cout:
Rout = 1 / (REAL(Yds)), extrapolated to 0Hz
then we can calculate
Zds = (Yds - 1/Rout)^-1
what finally gives Rds = REAL(Zds)
and Cds = -IMAG(Zds)^-1 / (2*PI*freq)
C out is then finetuned for high frequencies in S22
gm
Ygm: Ygm = Y21 - Y12
gm = MAG(Ygm)
TAU = -PH(Ygm) / (2PI * freq)
Developing A Small Signal Model For Mos Transistors
For frequencies below 1/10 fT, analytical expressions of the Z TEE_PI parameters can be
derived which represent the specific frequency behavior of (1):
615
Publications:
G.Dambrine, F.Heliodore, E.playez, A NEw Method for Determining the FET Small-Signal
Equivalent Circuit, IEEE Trans. on MIcrowave Theory and Techniques, vol.36, no.7,
July 1988
M.Berroth, R.Bosch, High Frequency Equivalent Circuit of GaAs FETs for LArge Signal
Applications, IEEE Trans. on Microwave Theory Tech., Vol.MTT-39, No.2, 1991, p.224
D.Lovelance, J.Costa, N.Camilleri, Extracting Small-Signal Model Parameters of Silicon

MOSFET Transistors, IEEE, MTT-S intern.Microwave Symp.Digest, 1994, pp.865-868
S.Lee, H.K.Yu, C.S.Kim, J.G.Koo, K.S.Nam, A Novel Approach to Extracting Small-Signal

Model Parameters of Silicon MOSFETS's, IEEE Microwave and Guided Wave Letters, vol.7,
no.3, March 1997, pp.75-77
J.P.Raskin, G.Dambrine, R.Gillon, Direct Extraction of the Series Equivalent Circuit

Parameters for the Small-Signal Model of SOI MOSFET's, IEEE Microwave and Guided
Wave Letters, vol.7, no.12, Dec.1997, pp408-410
C.C.Meng, G.-H.Huang, Unique Determination of the Equivalent Circuit Parameters for

Uniform-Doped GaAs MESFET, The Fifth Symposium on Nano Device Technology,
SNDT'98, National Nano Devices Laboratories, National Science Council, Chungli, Taiwan
R.O.C.
616
S2Y-Parameter Transistor Modeling GUI
This chapter is a cookbook about how to use the file
demo_features\1_BASIC_MDLG_EXAMPLES\
24_general_Transistor_DC_n_S2Ypar_Schematic_Mdlg\
Xtor_S2Ypar_schematic_Mdlg_PELdep.mdl
The underlying schematic of Y parameters is a PI schematic
Starting with the Y-parameters of the MOS transistor small-signal schematic
Ygd = -Y12Ygs = Y11 + Y12Yds = Y22 + Y12Ygm = Y21 - Y12 = gm * exp(-jTAU)
How to Proceed
The following individual steps explain how to get the MOS transistor schematics out of the
de-embedded S-parameters.
In the IC-CAP GUI, check boxes allow to distinguish between MOS transistors (Cgs in
series with Rgs) and bipolar transistors (Cbe in parallel to Rbe).
Also, two models for the output impedance are selectable in the IC-CAP GUI.
1. Convert de-embedded S-parameters to Y

2. Calculate complex impedances and Gm
Z_10 = (Y11+Y12)^-1
617
Z_20 = (Y22+Y12)^-1
Z_12 = -Y12^-1
Gm = Y21 - Y12 = gm * exp(-j*2PI*TAU)
3. Calculate finally
R10 = REAL(Z_10) C10 = -(IMAG(Z_10))^-1 // (2PI*f)
R12 = REAL(Z_12) C12 = -(IMAG(Z_12))^-1 // (2PI*f)
GM = MAG(Gm) TAU = PH(Gm) / (2PI*f)
R20 = (REAL(Z_20^-1)-1) C20 = IMAG(Z_20-1) // (2PI*f)
How to use the GUI
After starting the GUI, you can change the sweep order at your convenience (step 1.
above).
Then, a 2nd GUI pops up. On its left, you can select a subset of DC biases for data
inspection, but also for your later modeling using the right side of this GUI.
In this right side (step 3. in the slide above), you can specify which output impedance
model you want the plots to apply to, and which input inpedance (depending on MOS or
bipolar transistors) you want to use.
The following sections can assist you in better understanding the Impedance Plots of the
S2Y modeling procedure.
Interpreting the input impedance schematic
From the output impedance to its schematic
618
Note: While tuning Rout, keep an eye on the off-stripped impedance plot Zout,
representing the rest of the schematic !
What if the locus curve becomes inductive !!

If the impedance plots of de-embedded S-parameter measurements exhibit inductive
behavior (Sxx extend into the upper ohmic part of the Smith Chart, S12 shows a
resonance), you should first be suspicious about your de-embedding, and also about the
de-embedding layout. However, if a de-embedding verification gives good results (using
the THRU dummy), your inner transistor may indeed include inductances (the lines of
longer finger structures etc.).
Fine-tuning the series parameters so that the remaining inner Y parameters become
physically correct.
619
You can find the schematic and the parameter values of the series impedances by tuning
their values and inspecting the traces of the locus curves of the inner PI schematic.
Note: this method is not implemented in the IC-CAP GUI applied in this chapter.
Modeling Result Examples
Conclusions:
Y-parameter (PI schematic) modeling

is much more sensitive than S-parameter modeling
offers a nice separation of model parameters in the different Yxy plots
gives much inside view to the effects of model parameters
Y-parameter modeling is easy when applying the model-independent Y-Parameter

Modeling GUI in IC-CAP
In the demo_features directory, go to

\1_BASIC_MDLG_EXAMPLES\24_genl_Xtor_DC_n_S2Ypar_Schematic_Mdlgand load the
file Xtor_S2Ypar_schematic_Mdlg_PELdep.mdl
Try the S2y modeling yourself with the demo data included there
Of course it is most interesting when you apply it to your own, measured data.
Try it !!!
620
Testing S-Parameters for Small Signal
Transistors
IC_CAP file:
demo_features\1_BASIC_MDLG_EXAMPLES\29_TestSpars4Circuits\3transistor_PI_mdlg.mdl
This chapter is intended to present some matrix manipulations for the interpretation of S-
parameters of MOS and bipolar transistors. This is done by transforming the S-parameters
to Y-parameters.
PI circuit interpreted as a Y matrix
watch the signs!
Converting S-parameters into a small signal

schematic of a bipolar transistor:
Bipolar:
How to obtain the components values out of the S-parameters:
1. Transform the S parameters to Y

2. Calculate the complex conductance and the complex GM
Y_10 = Y11+Y12
Y_20 = Y22+Y12
Y_12 = -Y12
GM = Y21 - Y12 = gm * exp(-j**TAU)
3. Separate the results
R10 = REAL(Y_10)^-1 C10 = IMAG(Y_10) / (2PI*f)
gm = MAG(GM) Tau = PH(GM) / (2PI*f)

schematic of a bipolar transistor,
ZBC representing the off-state diode:
Bipolar:
621

2. Calculate the complex conductances and the complex GM
Y_10 = Y11+Y12
Y_20 = Y22+Y12
Y_12 = -Y12
GM = Y21 - Y12 = gm * exp(-j**TAU)
3. Separate the results
R12 = REAL(Y_12^-1) C12 = -(IMAG(Y_12^-1) * (2PI*f))^-1

schematic of a MOS transistor:

2. calculate the complex conductance and the complex GM
Y_10 = Y11+Y12
Y_20 = Y22+Y12
Y_12 = -Y12
GM = Y21 - Y12 = gm * exp(-j**TAU)
3. separate the results
622
VBIC
Contents
Introduction to VBIC (iccapmhb)
VBIC Background Information (iccapmhb)
Recapitulating the Gummel-Poon Model (iccapmhb)
The VBIC Model (iccapmhb)
Comparing the VBIC and Gummel-Poon Parameters (iccapmhb)
VBIC Modeling Strategy (iccapmhb)
Measuring and Extracting the VBIC Parameters (iccapmhb)
Regression Analysis (iccapmhb)
Acknowledgements, Web info, and Publications (iccapmhb)
623
Introduction to VBIC
VBIC - The Vertical Bipolar Inter-Company Bipolar Model
VBIC Model Physically
VBIC model Electronically
In 1954, Ebers and Moll have proposed a large signal model for bipolar transistors. This
model is still the background of today's bipolar transistor models. It describes the
fundamentals of the DC behavior. However, low and high current effects, as well as
parasitic resistors and dynamic behavior are not yet covered. Fig.1 depicts the topology of
the underlying equivalent schematic.
Equivalent Schematic for a Bipolar Transistor after Ebers-Moll
624
Based on the schematic of fig.1, an alternate, yet mathematically identical formulation has
been introduced. Instead of injection currents, it is based on a transport current. This
means, the two current sources of the forward and reverse current are combined into a
single current source (Fig. 2). The main difference between both models is the different
reference currents. The injection version model is based on both diode currents, while the
transport version makes use of the currents ICC and IEC, which give the current ICT.
Hybrid Equivalent Schematic of the Ebers-Moll Model
This large signal model has been enhanced and improved many times. Pedersen
introduced a classification into three Ebers-Moll model versions EM1, EM2 and EM3, see
Getreu.
The EM3 model covers already all essential effects, which are then included in the
Gummel-Poon model, published in 1970 see publications . The important advantage of
modeling the bipolar transistor with the Gummel-Poon model is especially the very clear
and standardized description of many effects by introducing the "integral charge control
relation". Therefore, with the years, the Gummel-Poon (G-P) model has become a
standard for the modeling of bipolar transistors.
It should be mentioned, however, that this modeling standard is usually a special version
of the G-P model, which has been implemented into the simulation program SPICE of the
University of Berkeley UCB, California. In some details, this implementation differs from
the original G-P formulation. This is especially true for the Early effect. See e.g. the
Agilent ADS implementation which allows to select either the simplified UCB SPICE
version, or the original Gummel-Poonm paper model.
For modern transistors with the continuous trend to smaller geometries, second order
effects become more and more important. Due to higher integration and the necessity to
improve the design yield, the need for more precise simulation results and thus to better
models has increased. Many companies have therefore developed in-house models, and in
some cases made them public. Such a model is the Philips MEXTRAM model. It was
developed in 1986 by de Graaff, Klostermann and Jansen.
Later, in 1995, an US industry consortium has proposed a new bipolar model, called
625
VBIC95. Its goal was to become an accepted standard for today's bipolar transistors.
Besides an improved modeling, also including the parasitic PNP transistor of integrated
NPN transistors, the VBIC95 is aimed to be as much as possible similar to the standard G-
P model. Today, it has changed its name to VBIC.
The following list gives the improvements of the VBIC model compared to the SPICE G-P
model:
Precise implementation of the Base width modulation

Parasitic Substrate transistor
Improved Kull model for quasi-saturation
Enhanced delay time modeling
Approximating a distributed Base region
weak avalanche current effects
consistent treatment of the additional phase factor for small signal and time domain
analysis
improved capacitance model
self-heating: improved temperature modeling
Since the VBIC is based on the Gummel-Poon model, and since we want to sketch the
similarities between both models, we will now recap the details of the standard SPICE G-P
model.
A Note on PNP Transistors

The VBIC model is focusing by its own name on a vertical bipolar transistor, i.e. per
definition to a NPN transistor. If applied to a PNP transistor, first check its physical
structure. Then decide, what kind of sub-circuit you may want to use for the modeling.
See /Getreu/.
Lateral PNP
Lateral PNP has a wide Base, therefore, its '' is low, and as a consequence, its 'ft' is low
too.
Substrate PNP
Substrate PNP has usually the following sequence of layers:
E p
B n
C p of wafer
This means, that its Collector is identical to the substrate, i.e. at negative supply voltage.
Such transistors usually feature a bad ''.
Vertical PNP (modeling possible using VBIC)
Its Collector current iC flows vertical, '' is high (because the Base can be made very
thin), and as a consequence, its 'ft' is high.
626
VBIC Background Information
VBIC Model History Notes
Rev.1.2 updates (Sept. 24, 1999)
3-terminal version defined
Base-emitter breakdown model added
Reach-through model added for B-C depletion capacitance
Homotopy version of code added
Limited exponential version added
Completely new code generation added
C, FORTRAN, Verilog-A, Perl, and MAST code provided
Bug in psibi mapping with temperature fixed
Bugs in electrothermal derivatives and solver stamp fixed
DTEMP local temperature difference parameter added
VERS and VREV (version revision) parameters added
NKF high current beta rolloff parameter added
Temperature dependence added to IKF
Ability to select SGP qb formulation added (QBM)
Ability to separate IS for fwd and rev added (ISRR)
Fixed collector-substrate capacitance added (CCSO)
Separate temperature coeffs added for RCX, RBX, RBP
tl node eliminated
POLARITY OF SOME BRANCHES REVERSED FOR VERILOG-A COMPATIBILITY
Ith flows from dt to ground and so is negative
Ixzf flows from xf1 to ground as so is negative of Itzf
Igc component moved into Ibc
Icc broken into forward and reverse components, Itxf or Itzf and Itzr
Release 1.2 Sept. 18, 1999

1.1.5 updates July 28, 1996
Dependence of Irbp on Vbci added to "branch currents"

Itf/Itr renamed Itfi/Itri to avoid name conflicts
Resistor collapse and code bypass condition changed from par = 0 to par <= 0
Branch current and charge dependencies separated for self-heating and no self-
heating
Depletion charge and avalanche routines that provide derivatives for self-heating
added
Self-heating solver and examples added (HBT)
Extra external node added for self-heating to allow coupling of thermal models
between devices
1.1.4 updates
Qbe diffusion term made equivalent to SGP (divide by qb)
Solver example including excess phase added (Icc separated into Itzf|Itxf and Itzr for
this)
Error in sgp_to_vbic in PTF to TD translation fixed
1.1.3 updates
Ith bug fixed and Igc term added
BFN exponent added to 1/f noise
RTH default changed to zero
parameter aliases added
1.1.2 updates
EAI bug fixed in PE/PC/PStemperature mapping
Single->double precision in decomp/solve/vbict/QCDEPL
Scale changed to vscale in solver to avoid name conflict
Avalanche model added, element Igc
Initialization changed in solver
AC solver and AC and temperature tests added
Missing term in derf_Vrci added
Potential numerical problem in Irci fixed
1.1.1 updates
VJ->V bug fixed in qj definition
Potential numerical problems with ITF fixed
Typo derf_Vcci fixed to derf_Vrci in FORTRAN code
SDD - Definition of the VBIC 1.1.4 Model in MDS

Although the toolkit is aimed to use the VBIC implementation in ADS (optional HSpice,
Eldo, Spectre etc.), the following MDS definition of the VBIC model (rev.4. version1.1.4),
627
implemented as "Symbolically Defined Device" (SDD), may be interesting to study.
This example is provided with the toolkit as vbicsdd.inc.
Note
The '\' sign means continuation with the next line.
sdd:CMP1 B 0 C 0 E 0 S 0 \
V_bi 0 V_ci 0 V_ei 0 V_si 0 V_cx 0 V_bx 0 V_bp 0 V_th 0 V_cxf 0 V_rxf 0 \
i[5,0]=ibi \
i[5,1]=qbe+qbc+qbcx \
i[6,0]=ici \
i[6,1]=-qbc \
i[7,0]=iei-ire \
i[7,1]=-qbe-qbex \
i[8,0]=isi-irs \
i[8,1]=qbcp \
i[9,0]=icx-ircx \
i[9,1]=-qbcx \
i[10,0]=ibx-irbx \
i[10,1]=qbex+qbep \
i[11,0]=ibp \
i[11,1]=-qbep-qbcp \
i[12,0]=irth-Ptot \
i[12,1]=qcth \
f[1,0]=ib-irbx \
f[1,1]=-qbeo-qbco \
f[2,0]=ic-ircx \
f[2,1]=qbco \
f[3,0]=ie-ire \
f[3,1]=qbeo \
f[4,0]=vs - irs*RS_T - vsi \
f[13,0]=itzf - itxf \
f[13,1]=-qcxf \
f[14,0]=vcxf - vrxf \
f[14,1]=-flxf
ics:CMP255 V_cx 0 idc=0
ics:CMP256 V_bp 0 idc=0
ics:CMP257 V_bx 0 idc=0
ics:CMP252 V_si 0 idc=0
ics:CMP253 V_ci 0 idc=0
ics:CMP251 V_ei 0 idc=0
ics:CMP250 V_bi 0 idc=0
ics:CMP254 V_th 0 idc=0
ics:CMP270 V_cxf 0 idc=0
ics:CMP271 V_rxf 0 idc=0
ie=_i3
irs=_i4
ic=_i2
ib=_i1
vci = _V6
vbi = _V5
vs = _V4
ve = _V3
vc = _V2
vb = _V1
delt = _V12
vrxf = _V14
vcxf = _V13
PE_T = psibi(PE,EAIE)
PC_T = psibi(PC,EAIC)
PS_T = psibi(PS,EAIS)
GAMM_T = GAMM * (rT ^ XIS * exp(-EA * (1.0 - rT) / vtv))
VO_T = VO * rT ^ XVO
CJCP_T = CJCP * (PS / PS_T) ^ MS
CJEP_T = CJEP * (PC / PC_T) ^ MC
CJC_T = CJC * (PC / PC_T) ^ MC
CJE_T = CJE * (PE / PE_T) ^ ME
NR_T = NR * (1.0 + TNF * (tdev - tini))
NF_T = NF * (1.0 + TNF * (tdev - tini))
IBCNP_T = IBCNP * (rT ^ XIN * exp(-EANS * (1.0- rT) / vtv)) ^ (1.0 / NCNP)
IBCIP_T = IBCIP * (rT ^ XII * exp(-EAIS * (1.0- rT) / vtv)) ^ (1.0 / NCIP)
IBCN_T = IBCN * (rT ^ XIN * exp(-EANC * (1.0- rT) / vtv)) ^ (1.0 / NCN)
IBENP_T = IBENP * (rT ^ XIN * exp(-EANC * (1.0- rT) / vtv)) ^ (1.0 / NCN)
IBEIP_T = IBEIP * (rT ^ XII * exp(-EAIC * (1.0- rT) / vtv)) ^ (1.0 / NCI)
IBCI_T = IBCI * (rT ^ XII * exp(-EAIC * (1.0- rT) / vtv)) ^ (1.0 / NCI)
IBEN_T = IBEN * (rT ^ XIN * exp(-EANE * (1.0- rT) / vtv)) ^ (1.0 / NEN)
IBEI_T = IBEI * (rT ^ XII * exp(-EAIE * (1.0- rT) / vtv)) ^ (1.0 / NEI)
ISP_T = ISP * (rT ^ XIS * exp(-EA * (1.0- rT) / vtv)) ^ (1.0 / NFP)
IS_T = IS * (rT ^ XIS * exp(-EA * (1.0- rT) / vtv)) ^ (1.0 / NF)
RBP_T = RBP * rT^XRC
RS_T = RS * rT^XRS
RE_T = RE * rT^XRE
RBI_T = RBI * rT^XRB
RBX_T = RBX * rT^XRB
RCI_T = RCI * rT^XRC
RCX_T = RCX * rT^XRC
AVC2_T = AVC2 * (1.0 + TAVC * (ftdev - tini))
iei = -icc-ibe-ibex
irbp = vrbp * qbp / RBP_T
irbi = vrbi * qb / RBI_T
irci = iohm / sqrt(1.0 + derf * derf)
derf = IVO * RCI_T * iohm / (1.0 + 0.5 * IVO * IHRCF * (sqrt(vrci * vrci + 0.01)))
iohm = (vrci + vtv * (kbci - kbcx - ln(rkp1))) / RCI_T
rkp1 = (kbci + 1.0) / (kbcx + 1.0)
kbcx = if (RCI==0.0) then 0.0 else sqrt(1.0 + GAMM_T * exp(vbcx / vtv)) endif
kbci = if (RCI==0.0) then 0.0 else sqrt(1.0 + GAMM_T * exp(vbci / vtv)) endif
vbcx = vbci - vrci
qbco = CBCO*vbc
628
qbeo = CBEO*vbe
qbcp = CJCP_T*qdbcp
qbep = TR*itfp+CJEP_T*qdbep
qbcx = QCO*kbcx
qbc = TR*itr+CJC_T*qdbc+QCO*kbci
qbex = CJE_T*qdbex*(1.0-WBE)
qbe = tff*it_f/qb+CJE_T*qdbe*WBE
slTF = if (ITF<=0.0) then 1.0 else 0.0 endif
IHRCF = if (HRCF<=0.0) then 0.0 else 1.0/HRCF endif
IITF = if (ITF<=0.0) then 0.0 else 1.0/ITF endif
IVTF = if (VTF<=0.0) then 0.0 else 1.0/VTF endif
IVO = if (VO_T<=0.0) then 0.0 else 1.0/VO_T endif
IIKP = if (IKP<=0.0) then 0.0 else 1.0/IKP endif
IIKF = if (IKF<=0.0) then 0.0 else 1.0/IKF endif
IIKR = if (IKR<=0.0) then 0.0 else 1.0/IKR endif
IVER = if (VER<=0.0) then 0.0 else 1.0/VER endif
IVEF = if (VEF<=0.0) then 0.0 else 1.0/VEF endif
tff= TF*(1.0+QTF*q1)*(1.0+XTF*exp(vbci*IVTF/1.44)*(slTF+(ritf/(ritf+1.0))^2)*sgitf)
ritf= it_f*sgitf*IITF
sgitf= if (it_f>0.0) then 1.0 else 0.0 endif
ibc = diode(vbci,IBCI_T,NCI)+diode(vbci,IBCN_T,NCN)
ibex = (1.0-WBE)*(diode(vbex,IBEI_T,NEI)+diode(vbex,IBEN_T,NEN))
ibe = WBE*(diode(vbei,IBEI_T,NEI)+diode(vbei,IBEN_T,NEN))
iccp = (itfp-itrp)/qbp
qbp = 0.5*(1.0+sqrt(1.0+4.0*q2p))
q2p = itfp*IIKP
itrp = diode(vbcp, ISP_T, NFP)
itfp = ISP_T*(WSP*exp(vbep/(NFP*vtv))+(1.0-WSP)*exp(vbci/(NFP*vtv))-1.0)
CEP = if (TD<=0.0) then 0.0 else TD endif
LEP = if (TD<=0.0) then 0.0 else TD/3.0 endif
icc= if (TD>0.0) then itxf-itzr else itzf-itzr endif
itxf = vrxf
qcxf = CEP*vcxf
flxf = LEP*itxf
itzr = itr/qb
itzf = it_f/qb
qb = 0.5*(q1+sqrt(q1*q1+4.0*q2))
q2 = it_f*IIKF+itr*IIKR
q1 = 0.5*(sqrt((q1z-0.0001)^2+0.0001*0.0001)+q1z-0.0001)+0.0001
q1z = 1.0+qdbe*IVER+qdbc*IVEF
itr = diode(vbci, IS_T, NR_T)
it_f = diode(vbei, IS_T, NF_T)
rT = tdev / tini
tini = TNOM + tabs
qdbcp = qj(vbcp,PS_T,MS,FC,AJS)
qdbc = qj(vbci,PC_T,MC,FC,AJC)
qdbep = qj(vbep,PC_T,MC,FC,AJC)
qdbex = qj(vbex,PE_T,ME,FC,AJE)
qdbe = qj(vbei,PE_T,ME,FC,AJE)
icx = irci+irbp
vrbp = vcx-vbp
vrs = vs-vsi
vbcp = vsi-vbp
vbep = vbx-vbp
vre = ve-vei
vrbi = vbx-vbi
vrbx = vb-vbx
vbex = vbx-vei
vrci = vcx-vci
vrcx = vc-vcx
vei = _V7
vbp = _V11
vbx = _V10
vcx = _V9
vsi = _V8
tabs = 273.15
qq = 1.602189E-19
kb = 1.380662E-23
qsingle(v,p,f,a,m) = ql0(v,p,f,a,m)+(1.0-f)^(-m)*(v-vl(v,p,f,a)+vl0(p,f,a))-q0(p,f,a,m)
vl(v,p,f,a) = 0.5* (dvh(v,p,f)-mv(v,p,f,a))+p*f
mv(v,p,f,a) = sqrt(dvh(v,p,f)*dvh(v,p,f)+a)
ql0(v,p,f,a,m) = -p*(1.0-vl(v,p,f,a)/p)^(1.0-m)/(1.0-m)
q0(p,f,a,m) = -p*(1.0-vl0(p,f,a)/p)^(1.0-m)/(1.0-m)
vl0(p,f,a) = 0.5* (dv0(p,f)-mv0(p,f,a))+p*f
mv0(p,f,a) = sqrt(dv0(p,f)*dv0(p,f)+a)
dv0(p,f) = dvh(0,p,f)
qspice(v,p,m,f) = qlo(v,p,f,m)+qhi(v,p,f,m)
qhi(v,p,f,m) = if (dvh(v,p,f)>0.0) then dvh(v,p,f)*(1.0-f+0.5*m*dvh(v,p,f)/p)/((1.0-f)^(1.0+m))
\
else 0.0 endif
qlo(v,p,f,m) = if (dvh(v,p,f)>0.0) then p*(1.0-(1.0-f)^(1.0-m))/(1.0-m) \
else p*(1.0-(1.0-v/p)^(1.0-m))/(1.0-m) endif
dvh(v,p,f) = v-f*p
qj(v,p,m,f,a) = if (a<=0.0) then qspice(v,p,m,f) else qsingle(v,p,f,a,m) endif
vtv = kb * tdev / qq
diode(v,is,n) = is*(exp(v/(n*vtv)) - 1)
vbci = vbi-vci
vbc = vb-vc
vbe = vb-ve
vbei = vbi-vei
isi = ibcp-iccp
ibp = -ibep-irbp-ibcp
ibx = ibex+irbi+ibep+iccp
psiio(p) = 2.0 * vtv * ln(exp(.5 * p / vtv)-exp(-0.5 * p / vtv))
psiin(p,e_a) = psiio(p) * rT - 3.0 * vtv * ln(rT) - e_a * (rT - 1.0)
psibi(p,e_a) = psiin(p,e_a) + 2.0 * vtv * ln(0.5 * (1.0 + sqrt(1.0 + 4.0 * exp(-psiin(p,e_a) /
vtv))))
tdev = if (RTH>0) then TAMB + tabs + delt else TAMB + tabs endif
Ptot = if (RTH>0) then ibe*vbei+(ibc-igc)*vbci+icc*(vbei-vbci)+ibep*vbep+ibcp*vbcp+iccp* \
629
(vbep-vbcp)+ircx*vrcx+irci*vrci+irbx*vrbx+irbi*vrbi+ire*vre+irbp*vrbp+irs*vrs+ibex*vbex
\
else 0.0 endif
irth = if (RTH>0) then delt / RTH else 0.0 endif
qcth = if (CTH>0) then CTH * delt else 0.0 endif
igc = if (AVC1==0.0) then 0 else (itzf-itzr-ibc)*avalm(vbci,PC_T,MC,AVC1,AVC2_T) endif
ibi = ibe+ibc-igc-irbi
ici = icc-ibc+igc-irci
vla(v,p) = 0.5 * (sqrt((p-v)^2 + 0.01) + (p-v))
avalm(v,p,m,av1,av2) = av1 * vla(v,p) * exp(-av2 * vla(v,p)^(m - 1.0))
ibep= if (IBEIP==0.0 && IBENP==0.0) then 0.0 else /
diode(vbep,IBEIP_T,NCI)+diode(vbep,IBENP_T,NCN) endif
ibcp= if (IBCIP==0.0 && IBCNP==0.0) then 0.0 else /
diode(vbcp,IBCIP_T,NCIP)+diode(vbcp,IBCNP_T,NCNP) endif
irbx = if (RBX_T <= 0.0) then 1E-15 else (vb-vbx)/RBX_T endif
ircx = if (RCX_T <= 0.0) then 1E-15 else (vc-vcx)/RCX_T endif
ire = if (RE_T <= 0.0) then 1E-15 else (ve-vei)/RE_T endif
R:CMP272 0 V_cxf r=1E20 OH
R:CMP273 0 V_bx r=1E20 OH
R:CMP274 0 V_bi r=1E20 OH
R:CMP275 0 V_rxf r=1E20 OH
R:CMP276 0 V_th r=1E20 OH
R:CMP277 0 V_cx r=1E20 OH
R:CMP278 0 V_bp r=1E20 OH
R:CMP279 0 V_ei r=1E20 OH
R:CMP280 0 V_ci r=1E20 OH
R:CMP281 0 V_si r=1E20 OH
SIMPLIFIED VBIC EQUATIONS, rev. 1.1.4:
Main transistor Collector current without the case check for

parameter=0, and without temperature effects
ici = icc - ibc + igc - irci

>> icc= itxf - itzr
itxf = vrxf, with vrxf = _V14
itzr = itr/qb
diode(v,is,n) = is*(exp(v/(n*vtv)) - 1)
qb = 0.5*(q1+sqrt(q1*q1+4.0*q2))
>> ibc = diode(vbci,IBCI_T,NCI)+diode(vbci,IBCN_T,NCN)
igc = itzf - itzr - ibc) * avalm(vbci,PC_T,MC,AVC1,AVC2_T)
itzf = it_f/qb
it_f = diode(vbei, IS_T, NF_T)
itzr = itr/qb
ibc =
diode(vbci,IBCI_T,NCI)+diode(vbci,IBCN_T,NCN)
>> irci = iohm / sqrt(1.0 + derf ^2)
iohm = (vrci + vtv * (kbci - kbcx - ln(rkp1))) /
RCI_T
vrci = vcx-vci
Main transistor Base current without the case check for

parameter=0, and without temperature effects
Different to the SGP model, the Base-Emitter current is defined independent from the
Collector-Emitter current. No beta parameter is used (i.e. no BF wit the VBIC model). The
Base Emitter current consists of an ideal and an nonideal part, both devided into an
intrinsic (Ibe ) and an extrinsic (Ibex ) part by the geometrical parameter WBE:
For version 1.1.4, there is:
and
The Base-Collector current is defined independent of the inverse transport a current as

well:
Paras transistor Collector current without the case check for parameter=0, and without
temperature effects
Ifp = Isp * (WSP*exp(vbep/(NFP*vt))+(1-WSP)*exp(vbci/(NFP*vt))-1)
Information on the VBIC Code, Release 1.2.

For a short synopsis of the changes see the chapter "History".
Here is a list of the major changes in version 1.2:
1. The name is now VBIC and not VBIC95.

2. The thermal network has been returned to its original form, which was how it was
implemented in all simulators anyway. The "tl" node was incorrect, the Ith current
had to circulate from dt to tl and so could not allow tl to funtion as a coupling node.
Ith has to have one end grounded.
Note
This means that the value of RTH used for single device self-heating differs from that used when a
thermal network couples more than one device.
3. All of the model additions agreed to at the BCTM meetings have been implemented
630
temperature dependence of IKF
separate temperature coefficients for intrinsic and extrinsic resistances
a 3 terminal version
base-emitter breakdown model (simple exponential)
reach-through model to limit base-collector depletion capacitance
VERS version parameter added (also VREV for version revision)
separate activation energy added for ISP
1. Additional changes were made based on feedback from many sources
errors in solvers and derivatives for electrothermal model fixed

simple continutation added to improve solver convergence
QBM parameter add to switch to SGP qb formulation
NKF added to parameterize beta(Ic) high-current rolloff
fixed collector-substrate capacitance added (CCSO)
for HBTs, ISRR added to allow separate IS for reverse operation
an error in the built-in potential temperature mapping was fixed
code bypass for efficiency, if some parameters are zero
limited exponential version provided
the transport current Icc was explicitly separated into forward and reverse
components
1. The automated code generation has been completely rewritten. All code, including
solvers, is now generated. Solvers exist for all combinations of the code.
2. IMPORTANT: note that the polarities of some of the current branches have changed.
This was necessary because Verilog-A supports (or appears to support) branches to
ground referenced from a node to ground, and not from ground to a node.
The Ith and Itzf branches in the thermal and excess phase networks are now defined
as the negative of what they were, but the connection polarity is switched. Ith is now
negative, but flows from dt to ground. This must be taken into account when setting
up the matrix stamp properly.
Equivalent Circuit Network for VBIC 1.2
Default Parameters rev.1.2

VERS 1.2
VREV 0.0
Parameter Extraction Temperature
TNOM 27.0
Local Temp. Dependence
DTEMP 0.0
CV
FC 0.9
CBEO 0.0
CJE 0.0
PE 0.75
ME 0.33
AJE -0.5
CBCO 0.0
CJC 0.0
QCO 0.0
CJEP 0.0
PC 0.75
MC 0.33
AJC -0.5
CJCP 0.0
PS 0.75
MS 0.33
AJS -0.5
CCSO 0.0
Resistors
RCX 0.0
RCI 0.0
VO 0.0
GAMM 0.0
HRCF 0.0
RBX 0.0
RBI 0.0
RE 0.0
RS 0.0
RBP 0.0
Early Voltage
VEF 0.0
VER 0.0
Main Transistor
IS 1.0e-16
NF 1.0
NR 1.0
IBEI 1.0e-18
WBE 1.0
NEI 1.0
IBEN 0.0
NEN 2.0
IBCI 1.0e-16
631
NCI 1.0
IBCN 0.0
NCN 2.0
AVC1 0.0
AVC2 0.0
IKF 0.0
NKF 0.5
IKR 0.0
ISRR 1.0
Parasitic Transistor
ISP 0.0
WSP 1.0
NFP 1.0
IBEIP 0.0
IBENP 0.0
IBCIP 0.0
NCIP 1.0
IBCNP 0.0
NCNP 2.0
IKP 0.0
Thermal Model
RTH 0.0
CTH 0.0
Transit Time
TF 0.0
QTF 0.0
XTF 0.0
VTF 0.0
ITF 0.0
TR 0.0
TD 0.0
Flicker Noise
KFN 0.0
AFN 1.0
BFN 1.0
Select SGP qB Formulation
QBM 0.0
Temperature & Misc.
EA 1.12
EAIE 1.12
EAIC 1.12
EAIS 1.12
EANE 1.12
EANC 1.12
EANS 1.12
XIS 3.0
XII 3.0
XIN 3.0
TNF 0.0
TAVC 0.0
VRT 0.0
ART 0.1
XRE 0
XRBI 0
XRCI 0
XRS 0
XRCX 0
XRBX 0
XRBP 0
XIKF 0
XVO 0
XISR 0.0
DEAR 0.0
EAP 1.12
VBBE 0.0
NBBE 1.0
IBBE 1.0e-6
TVBBE1 0.0
TVBBE2 0.0
TNBBE 0.0
EBBE 0.0
632
Recapitulating the Gummel-Poon Model
The figure GP-1 depicts the equivalent schematic of the Gummel-Poon large signal model.
GP-1: Equivalent Schematic of the Gummel-Poon Large Signal Model
Definitions
Currents are always considered to flow into the device. This means, for example, that IC is
a current into the Collector of the transistor. Voltages are indexed with their reference
nodes. In order to have a clear notation, voltage drops across parasitic resistors are only
considered, if this is required to better understand a specific detail or to prevent from
confusion. This means, that vBE stands usually for a voltage between inner Base and inner
Emitter vB'E'. Similarly, the parameter dependencies of temperature are neglected.
Finally, all notation is based on a NPN transistor.
The Normalized Base Charge Qb

An essential detail for the calculation of the DC, as well as the AC performance, is the
majority carrier Base charge normalized to its value without bias.
q b can be calculated as:
with
covering the Early effect (Base width modulation) and
covering the Webster effect (high current behavior).

The following simplifications are applied to all common implementations of the G-P model,
see e.g. the UCB SPICE version.
Equation (GP-2) is approximated by:
and charge q 1 is approximated by:
in order to obtain, similarly to the Ebers-Moll model, a constant output conductance.

Besides the approximation 1+x ~ 1/(1-x) for x<<1, equation (GP-6) contains especially
the assumption of invariable and constant space charge capacitors. The second
approximation in equation (GP-6) can lead to rather big modeling errors at low Early
voltages. Therefore, as mentioned earlier, the simulator ADS of Agilent Technologies
allows the user to select between one of these approximations.
DC Performance
The G-P Collector current IC comprises of a forward and reverse component following
633
Base width modulation and high current effects are modeled by the bias-depending Base
charge q b.
The Base current is composed of an ideal (I bei , I bci) and an non-ideal part (I ben , I bcn).
The latter covers the recombination current, a low-bias effect of I b.
In reverse mode, the Base current is described by
Figures GP-2 and GP-3 visualize the most important DC effects covered by the SPICE G-P
model.
GP-2: Typical output characteristic of a bipolar transistor
GP-3: Gummel plot of a bipolar transistor
Resistors
The parasitic resistors Re and Rc are assumed to be constant. R b, however, is modeled
bias-dependent. In the SPICE G-P model, it is implemented as:
with
Dynamic Behavior
The capacitors between Base and Emitter, as well as between Base and Collector,
represent each the space charge (depletion) and diffusion capacitor.

The space charge capacitors are described by:
To prevent from the pole in equation (GP-11) at vi = VJE , a linear continuation of (GP-11)
is used for voltages vi > FC*VJE. The Base-Collector capacitor Cjc is distributed between
inner and outer Base node of the model by:
and
634
In addition to the capacitors depicted in fig. GP-1, an additional capacitor between
Collector and Substrate is added to the SPICE G-P model. This substrate capacitor is either
considered to be a constant or it is also modeled after equation (GP-11).
Diffusion Charge Capacitors:
The other charge, stored with the diffusion capacitors, is calculated by:
and for B-C:
where IF stands for the current through Collector-Emitter, and IR for Emitter-Collector,
see equation (GP-7). This means for the capacitors
and
The transit time TFF itself is modeled by the empirical equation
The figure GP-4 shows the trace of TFF vs. the Collector current Ic with vbc as secondary
sweep
Fig. GP-4: Trace of TFF (I c, vbc)
Since the measured phase shift of Ic is usually bigger than what is covered by the model
equations, the UCB SPICE G-P model has an additional parameter PTF. When simulating in
the frequency range, the additional phase shift is added to the phase of Ic. For the
transient simulation in the time domain, however, a different implementation with Bessel
function is used in UCB SPICE.
From fig. GP-1, the small signal schematic for high frequency simulations can be derived.
This means, for a given operating point, the DC currents are calculated and the model is
linearized in this point, see the figure below. Such a schematic is used later for SPICE S-
parameter simulations.
It must be noted that this schematic is a pure linear model. It cannot be used to predict
non-linear high-frequency behavior of the transistor. In order to do this, RF simulators like
ADS apply harmonic-balance techniques to perform nonlinear RF large signal simulations.
GP-5: AC small signal schematic of the bipolar transistor
Note
XCJC effect neglected.
A Quick Tutorial on the Gummel-Poon Parameters
635
636
And, finally model the dependence of fT on VCE:
Gummel-Poon Model Parameter List

The following table compiles the G-P model parameters and their SPICE default values.
These default values basically switch off the effect covered by these parameters. This
enables to simulate a transistor behavior also in the case of not knowing all the model
parameters.
However, these default values should not be confused with typical parameters!
637
638
The VBIC Model
The following figure (VBIC-1) shows the complete equivalent schematic of the VBIC
model. Besides the main NPN transistor between Base, Collector and Emitter, there is an
additional parasitic PNP transistor between Base, Collector and Substrate. An additional
circuit is added to cover self-heating (Rth,Cth). Also, an extra circuit for adding phase shift
to Icc is available. The resistances Rc, Rbi/qb and Rbip/qbp are modeled non-linearily.
Therefore, they should rather be considered as nonlinear voltage dependent current
sources.
VBIC-1: Equivalent schematic of the VBIC model
A note on the current flow direction of the ICP source of the parasitic PNP. As can be seen
from the main NPN transistor, the transport current is always in the same direction like its
driving currents, i.e. the diode currents Ibei and Iben. For the parasitic PNP, this has to be
true also: flowing from its Emitter to the Collector (what is the common forward operating
condition for a PNP!).
We will now look into the details of the VBIC formulation. Thermal effects will be
neglected. Currents will be indexed with their corresponding node names and flowing into
the nodes. Voltages, indexed by their two node names, will additionally be named by
abbreviations like 'i' for internal, 'x' for external, 'o' for outer and 'p' for parasitic.
The following formulas refer to VBIC rev.1.1.4
The Normalized Base Charges q b and q bp
Since the VBIC has one of its major roots in the Gummel-Poon model, it is again the Base
charge qb, which is one of the most important internal model parameters. However, with
the VBIC, approximations like in the SPICE G-P model are not used at all.
As a consequence, this allows a bias dependent modeling of the output conductance g CE !
In the Gummel-Poon model, with its UCB-SPICE definition of the Early voltages, g CE is a
constant !
Back to q b , this charge is implemented as:
with q1 as a function of the space charge capacitors
and
with the temperature voltage VT after (VBIC-29). Here, q je and q jc are the normalized
charges of the space charge (depletion) capacitors Cje and Cjc. This implies that the space
charge capacitors have to be modeled before the Early voltages are extracted!
For the parasitic PNP transistor, its Base charge q bp is given by:
639
yet neglecting the Early effect!
The charge
covers the Webster effect of the parasitic PNP transistor, but only in forward bias mode.
Coming back to the space charge capacitors, they can be modeled either after the SPICE
G P model, or, depending on the switch parameter AJx, after a VBIC-specific new
approach. For the case of the SGP model, and in reverse bias (Vx < FC * Px) the
normalized depletion charge function is defined as:
where Px represents the built-in potential, Mx the junction grading coefficient and Vx the
applied voltage of the appropriate junction.
For a forward biased junction ( Vx > FC * Px ) the depletion charge function is given by:

Integrating the charges from above, the Base-Emitter space charge capacitance is given
for vBE < FC*PE by
and else,
The same formula applies also to the other two space charges capacitors.
For these inner space charge capacitors, there is, besides this SPICE G-P formulation, also
an alternate continuous formula implemented in VBIC. This alternate method requires no
linear continuation in order to avoid the pole in the SPICE-like equation. The model
parameters AJx with x=C, E or S, select the preferred model (-0.5, the default value,
selects the G-P formula). The Base-Emitter space charge capacitance is, in analogy to the
partitioning of the Base-Emitter current by the DC parameter WBE (see further below),
distributed between internal and external Base.
The Base-Collector space charge capacitance is distributed between the capacitances C jc

(Base and Collector of the main transistor) and Cjep (Base and Emitter of the parasitic
one). Another space charge capacitance, C jcp, is located between Base and Collector of
the parasitic PNP. It corresponds to the substrate capacitance in the SPICE G-P model.
DC Performance of the Main NPN Transistor

NPN Collector-Emitter Current
The Collector current source I CC is determined, similar to the SPICE G-P model, by the
two transport current terms I TF and I TR
with the temperature voltage VT after (VBIC-29).

This means that the Collector-Emitter current in forward operation is given by:
and in reverse operation by:
NPN Base-Emitter Current

Like with the SPICE Gummel-Poon model, the Base-Emitter current covers the ideal and
non-ideal (recombination) behavior. However, with the VBIC model, it is not coupled to
the Collector current ICC like with the Gummel-Poon model, where we have the model
parameters BF and BR. Here, two separate, parallel diode currents are used instead, one
for the 'ideal' part (index I ) and the other for the 'non-ideal' part (index N) of the Base-
Emitter current:
640
Note
The VBIC parameter naming is very mnemonic.
I ideal part of the Base current
N non-ideal part of the Base current
X external part of the Base current
The VBIC model features a split of the Base-Emitter current 'to the right and to the left' of
the internal Base resistor RBI. This means that the very inner Base-Emitter current from
node 'Bi' to 'Ei' is given by:
and the outer Base-Emitter current from the node 'Bx' to 'Ei' is given by:
Note
WBE, does not affect the DC fitting, but the S-parameter fitting instead.
NPN Base-Collector Current

The Base-Collector current is calculated in analogy to (VBIC-13):
With the VBIC, a weak avalanche current is described by
It overlays the current iB (see iB vs. vCE in the foutput setup), what then shows up in the
forward output characteristics of iC vs. vCE.
DC Performance of the Parasitic PNP Transistor

PNP Collector-Emitter Current
The current Icp of the parasitic PNP transistor, again, is the difference between the
forward and reverse transport currents Itfp and Itrp.
However, the control of the forward transport current Itfp is split into the voltages vBCI of
the NPN and vBEP of the parasitic PNP transistor:
The parameter WSP controls that splitting. Its default value is 1.

WIth WSP=0, a second knee in the log(iB) vs. vCE and log(iS) vs. vCE is modeled, with
WSP=1, the log(iB) and log(iS)drop in a single step, see file VBIC_TUTOR.mdl for details.
The reverse transport current is defined by:
Note once again the VBIC mnemonic:
P parasitic PNP transistor
PNP Base-Emitter Current

The Base-Emitter current is, like with the main NPN, split into an ideal and non-ideal part
in forward and reverse operating mode.
Note
Since the parasitic Base-Emitter is identical to the Base-Collector of the main NPN transistor, the reverse
emission coefficients of the 'MAIN NPN', i.e. NCN and NCI, are also used for the parasitic forward Base
current formula.
PNP Base-Collector Current
641
Finally, the Base-Collector current for reverse operation of the PNP is
Resistors
The parasitic resistors RE of the Emitter and RS of the Substrate contact are modeled with
a constant value. The Base resistance, however, is comprised of a constant part RBX and
a variable part RBI/qb
with qb after equations (VBIC-1) ..(VBIC-3)

The parasitic PNP Base resistance, RBIP is modeled by:
with qbp after equations (VBIC-4) ..(VBIC-5)
The Collector resistance, different to the G-P model, now also consists of a constant part
RCX and a current-dependent, non-linear part RCI. The later models the quasi-saturation
of the output characteristics, see below.
Quasi-Saturation
Quasi-saturation, as implemented in VBIC, is essentially based on the Kull model. This
means that the current through the resistor RCI is depending on the inner and outer
Collector voltage. With some smaller modifications to the Kull model, which essentially
refer to the modeling of the velocity saturation at high voltages, the current Irci through
resistor RCI is given by
In order to better understand this complex formulation, we consider its terms individually
after private communication with J.Berkner, Infineon, Munich/:
First of all, the slope for low vBE is determined, like with the Gummel-Poon model, by the
external Collector resistor RCX. For higher bias levels, the internal resistor RCI comes into
play. This resistor RCI is bias dependent, while RCX is not. The bias dependency of RCI is
related to both, the voltage drop across it, plus the current thru it.
We commence with the dependency versus voltage. To simplify things, we start with the
ohmic law, applied to RCI and considering vrci = vCx - vCi.
This means, the voltage drop across RCI is depending on a correction voltage. For non-
saturated bias, Vcorr = 0, and the internal Collector resistor is constant and equal to RCI.
When quasi-saturation occurs: Vcorr > 0. This means that Iohm increases and the
effective internal Collector resistor is therefore reduced.
This correction voltage is given by:
with the coefficients
642
and
with
In addition to this reduction of the effective, internal Collector resistor, its value can
further be reduced due charge carrier velocity saturation, and we obtain equation VBIC-25
from above.
Simplified for HRCF -> infinite, VBIC-25 reduces to
From this simplification, we see that Irci is further reduced if Iohm * RCI ~ VO, i.e. for
increasing values of Iohm.
Dynamic Behavior
The additional charge caused by quasi-saturation (Base widening), also influences the
dynamic behavior of the transistor (S-parameter). This is covered by parameter QCO
which models the charges stored with the capacitances Cbcx and Cbcq.
Therefore, QCO affects the fT modeling for low vCE, i.e. in quasi-saturation.
with Kbcx and Kbci from (VBIC-27) and (VBIC-28).

Besides the quasi-saturation charges of (VBIC-30) and (VBIC-31) and the constant
parasitic capacitors CBC0 and CBE0, the model features for both, the main NPN and the
parasitic PNP transistor, individual space charge and diffusion capacitors.

Since this capacitor contributes to the Early effect modeling, it has been already discussed
in the very first section of this chapter.
Diffusion Charge Capacitors of the NPN Transistor:

The diffusion capacitance of the main transistor are modeled basically in analogy to the G-
P model. The charge, stored in the capacitors Cbe and Cbc, is calculated by
and for B-C:
The forward transit time is calculated by
very similar to the SPICE G-P model.

The term (1+QTFq1) covers additionally a dependency of the Base width modulation.
Diffusion Charge Capacitors of the PNP Transistor

Related to the parasitic PNP transistor, the model features only one diffusion capacitance
Cbep, which is described by
Additional Phase Shift

The additional phase shift is calculated with the VBIC model by a separate network. From
the parameter TD, the phase is generated by a Bessel function of second order. The
advantage of this method is essentially the consistent description of additional phase in
the small signal and transient analysis.
643
Self-Heating Modeling
The most important difference to the Spice Gummel-Poon is that with VBIC, the
temperature is no longer a static 'environment parameter'. Based on the additional
temperature network consisting of the thermal resistor RTH and the temperature time
constant 1/(RTH*CTH) , the actual, bias-dependent temperature conditions of the
transistor are reflected.
The temperature modeling is comparable to SPICE G-P. However, some enhancements
have been included. See the sdd-Definition further below.
VBIC Model Parameter List

Compared to the SPICE Gummel-Poon model, we have to handle a lot more parameters
with the VBIC. However, please keep in mind that the naming of the VBIC parameters is
really very mnemonic:
I Ideal part of the Base current

N Non-ideal part of the Base current, recombination effect, below the 'knee'
P A model parameter of the parasitic PNP
With this in mind, the VBIC can be understood very easily with its similarity to the G-P
model.
Note
If RTH is specified, CTH=0 (the common default value) is not allowed. Otherwise, the transistor operated
e.g. in pulsed mode, would react with self-heating, faster than its electrical delay.
644
Comparing the VBIC and Gummel-Poon
Parameters
The following table gives a comparison between the VBIC and the G-P model parameters.
Table: Comparison between VBIC and Gummel-Poon Parameters
VBIC G-P Remarks

Parasitic - -
Capacitors
CBE0 - External capacitors, not included in the G-P model
CBC0 - External capacitors, not included in the G-P model
Space Charge - -
Capacitors
AJE - AJE selects one of the space charge models (-0.5 for G-P version)
CJE CJE -
PE VJE -
ME MJE -
AJC - AJC selects one of the space charge models (-0.5 for G-P version)
CJC CJC * XCJC -
PC VJC -
MC MJC -
CJEP CJC*(1-XCJC) -
CJCP CJS GP only includes the CV path from C->S, however not the DC C->S diode
PS VJS -
MS MJS -
AJS - AJS selects one of the space charge models (-0.5 for G-P version)
FC FC Default for VBIC: 0.9; at G-P: 0.5
WBE - WBE distributes the Base-Emitter current space charge capacitor
between inner and outer Base
Early Modeling - -
VEF VAF The Early effect is modeled differently
VER VAR With the VBIC model
DC Forward
Main Transistor
IS IS -
NF NF -
Forward Base:
IBEI IS / BF With the VBIC, the forward Base current is not coupled to the Collector
current. Instead, two parallel diodes, one to model the ideal (I) Base
current and another to cover the non-ideal or recombination (N) effect, are
used.
NEI NF
IBEN ISE
NEN NE
IKF IKF
VBIC G-P Remarks
DC reverse Main Note: IS from the forward Collector current model is used.
Transistor
NR NR -
IBCI - IS / BF With the VBIC model, the reverse Base current is not coupled to
the Emitter current. Instead, two parallel diodes, one to model the ideal (I)
Base current and another to cover the non-ideal or recombination effect
(N) effect, are used.
NCI NR -
IBCN ISC -
NCN NC -
IKR IKR -
Parasitic
Transistor
ISP - G-P does not cover a parasitic transistor
NFP - -
WSP - Distributes parasitic collector current control to vbci of the main transistor
and vbep of the parasitic transistor
Forward Base - -
BEIP - For the exponential coefficient, NCI of the main transistor is used
IBENP - For the exponential coefficient, NCN of the main transistor is used
Reverse Base
IBCIP - With the VBIC model, the reverse Base current is not coupled to the
Collector current. Instead, two parallel diodes, one to model the ideal (I)
Base current and another to cover the non-ideal or recombination effect
(N), are used.
NCIP - -
IBCNP - -
NCNP - -
IKP - -
645
Avalanche Effect - -
AVC1 - -
AVC2 - -
Resistances - -
RE RE -
RBX RBM The bias-dependent Base resistance is modeled differently in both models.
RBI RB - RBM -
- RB -
- IRB -
RS - -
RBP - -
RCX RC Constant, external Collector resistance
VBIC G-P Remarks
Quasi-
Saturation
CI - VBIC has a modified Kull model implemented
GAMM - -
VO - -
HRCF - -
QCO - -
Transit Time
Modeling
TF TF -
QTF - Describes the additional dependency of the transit time from qb.
XTF XTF -
ITF ITF -
VTF VTF -
TR TR -
Excess Phase - -
TD *TF*PTF/180 The VBIC implementation is consistent between small signal and transient
analysis
Temperature - -
Dependence
CTH - VBIC includes self-heating effects
RTH - -
TAMB - Environmental temperature
TNOM TNOM Measurement temperature for parameter extraction
EA EG The G-P model only contains one energy gap
EAIE - -
EAIC - -
EAIS - -
EANE - -
EANC - -
EANS - -
XRE - Temperature coefficients of the resistors are not covered with G-P
XRB - -
XRC - -
XRS - -
XVO - -
XIS XTI -
XII XTB -
XIN XTB -
TNF - -
TAVC - -
Since the VBIC model is based essentially upon the G-P model, most of the G-P
parameters can be converted to VBIC. However, the following details have to be kept in
mind strictly:

For all space charge capacitances, the parameter Ajx with x = E, C and S must be set
to a value less than or equal zero, in order to obtain the same formulation in both
models. In this case, the values of the Base-Emitter capacitance can be transferred.
With the CJC parameter of the Base-Collector capacitance, it must be reflected that
and If the actual SPICE implementation of the G-P model only contains a constant
value for the Substrate capacitance, CJCP will hold that value and MS is set to 0.
Finally, FC, modeling the transition between the hyperbolic formulation and the linear
continuation, has different default values in both models.
Diode Currents
The forward and reverse parameters IS, NF and NR can be transferred directly. The
ideal Base current sections are not coupled to the transport currents. For VBIC, there
is IBEI = IS/BF and IBCI = IS/BR . The G-P parameters of the non-ideal or
recombination section of the Base current can, however, be transferred directly to
VBIC. Finally, setting WBE =1 , the Base current distribution (inner and outer Base)
is switched off.
646
Early Modeling
The implementation of the Early effect in the Gummel-Poon and the VBIC model is so
different, that the parameters cannot be converted (different modeling of the
normalized Base charge qb). Especially with small G-P Early voltages, the resulting
error can be considerably big. Yet, as a general rule, the G-P Early parameters are
usually bigger than those of the VBIC model. For rather big values of the G-P Early
voltage, only slight modifications should be required.
Parasitic Transistor and Avalanche effect
Using the mentioned default values in VBIC, the parasitic transistor and the
Avalanche effects are switched off. Exception: Base-Collector space charge
capacitance.
Resistances
The constant resistors of Emitter and Collector can be overtaken. From the
parameters RBM, RB and IRB of the G-P model, suitable values have to be generated
for the VBIC parameters RBI and RBX, since the models differ here.
Quasi-Saturation
Using suitable parameter values, no quasi-saturation effects are taken into account
with the VBIC. Setting GAMM = 0, Rc is reduced to an ohmic resistance. QC0 = 0
eliminates the influence of the additional capacitances Cbcx and Cbcq.
Transit Time Parameters
Setting QTF = 0, the G-P parameters of the transit time TFF as well as the excess
phase can be transferred without any change.
Setting Rth = 0, the VBIC temperature model including self-heating is reduced to G-
P, which only covers a constant ambient temperature.
647
VBIC Modeling Strategy
For a good modeling result, it is essential to follow a good parameter extraction strategy.
It is especially important to follow a certain sequence of extractions, since most model
parameters depend on each other. Usually, the first parameters to be extracted are those,
which do not or only lightly depend on others. Then, when proceeding through the
extraction strategy, the more nested parameters are extracted subsequently, and the
model fits more and more accurately.
Related to the VBIC model, the following parameter extraction sequence is proposed. See
also the VBIC toolkit VBIC_EXTRACT.mdl file and its macros.
CV
Since the Base charge is the basic relationship of the VBIC Early effect description, it is
the space charge capacitors which have to be modeled first.
DC
First, the ohmic parasitics are extracted from specific measurement setups.
Then, the Early voltages are extracted from the DC output characteristics for non-
quasisaturation and no avalanche effect. The parameters, however, are not yet optimized.
This is due to the fact that the other DC parameters are not yet know.
We will optimize the Early parameters after the fitting of the Gummel plots. The diode
parameters ISx and Nx as well as the knee currents IKx of the main NPN transistor are
extracted from forward and reverse Gummel-Poon measurements. The transistor should
not be in quasi-saturation.
From measurements with either an open Emitter contact, or vEC=0 for the main NPN, the
diode parameters ISx and Nx and the knee current IKP of the parasitic PNP transistor
are extracted. The quasi-saturation parameters, except QC0, are calculated from the
output characteristics iC(vCE, vBE), as well as the avalanche parameters, and the
selfheating parameter RTH
DC Finetuning
Finally, the output characteristics, especially the quasi-saturation region fitting, is fine-
tuned by optimization.
S-parameters
The transit time parameters TF, XTF, ITF, VTF are extracted from S-parameter
measurements in setup 'tf_vbe_vce' with non-quasi-saturated bias conditions. Since the
dynamic model description for this bias conditions are identical to the SPICE Gummel-
Poon model, the same extraction strategy is applied here too. For this condition, we set
QTF=0
In VBIC, the Base resistance is modeled by RBB'=RBX + RBI/qb. Since qb represents a

quite complex formula (see equ.(VBIC-1 ff.), the 'input-impedance-circle method' from the
Gummel-Poon model cannot applied easily to get a starting value for the inner Base
resistance RBI. Therefore, RBI is obtained by optimizing the S11 plot of setup
'biased_Spar'.
The S-parameter quasi-saturation parameter QCO, which affects the high frequency
performance, is determined from S-parameter measurements under quasi-saturation DC
bias condition. This is done in setup 'biased_Spar', referring only to the highest vBE,
including the saturated vCE bias condition.
S-Parameter Finetuning
The S-parameter fitting for all DC bias conditions is fine-tuned using optimization.
648
Measuring and Extracting the VBIC
Parameters
Measurements
For the measurements, please refer to IC-CAP file BJT_MEAS_MASTER.mdl.
Perform the measurements of all setups, and execute finally macro 'EXPORT_DATA' in
order to generate mdm files for import into file VBIC_EXTRACT.mdl.
Parameter Extractions
The IC-CAP model file VBIC_EXTRACT.mdl contains the steps described in this chapter.

CJx, Mx, Px, AJx, FC
The space charge capacitances are determined from CV measurements between Base-
Emitter, Base-Collector and Base-Substrate. Since we will use the G-P description, the
parameters AJE, AJC and AJS have to be set to -0.5 (default!).
Note
it has been observed that the trace of C(v) above Pj does not influence the S-parameter fitting, because
the diffusion capacitance (parameters TF, ITF, XTF and VTF) usually overlay the space charge capacitance
at this bias condition.
We will refer to the modeling of the BE capacitance. The BC and CS capacitors are
modeled correspondingly.
For the measurement of the Base-Emitter capacitance, the Collector and Substrate are left
open.
Note
On the DC bias for the CV measurements:
A reasonable rule of thumb is to include biases up to where the capacitance is 2 to 3 times the zero bias
capacitance. You want to be less than FC*VJ.
The default FC for VBIC is 0.9, rather than the 0.5 of Spice Gummel-Poon.
Note
On the influence of the remaining capacitances of the open pin:
as one of the transistor pins is left open, the measurements of the actual C(v) trace is overlaid by the
other capacitances CSBi ( i = E,C )
The total measured capacitance is therefore CSBi in parallel with the parasitic ones. This
means that the measurement results are always a bit too big.
When using a capacitance meter like the HP4284, that eliminates by its measurement
principle parasitic capacitances to ground, this effect can be avoided by applying an AC
short to the open transistor pin versus ground (big capacitor, 1nF ).
The behavior of the space charge capacitor is given by equations (VBIC-8) and (VBIC-9):
For vBE < FC*PE, we have:
649
(CV-1)
and else
(CV-2)
with
CJE : space charge capacitance at vBE = 0V
PE : built-in potential or pole voltage (typ. 0,7V)
ME : junction exponential factor, determines the slope of the cv plot

(abrupt pn junction (<0,5um): ME = 1/2)
(linear pn junction (> 5um): ME = 1/3)
FC : forward capacitance switching coefficient, default 0,9
Determination of the CV parameters
For simplicity, we only use the measurement data from the negative bias, i.e. we begin
with the equation (CV-1):
A logarithmic conversion yields:
ln(CSBE) = ln(CJE) - ME ln(1 - vBE / PE ) (CV-3)
This equation can be interpreted as a linear function according to the ideas of linear
regression analysis:
y=b+mx
with
y = ln(CSBE) (CV-4)
b = ln(CJE) (CV-5)
and
m = - ME (CV-6)
x = ln(1 - vBE / PE ) (CV-7)
Linear regression means to fit a line to given measurement points. Therfore, the three
main equations of a linear regression are b=f(xi,yi) and m=f(xi,yi), together with a fitting
quality factor r=f(xi,yi,m,b). For a good fit, r~0.9...0.9999. See also the appendix on
linear regression.
How to proceed
The measured values of CSBC are logarithmically converted according to (CV-4). Following
(CV-7), the stimuli data of the forcing voltage vBE are nonlinearily converted too.
This is done using a starting value for the unknown parameter PE (e.g. 0,2V). These two
arrays are now introduced into the regression equations as corresponding yi- and xi-
values. A linear curve is fitted to this transformed 'cloud' of stimulating and measured
data. Thus we get the y-intersect b(PE) and the slope m(PE) for the actual value of PE. In
the next step, this procedure is repeated with an incremented PE, and we get another pair
of m(PE) and b(PE). But now the regression coefficient r 2 will be different from the earlier
one. I.e. depending on the actual value of PE, the regression line fits better or worse the
transformed data 'cloud'. Once the best regression coefficient is found, the iteration loop is
exited and we finally get PE_opt as well as the corresponding b(PE_opt) and m(PE_opt).
Thus we get from (CV-6):

ME = - m(PE_opt)
and from (CV-5):

CJE = exp [ b(PE_opt) ]
After that, we apply the same methodology to the other two CV curves.
Note
The Base-Collector capacitance is distributed between intrinsic and extrinsic transistor with the program
'partition_CBC'. The main part of this capacitance is usually associated with the parasitic transistor
(default settings CJC=0.05*CBC_total and CJEP=0.95*CBC_total).
From the C(v) measurements, however, the partitioning cannot be fine tuned. This is done with the S-
parameters. Bsically, the partitioning affects the knee in S22, but also the magnitude of S21 for higher
frequencies. See file VBIC_TUTOR.mdl.
Parasitic Resistors From DC Measurements

RE
650
The ohmic emitter resistor is physically located between the internal Emitter E' and the
external Emitter pin E. When we apply a Base current and have the Emitter pin grounded,
we get a voltage at the open Collector that is proportional to the Base current through this
Emitter resistor. For this mesurement, we leave the substrate contact open. We then
derivate vCE with respect to iB, we get the equivalent RE for each operating point. The
result is displayed in a separate plot. The value of RE is then the mean value of the flat
range in this plot.
RBX
An interesting method to determine RBX is to use the RE-flyback method, with additionally
measuring vBE /T.Zimmer/. This method is applied now.
The theoretical values of the measured voltages are:

vCE = VT * ln(1/AI) + iB * RE
with AI: reverse current amplification in common Base
and vBE = iB * RE + iB * RBX + vB'E'

Subtracting these equations and dividing by iB yields:
i.e. after plotting the measured data accordingly, we get RBX as the y-intersect.
In a parameter visualization step, we apply a loop to these data, in which a line is fitted to
two adjacent points, and the local y-intersect is calculated. The incremental y-intersects
are then displayed against the stimulus iB, and represent RBX vs. iB. RBX is extracted
651
from the most constant range in this plot.
Note
RBX may also be obtained form a flyback measurement on the parasitic PNP with iE_MAIN = 0, i.e Emitter
pin of the MAIN transistor left open.
RBI
Note
RBI is modeled using S-parameters. See further down.
RCX
RCX is extracted from the slope of 'iS' of the Gummel-Poon plot.
The method is after the Isub-method in:
J.Berkner, A Survey of DC-Methods for Determining the Series Resistances of Bipolar
Transistors including the New delta-Isub Method, Proceedings of the European IC-CAP
user meeting 1994, Colmar, France
Referring to two alues of iS and iC at two voltages vBE for which iS rises linearily on a
log(iS)_vs_vBE plot, it is
As an example, here a Gummel-Poon plot including iS:
As explained, RCX can be extracted from the slope of iS, see next plot:
RCI
A first-guess RCI is extracted from the slope of the quasi-saturated output characteristics
at highest iB
Note
RCI is typically bigger than RCX (5-10 fold), and its value is typically not representing a physical ohmic
resistance value.
RS
From the reverse Gummel plot of the parasitic transistor, the substrate resistor Rs can be
extracted. It is visible as a decline of the slope for high currents.
RS is tuned-in and optimized in the parasitic reverse Gummel setup.
Early Voltages
VEF, VER
When extracting the VBIC Early voltage parameters, their interaction with the space
charge capacitances must be considered. Referring to publication
C.C.McAndrew, L.W.Nagel, "Early Effect Modeling in SPICE",
IEEE Journal of Solid-State Circuits, vol. 31, Nr. 1, Januar 1996, equ.9
we start with
652
with
gof, gor output conductance in forward and reverse mode
qje, qjc normalized charges
cje, cjc normalized capacitances
VEF, VER Early voltages
These two equations can be rearranged into:
They can be solved for the Early voltages and we get:
The operating point has to be selected so that no quasi-saturation or high-current effects

(avalanche or thermal) disturb the measured data.
See the DUT 'VEF_VER' in file VBIC_EXTR.mdl for details.
Note
VER is often quite low. It is approximately VER ~ VEF * CJC / CJE and CJC is a lot less than CJE in
"normal" BJTs because the B-C doping is a lot less than the B-E doping.
Diode Parameters
The DC parameters of the Base and Collector current of the main NPN transistor are
determined from Gummel-Poon measurements. Referring to the figure below, the
parameters are extracted from their dominant bias sweep ranges using regression
techniques or visual extraction techniques. The following graphic depicts the basic
strategy:
The extraction of the IS, NF, IBEx and NEx parameters of the VBIC model is probably the
most tedious task in the whole modeling process. Because this 'diode fitting' has to be
applied to both, the iC and iB in both, forward and reverse opoeration, and additionally
also once again to the parasitic PNP transistor, it can become a bit confusing. Therefore, it
is most important to always remember the VBIC nomenclature:
I ideal part of the Base current

N non-ideal part of the Base current
P parasitic PNP
With this in mind, the modeling of the different diode currents becomes much more
transparent.
Because of this repetitive task, we refer here only to the iC(vBE) modeling of the main
NPN transistor. And this methology can then be applied to all the other parameters like:
NPN forward: IS, NF, IBEI, NEI, IBEN, NEN, IKF

NPN reverse: NR, IBCI, NCI, IBCN, NCN, IKR
653
PNP forward: ISP, NFP, IBEIP, IBENP, IKP
PNP reverse: IBCIP, NCIP, IBNCNP, NCNP
IS, NF
IS transport saturation current
NF forward current emission coefficient
NF determines the slope and IS the y-intersect of the half-logarithmically plotted iC(vBE).
Provided that vB'E'=vBE and vB'C'=vBC , we start with the ice formula (VBIC-11):
We simplify this equation by setting the normalized Base charge qb=1. In other words, we
neglect the Early effect for this extraction. This assumption, which is in most cases no big
simplification, will be corrected later by applying a quick optimization to the extracted
parameters.
Also, we select a low vCE for the IS and NF, IBEx and NEx parameter extractions.
Otherwise, it might hgappen that due to the reverse voltage vCB, a leakage current from
the Collector into the Base overlays the positive Base current and, therefore, does not
allow to accurately extract the parameters IBEN and NEN.

We begin with
For v be > 0.2, a very typical condition to obtain noise-free measurement data, this
simplifies further to
We first apply a non-linear transformation to this equation, i.e. the measured data, in
order to obtain a linear context between the measured values of iC and the stimulating
values of vBE:
A log10 conversion gives:
or
This can be considered as a linear form:
with substituting:
654
How to proceed
We select a sub-range of the measured data, where the half-logarithmicly plotted data
represent a straight line. Then the logarithmically converted icei of this sub-range are
interpreted as y- and the linear vBEi values as x-data for the regression formula. Applying
these formulas, (see the appendix), we obtain y-intersect 'b' and the slope 'm' of the
straight fitted line.
A final re-substitution gives the parameters I S and N F out of 'b' and 'm':
IS = 10 b
and
Validity of the extraction

vBE between 0,2V (no noise) and 0,7V (no high current effects), low vCE (to keep the
Early effect low).
Note
many modeling engineers do not extract NF, but keep it rather NF=1. The reason is that for NF NR, the
power balance of the transistor is violated (it generates power instead of behaving like a controlled
resistor). They instead extract TNOM from the fitting of the slope.
In this case, the above formula for NF changes to
A hint on visualized parameter extractions:
Transforming the measured data such that the model parameter can be displayed directly
against the stimulating voltage or current is another smart way to determine model
parameters. In the case of NF this would mean to start with
This is the mathematical representation of the half-logarithmic Gummel plot for i C.

The parameter NF is proportional to the slope and we have therefore to differentiate ln(i C)
with respect to v BE and obtain:
Therefore, if we display the calculated NF (what is the 'effective NF' for every measured
data point) versus v BE, we get
655
Note
Applied to modeling iB(vBE), this method allows to determine the exact sub-range of data from which to
extract NEN and NEI as the mean value of that flat range. The same principle cal also be applied to extract
IBEN and IBEI.
IKF
Referring to the log(iCE)_vs_vBE curve from the top of this chapter, for which we had set
vBC ~ 0, we extract IKF from vBC > 0, typically a vBC which corresponds to vCE = vCE
max of the foutput setup. This is because for vBC ~ 0, quasi-saturation and Webster effect
(IKF) are overlaying each other.
Therefore, IKF is extracted from a separate Gummel-Poon measurement, with a

sufficiently high vCE. Before we do so, we first perform a quick fine-optimization of the so
far determined Gummel-Poon parameters IS, NF and IBEx, NEx for the actual DC bias
conditions.
Then, for the determination of the parameter value, we follow the simple idea that IKF
represents that iC value where beta=ic/ib has fallen to half its maximum value. Although,
there is always an overlay of IKF and RE, this method gives usually quite reliable results.
Again, like with the IS, NF, IBEx and NEx parameters, a quick optimizer finetuning is
applied to match both, the RE and the IKF value above the knee in the log(iC)_vs_vBE
plot.
Applying the same methodology, we extract the main transistor reverse parameters as
well as the forward and reverse Gummel parameters of the parasitic transistor. See the
setups in DUT 'dc_gummel' for details.
Output Characteristics
As a quick remark on the measurement setup, the question is to either stimulate the
output characteristic with a Base current or a Base voltage. It was found that the quasi-
static behavior of the output characteristics shows up much better when forcing a Base
voltage than a current.
However, with a forcing vB, thermal effects show up much more. While this is a problem
for Gummel-Poon, the VBIC model features the parameter RTH.
Note
Typical thermal resistance on wafer is ~200'K/W
VEF, VER
These parameters were already extracted at the beginning, and therefore, need now only
be fine-optimized. With all the Gummel-parameters (IS, NF, IBEx, NEx) extracted, the
output characteristics should fit now well for medium vCE and no avalanche or thermal
runaway
RCI, V0 and GAMM (the Quasi Saturation Parameters)

In setup foutput, the quasi-saturation parameters RCI, V0 and GAMM are determined.
The figure above visualizes, how the parameter RCI affects the output characterization
fitting. It basically determines the slope of the saturated range. Therefore, its value can be
determined from the transition to quasi-saturation. It should be mentioned that RCI does
not necessarily relate to a physical value.
With the parameter GAMM, the effect of quasi-saturation is deferred to higher currents,
see the figure below. However note that GAMM also affects through the capacitors Cbcq
and Cbcx the dynamic behavior of the transistor.
656
VO determines the begin of velocity saturation. This means a smoothing at the high-end
of the quasi-saturation. For big values of VCO, its influence on the curve vanishes.
Finally, HRCF was added empirically to the quasi-saturation model in VBIC, to reflect an
increase of iC with higher vCE. The influence of HRCF is increasing with smaller parameter
values. Over a wide range, therefore, HRCF conflicts somehow with VO. We therefore set
it to its default value of 1000.
The high-frequency quasisaturation parameter
QC0
not affecting the DC performance, will be determined later from S-parameter
measurements.
WSP (DC Current Distribution Between Main NPN and

Parasitic NPN)
Parameter WSP, which partitions the control of Itfp between intrinsic main and parasitic
transistor, is used to model the 2nd knee in plots log(ib)_vs_vce and log(is)_vs_vce. It
can be used for fine-tuning. Its default value '1' represents no such 2nd knee, while the
other limit '0' includes it. See setup 'foutput_aval_therm_is_ib' for details.
657
AVC1 and AVC2 (Avalanche Effect Modeling)

The weak avalanche current is modeled as a special current from Collector to the Base
following:
Here, AVC1 and AVC2 are the avalanche model parameters, PC and MC are the Base-
Collector space-charge capacitor model parameters.
See the transforms in setup 'foutput_aval_therm_is_ib' for details.
RTH (Output Characteristics Thermal Effect

Modeling)
See the transforms in setup 'foutput_aval_therm_is_ib' for details.
658
Note
In Spectre, RTH is only active if parameter selft=1
S-Parameter Modeling
The transit time T FF is calculated from transit (cutoff) frequency measurements following
the known formula
fT is the cutoff frequency where H21 = 1 (0dB). This is usually extrapolated from H21
measurements by fitting a -20dB/decade slope, see below.
This means, we get
or solved for the parameter of interest:
where fT1-pole is a function of the bias current iC and the bias voltage vCE .
The modeling equation for TFF (VBIC-34) is
and
This means that, except for QTF, we can apply the known Gummel-Poon extraction
methods also to the VBIC model.
For small values of the forward transport current IF, the above equation simplifies to
This would allow to model TF and QTF. However, the effects are difficult to separate.
Therefore, we start with QTF=0. If required, we obtain its final value from S-parameter
fine-tuning.
659
Since the VBIC model includes the quasi-saturation effect, we need to watch out for the
right DC bias conditions to extract the HF parameters of TFF. Therefore, we first extract
TF, XTF, ITF and VTF from values of vCE where there is no quasi-saturation and no
avalanche or thermal effect. Then, we select an S-parameter measurement in DC
quasisaturation, and model QCO by optimization.
The graphic below gives detailed information about the best parameter extraction bias
ranges.
DC bias condition for extraction of the transit time parameters
Parameter extraction in detail

We proceed like with the Gummel-Poon model, and consider first the obtained trace of TFF
for as low as possible vCE, but not in quasi-saturation !
The above figure shows the theoretical curve in addition to a typically measured one. For
low frequencies, the real measured curve is overlaid by the space charge capacitor effects
for low collector currents.
On the other hand, the DC bias conditions for which the TFF parameter ITF and XTF show
up, cannot be measured without self heating effects, because the required bias current iC
is usually well above ~50mA. However, because the VBIC includes RTH, and if this RTH
modeling was performed carefully in the foutput setup, this should not influence the TF,
ITF, XTF extraction.
Due to these overlay and measurement problems, it had been found that a pretty simple
and straight-forward extraction technique can be applied that gives nevertheless quite
reasonable results. This method is explained below. There exist some more complex
strategies, but the extraction results may be not much better.
How to proceed
TF
is extracted as the minimum value of TFF.
XTF
The behavior of TFF was given in the above figure. In many cases, measurement data for
a higher Collector current are not available due to compliance. So XTF is estimated from
the trace of TF at max. available Collector bias current under the assumption that it would
be TFF at infinite current:
660
This usually gives a pretty good first-order estimation. Due to the Collector current
limitations, an estimation correction like XTF = 5..10 * XTF_extracted can improve the
starting conditions for the optimizer.
ITF
Referring to the same measurement restrictions as above, a good first-order estimation of

ITF is to use the max. Collector current measured:
ITF = MAX(iC_meas)
Again, since the end of the TFF trace is often not measurable, correct this estimation by
ITF = ~5*ITF_extracted.
Note
This XTF and ITF extraction method follows the idea of the Gummel-Poon Base resistor parameter
extraction!
Note
In the TFF equation VBIC-34, when TFF = TF ? (1 _ XTF / 2), i.e. TFF is in the middle between its
minimum and maximum value, the corresponding IF bias current is iC_meas = 2,41 ? ITF.
VTF
Finally, we consider also the vCE sweep, but, again, not in quasi-saturation:
VTF can be obtained from (VBIC-34) for a fixed value of iC:
This gives:
661
QCO
this parameter is very dominant in the gm plot, obtained from S-to-Y parameter
conversions. We select a bias condition in quasi-saturation and apply an optimization.
TD
The delay time parameter affects mostly the ttrace of S21 in the 1st quadrant. It is well
visible in the phase of gm. Its value is optimized for S21 in setup 'biased_Spar'.
RBI
In VBIC, the Base resistance is modeled by RBB'=RBX + RBI/qb. Since qb represents a

quite complex formula (see equations (VBIC-1 ... VBIC-3), the 'input-impedance-circle
method' from the Gummel-Poon model cannot applied easily to get a starting value for the
inner Base resistance RBI. Therefore, RBI is obtained by optimizing the S11 plot of setup
'biased_Spar'.
Final Parameter Tuning
Since the high-frequency model parameters can affect back the DC model results, all
modeling setups should be re-simulated and fine-tuned by optimization. Using the
measurements of setups verify_DC and verify_S, the error between measured and
simulated curves can be analyzed easily for all relevant bias points.
Parametrized Modeling vs. Transistor Geometry

The following info is a copy of the VBIC committee extraction suggestions:
CV
From devices of different base-emitter area to perimeter ratios CJE can be modeled as
function of emitter-base area and perimeter, by extracting area, perimeter, and if
necessary constant (corner) components from the different area/perimeter structures. CJE
is then calculated as the sum of area, perimeter, and constant components, based on
specific device geometry.
If no layout information is known, the CJC/CJEP splitting should be done so that ft is

modeled well (note that typically most of the capacitance should be in CJEP).
If the relative areas of the b-c junction under the emitter and not under the emitter (e.g.
intrinsic and extrinsic b-c junction areas) are known, partition the extracted CJC between
CJC and CJEP accordingly.
Preferred approach
If devices of different base-collector intrinsic/extrinsic areas and perimeters are available,
area and perimeter components of CJC+CJEP can be easily determined. CJC is then
calculated from the area of the base-emitter, and CJEP from the base-collector perimeter
and the excess of the base-collector area over the base-emitter area, plus a constant
(corner) component if so modeled.
DC
Ibc components can be split into intrinsic (IBCI/IBCN) and extrinsic (IBEIP/IBENP) in
manner analogous to the split for CJC/CJEP. IS, IBEI, IBEN, etc. can all be related to
geometry (area and perimeter) by determining them for two or more area/perimeter
ratios and then calculating the area and perimeter components.
Thermal Modeling
The VBIC model includes many parameters which allow the modeling of the transistor
behavior at different operating temperatures.
Here some hints on this kind of modeling from the VBIC 1.1.4 extraction
recommendations of the VBIC committee:
The temperature dependence of the junction built-in potentials and capacitances is

determined by the activation energies, which are determined from the temperature
dependence of saturation currents.
From low-bias FG Ic data over temperature, determine EA by optimization. XIS can also
be included in the optimization, but from my experience the optimization is relatively
insensitive to XIS, and EA is by far the major controlling factor. Therefore, XIS should be
left at default (XIS=3).
From low-bias FG Ib data over temperature, determine EAIE and EANE by optimization.
Again, XII and XIN do not affect this optimization much, and so should be set to 3 rather
than being included in the optimization. Different temperature dependences for IS, IBEI
and IBEN are necessary to model the variation of beta with temperature properly, the
beta roll-off at low Vbe, caused by the non-ideal component of Ibe, has a different
temperature variation than the variation of the peak/flat beta with temperature.
662
There are no "free" parameters for the temperature variation of Ie in reverse mode
operation, they are fixed by EA and XIS, determined from the temperature variation of Ic
in forward mode operation. EAIC and EANC are determined by optimizing the fit to low
bias RG Ib data. Ibc has separate temperature parameters from Ibe, as it has a slightly
different variation with temperature.
From low-bias measurements of the parasitic b-c current (over temperature) determine
IBCIP/NCIP/IBCNP/NCNP (EAIS/EANS) as done for intrinsic device Ibe/Ibc parameters.
These parameters are really only used to flag improper biasing of a device, so reasonable
estimated could be used instead of making measurements and extracting parameter
values.
APPENDIX
Plots of transistor characteristics like MAG and FMAX can be used to check the final S-
parameter fitting. Since they include all the S-parameter, i.e. S11 ... S22, they can be
used to show the general S-parameter fitting at a glance.
MAG
The maximum amplification gain is obtained when the transistor or amplifier is inserted
between matching load/source resistance. I.e. for conjugate complex impedance matching
at both sides of the transistor, the amplifier gain is different, usually higher than for a Z0
= 50 environment..
The maximum available gain is
MAG is only valid if K>1
MAG calculates the gain of the entire network (DUT plus matching networks).
Note
it is a typical characteristic of RF transistors that they are instable for lower frequencies (K<1). Calculating
MAG for those frequencies makes therefore no sense. In other words, only for frequencies with K>1, an
impedance matching for max. gain MAG can be designed.
It is further helpful to draw |S21| 2 and MAG together in the same plot versus log(freq).
Usually, MAG is bigger than |S21| 2, what means that the power amplification can be
improved with impedance matching, compared to the conventional modeling
measurement condition Z0=50. See the following figure. Note that for a wide frequency
range, the transistor is not unconditionally stable (K<1), and therefore MAG is not
defined.
Comparing power amplification |S21|2 to the maximum available gain
FMAX
The transit frequency ft is defined as the frequency at which |h21| hits '1'. This frequency
ft determines the max. switching frequency and is therefore most interesting for digital
applications (output AC-wise shorted).
Yet, this is not the max. frequency for the transistor in analog operation and without the
assumption of an AC-wise short at the output of the twoport.
This max. frequency is defined by FMAX , which refers to the maximum available gain
MAG. As long as MAG> 1, the transistor can still generate oscillations. The max. possible
oscillation frequency is therefore
663
Regression Analysis
Regression Analysis is a technique which provides a best curve fit for given data sets.
Mathematical Basics of Curve Fitting

Let's assume we made 'N' measurements yi at the stimulating points xi.
I.e. we obtained the array [ xi, yi ], which is then shown in a plot.
A curve Y( x ) shall be fitted to this array of measured data points using least square
curve fitting technique.
Referring to an individual measurement point, the fitting error is:
and for all data points:
The fitting will be done by varying the coefficients of the fitting curve of equation (2). The
minimum of the total error E depends on the values of these coefficients. This means, we
have to differentiate E partially versus the curve coefficients and to set the results to zero.
We obtain a system of equations, solve it, and get the values of the coefficients for a best
curve fit. This is known as regression analysis.
Note
Linear Curve Fitting

IC-CAP File:
demo_features/5_PEL_PROGRAMMING/3_PARAM_EXTRACTIONS_CURVE_FITTING/1_basic_PEL_extractions/1fit_line.mdl
Provided we have got an array of N measured data points of the form [ xi,yi ].
A linear curve with the equation
Linear regression applied to measurement points
Using the least means square method following equ.(2) yields:
664
We obtain from (5) after a re-arrangement:
and from (6):
Multiplying (7) by -N and (8) by xi and adding these two equations allows the elimination
of the coefficient 'a', and we can separate the slope 'b':
or:
and from (8) for the y-intersect 'a':
with 'b' according to (10).
With equations (10) and (11), we determine the values of the two coefficients of the line
which fits best into the 'cloud' of measured data.
Finally, a curve fitting quality factor r2 [ regression coefficient ] is defined.

Its value ranges from [ 0 < r2 < 1 ].
The closer it is to 1, the better is the fit of the linear curve.
with 'b' from (10)
General Formula for Regression Analysis

Linear Regression:
Quadratic Regression:
Cubic Regression:
continue with the evolution of matrices like above.
665
Application: Calculating the quadratic polynomial
parameters from measurement data
IC-CAP File:
demo_features/5_PEL_PROGRAMMING/3_PARAM_EXTRACTIONS_CURVE_FITTING/1_basic_PEL_extractions/1fit_quadratic.mdl
Starting with equ.(14) from above
we get the following system of 3 equations:
Equations (15a) .. (15c) are of the form
for which we have already a solution for a, b and c in chapter 6.Data_Interpolation.
Therefore, with the following abbreviations:
we can re-use the quadratic data interpolation formulae of the chapter Data Interpolation:
and
and are ready to get the quadratic fitting function y( x )
In the plot below, the red curve (dummy_data) has been fitted by the magenta curve
(QUADFIT), inside the marked box.
666
667
Acknowledgements, Web info, and
Publications
Acknowledgements
The author would like to thank especially F. X. Sinnesbichler of the Technical University
Munich, Germany, Institute of High Frequency Techniques. His IC-CAP workshop about
VBIC in Munich, Jan.1998, in which he emphasized on introducing the VBIC model with its
similarities to the Gummel-Poon toolkit, inspired much the beginning of this VBIC toolkit.
Very special acknowledgments also to J.Berkner of Infineon Technologies, Munich,

Germany and E.Gerhardt of Alcatel SEL, Stuttgart, Germany, for many helpful discussions.
Also, many thanks to Fujiang Lin, formerly working at the Agilent Singapore Modeling
Center, for informations about the new version 1.2 model and special parameter
extractions methods.
Web Info
Info on the web since July, 2000:http://www.fht-
esslingen.de/institute/iafgp/neu/VBIC/index.html
Publications
J.J. Ebers, J.L. Moll, "Large-Signal Behaviour of Junction Transistors", Proceedings of
the IRE, pp. 1761-1772, Dezember 1994.
I. E. Getreu, "Modelling the Bipolar Transistor", Elsevier, Amsterdam, 1978, also
available as Tektronix book, Tektronix part no. 062-2841-00.
B. K. Gummel, H. C. Poon, "An Integral Charge Control Model of Bipolar Transistors",
Bell Systems Technical Journal, vol. 49, pp827-853, 1970.
H. C. de Graaff, F.M. Klaassen, "Compact Transistor Modelling for Circuit Design",
Springer Verlag, Wien, 1990.
C. McAndrew et al, "VBIC95: An Improved Vertical, IC Bipolar Transistor Model",
Proceedings of the 1995 BiCMOS Circuits and Technology Meeting, pp.170-177,
1995, Minneapolis.
C. McAndrew et al, "VBIC95, The Verticl Bipolar Inter-Company Model", IEEE Journal
of Solid-State Circuits, vol. 31, Nr. 10, Oktober 1996.
C.C.McAndrew, L.W.Nagel, "Early Effect Modeling in SPICE", IEEE Journal of Solid-
State Circuits, vol. 31, Nr. 1, Januar 1996.
F. Najim (ed.), "VBIC95: An Improved Bipolar Transistor Model", IEEE Circuits and
Devices Magazine, vol. 12, pp. 11-15, March 1996.
J. Parker, M. Dunn, "VBIC95 Bipolar Transistor Model andAssociated Parameter
Extraction", Proceedings of the 1995 HP EEsof US ICCAP Users' Meeting, Washington,
Dezember 1995.
P. Antognetti, G. Massobrio, "Semiconductor Device Modeling with SPICE", McGraw-
Hill, New York, 1988.
G.M. Kull et al, "A Unified Circuit Model for Bipolar Transistors Including Qasi-
Saturation Effects", IEEE Transactions on Electron Devices, vol. 32, Nr. 6, pp. 1103-
1113, Juni 1985.
P.B. Weil, L.P. McNamee, "Simulation of Excess Phase in Bipolar Transistors", IEEE
Transactions on Circuits and Systems, vol. 25, Nr. 2, pp. 114-116, Februar 1978.
F.van der Wiele, W.L.Engl, P.G.Jespers, "Process and Device Modeling for Integrated
Circuit Design", NATO Advanced Study Institute, Louvain-la-Neuve, Belgium, 1977
J. Berkner, "Parameter Extraktion for BJT Quasisaturation Models", Proceedings of
the 4th European IC-CAP User Meeting, Berlin, Oktober 1997.
F. X. Sinnesbichler, G. R. Olbrich, "Early Effect Modelling of Bipolar Junction
Transistors with Low Early Voltages", Proceedings of the 4th European IC-CAP User
Meeting, Berlin, Oktober 1997.
G. Neaves et al, "Characterization of Low VAR Bipolar Transistors Using a Revised
SPICE Simulator", IEEE Proceedings of the 1992 Bipolar BICMOS Circuits and
Technology Meeting, pp 229-232, 1992, Minneapolis.
T.Zimmer: Contribution la modlisation des transistors haute frquence, Thse
l'Universit de Bordeaux I, 17.7.1992.
Zimmer, Meresse, Cazenave, Dom, 'Simple Determination of BJT Extrinsic Base
Resistance', Electron.Letters, 10.10.91, vol.27, no.21, p.1895
F. Sischka, "Gummel-Poon Bipolar Model Parameter Extraction Toolkit", Agilent
Technologies GmbH, Mnchen, Germany
668
RF Passive Linear Components
Contents
Test Fixture
Time Domain Modeling (iccapmhb)
Frequency Domain Modeling (iccapmhb)
RLC
RLC Modeling (iccapmhb)
Resistor (iccapmhb)
Capacitor (iccapmhb)
Inductor (iccapmhb)
Strip Lines (iccapmhb)
Modeling Vias (iccapmhb)
Dummy Chip (iccapmhb)
Packages
Package Modeling (iccapmhb)
SOT23 Package (iccapmhb)
TSOP44 Package (iccapmhb)
Modeling Connector (iccapmhb)
Frequency Dependent Losses (iccapmhb)
Test for Passive Circuits (iccapmhb)
669
Modeling of a Chip Capacitor up to 26
GHz
First, the NWA is calibrated with the method of /Gronau 1992/ , see chapter on network
analyzer measurements and calibration. This method allows a shift of the calibration plane
up to the pins of the capacitor chip. The following figure shows the measurement setup,
the theoretical curve for an ideal capacitor from DC to infinite frequency as well as the s21
measurement curve of the SMT capacitor up to 26GHz. The chip capacitor was pressed to
two 50 Ohm strip lines.
Some pre-considerations:
For the frequency '0', the capacitor blocks the power flow. In the above figure, the
theoretical curve starts at (0 + j*0). At infinite frequency, all power is passing the
capacitor, and that's why Sxy should end at (1 + j*0). Between these points, a semi-circle
should be expected. Yet, the real measurement gives a different picture: the lower
measurement frequency of 45MHz causes a certain phase shift of the first measurement
point. At higher frequencies, parasitic inductors show up, since the curve exceeds the
point (1 + j*0).
Therefore, the modeling can be performed in the following sequence:

The equivalent schematic starts simply with a capacitor (CMAIN). Its value is the specified
manufacturer value (see the following figure). A simulation should show a good
agreement with the measured curve for low frequencies. Next, we are concentrating on
the fit at higher frequencies. We determine the value of RMAIN (since the S21 curve does
not hit exactly (1 + j0), then the value of LMAIN (for the raw trace beyond (1 + j*0). We
also evaluate the parasitic connectivity parameters L1,C1 and L2,C2.
The fitting quality of the model is depicted in the following figure for the magnitude and
phase of S21.
670
Modeling of a Dummy Chip on a Wafer
Using on-wafer calibration techniques allows to correct for all measurement errors down to
the chip pads. Nevertheless is it important for frequencies above some 100MHz to
accurately model the parasitic effects like bond pad capacitances on the chip itself,
because they affect and distort the correct measurement of the 'inner device' on the chip.
Once they are known, they can be eliminated from the measurement of the whole chip
using de-embedding techniques, see chapter on de-embedding in section Vector Network
Analyzers.
The following shows the layout of a so-called 'open dummy chip'. To the left are the 3
pads for the ground-signal-ground connection of the network analyzer's port 1, and to the
right those of port 2. The transistor itself is left 'empty', so that we measure nothing but
the parasitics of these pads.
The following figure shows layout of an 'open dummy chip' device on a wafer (the active
part of the bipolar transistor is 'left empty').
Following figures show the S-parameter measurement result of a structure like in above
figure. The frequency range is from 50 MHz to 20 GHz. The curves can be considered as
typical, and as will be seen shortly can be interpreted quite conveniently.
To get a first idea about the equivalent schematic, the S-parameters are first transformed
to Z-parameters. Fig.4 gives the magnitude of Z11 and Z22. There is no increase in value
visible for high frequencies. This leads to the first modeling result: we will probably be
able to neglect inductive elements for the equivalent schematic to be developed.
Next, we will test the applicability of a PI structure. So the S-parameters are converted to
Y-parameters.
Assuming the PI equivalent schematic (fig.5) and expecting capacitors and dielectric
losses as its elements (C in parallel to G), we plot
REAL(A) = REAL(Y11+Y12),
IMAG(A) = IMAG(Y11+Y12) / (2*PI*freq)
REAL(B) = REAL(Y22+Y12),
IMAG(B) = IMAG(Y22+Y12) / (2*PI*freq)
REAL(C) = REAL(-Y12)
IMAG(C) = IMAG(-Y12) / (2*PI*freq)
versus frequency. The results are given in following figure.
671
Unfortunately we cannot get the expected curves like REAL(A) and REAL(B) to be a
frequency-independent resistor and IMAG(A) and IMAG(B) to be a capacitor with a
constant value over the frequency. Therefore we have to reject the PI equivalent
structure.
We will now test the applicability of a T structure, consisting of capacitors in series with
resistors. Such a capacitor is typical for lossy silicon capacitor parasitics. The simplest
representation of such a structure is a Z matrix, see figure below.
So we convert the measured S-parameters to Z, and plot
REAL(A) = REAL(Z11-Z12)
REAL(B) = REAL(Z22-Z12).
This corresponds to the ohmic resistors RA resp. RB of components A resp. B in fig.10. As

given in fig.11, there is no frequency dependency. Therefore, the assumption of a TEE
structure for these resistors is valid and we get RA = 110 Ohm as a mean value for
frequencies above 3GHz. Using the same strategy, RB is calculated out of the plot
REAL(Z22-Z12) as 80 Ohm.
Then, we plot the inverse of the imaginary parts divided by 2*PI*freq as depicted in above
figure. Also this type of transformed data looks quite constant, what can be interpreted
with respect to fig.10 as capacitors CA and CB in series with the already identified resistor
RA and RB. Their values are CA = 140fF and CBC = 160fF from above figure.
Finally in above figure, we see the real part of Z12 and Z21. Also here, we have an ohmic
672
behavior for frequencies above 3 GHz, and the mean value is RC = 30 Ohm. On the other
hand, the imaginary part of Zxy gives no clear picture for our purpose and is therefore
neglected.
We are now finished with the development of the equivalent schematic. The resulting
lumped circuit is depicted in following figure:
Finally, we define a netlist using the appropriate simulator syntax like UCB-Spice, MDS,
Eldo, Saber, Spectre, Hspice, or Pspice. The simulation results are then plotted in
following figures together with the measurement results. Using the optimizer will give us
the fine-tuned parameter values.
The final rms fitting error using the proposed equivalent schematic is about 15%. With
respect to our goal of simply subtracting the Y matrix of our 'open dummy structure' for
de-embedding the DUT on the wafer, we can state that the developed equivalent
schematic is totally in parallel with the 'inner device'. Therefore a simple subtraction of the
S-to-Y transformed 'dummy device' data from the also S-to-Y transformed 'total device'
data will be sufficient for the de-embedding of the 'inner device'.
Physical Interpretation of the Equivalent Schematic:

The resistors RA and RB can be interpreted as the loss in the oxide capacitor, while RC is
basically the resistance in the bulk of the substrate. The figure below shows the physical
interpretation of the OPEN dummy schematic. Shown are the test pads, the probes and
the silicon substrate.
Modeling the Second Order Effects of an Open

Dummy Structure
Like already a bit visible in above figures, the measured traces may deviate from the half-
circle for high frequencies. This indicates that the underlying schematic is too simple.
Fig.18 depicts such a situation for a 50GHz open dummy measurement.
673
This means, that for frequencies above 7.5GHz, the Sii plots tend to another half circle,
which will hit the X axis for infinite frequencies at a much lower ohmic value. This effect,
also explainable with HF power finding a lower-resistance path to ground for higher
frequencies than with the simple equivalent schematic, is now modeled with the
components C10, R10 resp. C20, R20. As stated, their capacitance values are much
smaller than those of C1 and C2. (See also chapter 3.3.4., section 'About the
interpretation of S-parameters of an asymmetrical TEE structure).
Physically, this new HF power path is cross-coupling between the pads of the dummy
structure.
For the Sij plot in above figure, the trace showing up at frequencies above 7.5GHz, means
more cross-over power from port1 to port2 and vice versa than for frequencies below
7.5GHz. Since the so far added C10, R10 do definitively not lead to more HF power flowing
between the ports, this effect can only be described with an additional cross-over
capacitance C12. The Figure below explains the improved schematic:
The parameter extraction is done as follows:

Firstly, the model parameters of R1, R2, R3 and C1, C2 are extracted as mean values
from the transformed S-parameter curves, for a frequency range up to ~10..15GHz and
following the procedures mentioned in the section above.
Note
Since the underlying simple TEE structure is not valid for the data, the traces will not be flat, but
frequency dependent!
Then, the additional components C10, R10, C20, and R20 are added to the circuit and
tuned with the tuning feature in IC-CAP, mainly fitting the Sxx traces. Then, C12 is added
too and fine-tuned basically to fit the Sxy traces, and not distorting the Sxx plots.
Finally, an optimizer run will find the best parameter values.
PUBLICATIONS:
an interesting publication on OPEN dummies on different silicon substrates is:

L.P.Chen et.al., Extraction of Parasitic Parameters of Dummy Devices on Different Silicon
Substrates, 1998 IEEE Radio Frequency Integrated Circuits Symposium, June 7-9, 1998,
Baltimore, pp. 321-324, ISSN 1097-2633
674
Modeling Frequency Dependent Losses
IC_CAP demo files: skin_tdr.mdl, skin_spar.mdl
Time Domain:
Frequency dependent losses of lines show up with TDR measurements as a decline of the
slope of the TDR step function.
Frequency Domain:
Regarding S-parameter measurements, the effect is represented by an S-shaped decrease
in the MAG(Sxy).
Modeling
Conventional SPICE models, however, cannot cover this effect. They only allow the
modeling of ideal delay lines without a decrease in magnitude. Modern high-frequency
simulators like ADS feature several models to cover different non-ideal effects of lines.
/Katzier/ proposes a special SPICE subcircuit for modeling this effect by using
conventional SPICE components.
Lacross and Racross are modeling the damping of |S21|. Here, Lacross is the most
important fitting parameter and Racrossis used for fine-tuning. Note that the main effect
is the variation in the magnitude of Sxy, while the phase of Sxy is nearly not affected.
There are no analytic methods to extract the values of Lacross and Racross and therefore,
their values are either tuned or optimized.
On the other hand, the proposed schematic can easily be cascaded, due to its nice feature
of exhibiting Z0 at its input and output independently of any parameter settings.
The following figure depicts such a scenario:

the modeling of a 25cm long line using a cascade of 4 schematics. the model of above
figure can easily be cascaded for an improved fit of geometrically long devices, e.g. strip
lines etc. (plot by courtesy of H.Katzier)
.SUBCKT mdlfreqlosses 1=P1 20=P2

Rs1 1 3 T1.Z0
Rs2 3 2 T1.Z0
Lacross 1 2 1n
Racross 1 2 10
Rp 3 33 T1.Z0^2//Racross
Cp 33 0 Lacross//T1.Z0^2
T1 2 0 20 0 Z0=50 TD=200p
.ENDS
The above table shows the SPICE input deck for the circuit of the figure shown above.
The performance of this model including the ideal delay line is given with the next 3 plots.
Note that the model affects mainly the decrease in magnitude of S21.
675
Note
Sxx is not affected by the schematic of the first figure showing SPICE model.
The above figure shows Frequency-dependent losses with S-parameters: the s-shaped
decrease of MAG(Sxy) is modeled by the proposed schematic.
Time Domain
In the time domain, the frequency-dependent losses effect is visible as a quasi-'lowpass'
slope of the back-reflected step response. The above figure shows a modeling of a chain
of delay lines with and without this effect. Again, LACROSS and RACROSS were used to
tune the slope.
The above figure shows measuring and modeling the end of an open delay line:
frequency-dependent losses effect with TDR measurements.
Applied to a Gaussian impulse, the effect becomes visible with a unsymmetrical impulse
response with an undershoot. This proofs nicely the plot of fig.4, since a step function is
the integral of a pulse.
676
Note
It should be noted that the proposed schematic features an excellent fitting capability in both, TDR and S-
parameter domain. However, when keeping the length parameter of the preceding strip line unaffected, it
requires to modify the strip line's delay time, or when applying a more physical strip line model like ADS's
MLIN2, to modify the ?r of the substrate and the epsr_losstangent. In other words, the combination of
delay line plus the proposed TEE schematic should be as a blackbox which then in total fits the delay line.
The figure below shows the effect in the time-domain and the frequency-domain.
Publications:
H.Katzier: Analyse und Modellierung der Gesamtdaempfungsverluste von
verlustbehafteten Streifenleitungen, 4.GMM/ITG Diskussionssitzung, October 1996, Berlin
677
Test Fixture Modeling in Frequency
Domain
This chapter gives an example on the modeling of a test fixture for packaged transistors.
In a first step, the HP8753D NWA has been calibrated up to the ends of its 3.5mm
connectors. Then the test fixture has been connected and measured. Figure 1 shows the
result.
As the next step, we will check the applicability of delay lines to the equivalent schematic.
Therefore, FFT transform is applied to get an idea on the time delays of the microstrip
lines on the fixture. From fig.2 we get an estimation for the delay lines of our equivalent
schematic. We assume TD=250ps. An optimizer run is used to overcome the uncertainties
with this method that stem from the limited bandwidth of 3GHz of the HP8753D, and we
get a delay of ~290ps. The phase fitting after this step is depicted in fig.3 below:
Next, the schematic for the simulation is developed further. From Sxy, the cross-coupling
can be modeled with a capacitor C and a resistor R.
678
Then, back to Sxx and the estimation of 28fF for the non-idealities of the open strip lines
(which will be contacted by the SOT23 package later), Cx and Rx have been added to help
in the modeling of the decrease in MAG_Sxx. During this procedure, the delay times of the
strip lines at port1 and port2 were re-optimized manually.
From the phase plots in ph_Sxx, it can be found that a split of each delay line into two
parts with different Z0 improves the fit. This tiny improvement affects very much the de-
embedding quality later.
.SUBCKT MEAS 2=P2 1=P1 3=E

T1 1 0 11 0 Z0=50 TD=76.6p
T11 11 0 111 0 Z0=47 TD=76.6p
C1 111 1111 85f
R1 1111 0 1k
Copen1 111 0 28f
T2 2 0 22 0 Z0=50 TD=69p
T22 22 0 222 0 Z0=48 TD=69p
C2 222 2222 85f
R2 2222 0 1k
Copen2 222 0 28f
C 111 12 2.6f
R 12 222 6.9k
.ENDS
The final modeling results are given below.
679
Modeling of Spiral Inductors
On-wafer spiral inductors are a commonly used component in integrated high frequency
amplifiers. Their performance contributes a lot to the overall performance. Additionally,
they occupy a big space on the wafer, what makes them ext4remely expensive
components. Therefore, a thorough modeling, usually for both, the S-parameter and the Q
factor, has to be applied.
De-embedding the measured data before modeling

It must be emphasized that one of the most important issues for accurate spiral inductor
modeling is a preceding, very thorough de-embedding. Since basically all passive
components exhibit a pretty large geometry compared to the transistor area, the
connection lines from the test pads to the component itself cannot be neglected.
Therefore, a dummy THRU has to be measured besides the conventional OPEN dummy.
Such a THRU dummy can be considered as an equivalent to the SHORT dummy for
transistors.
Altogether, the de-embedding procedure is like this:
1. de-embed the total measurement from the OPEN dummy first,

2. then, the THRU dummy is de-embedded from the OPEN dummy and its phase shift
and if required, the loss is modeled with respect to the length of the THRU strip line.
After that, we can calculate the equivalent ABC matrices of stub at port 1 and of port
2.
3. finally, de-embed the result of step (1) from the de-embedded THRU of step (2).
While the OPEN de-embedding is performed applied to Y matrices, the THRU is best
represented by either an S matrix or an ABC matrix. If the THRU can be modeled by a
lossless delay line, we calculate the de-embedded performance of the spiral inductor
directly out of the total S-parameters by
with TD1= delay time of strip line at port1

TD2= delay time of strip line at port2
If the THRU is represented by a lossy delay line, we apply ABC matrix de-embedding. This
is done by firstly modeling the lossy delay line with an ABC matrix each like
where
L = length of the strip line
= damping factor
= phase shift factor
and then, secondly, applying the de-embedding for both strip lines at both connections of
the spiral inductor as
See the chapter 3.3.3.3 on De-embedding for more details.
1-port modeling of a spiral inductor

This chapter presents a tutorial about how to use IC-Cap's PEL language for the flexible
and accurate modeling of spiral inductors, measured as a 1-port component (See fig.1). In
this example, PEL is used for extractions and also for the implementation of equations into
the SPICE netlist (e.g. resonance formula).
Fig.1 shows the test setup and the measurement result of the spiral inductor. The
measurement was taken from 300kHZ to 3GHz using an HP8753 network analyzer under
IC-CAP.
680
The modeling was performed in the following order:

Looking at the lowest frequencies in fig.1, we assume a resistive behavior at 300kHz. The
real part of S11 is
REAL(S11) = -0,387
or
This is our starting point for the equivalent schematic of fig.5 below.
Then, in order to get the value of the designed inductor, we convert the measured S11
parameters to the complex resistance plane following
The result is given in fig.2.
For low frequencies, LP is without any doubt affecting the curve the most. Therefore, we
divide the imaginary part of the complex resistance of equation (2) by (2*PI*freq). The
result is shown in fig.3. From the lowest frequencies, we obtain Lp=120nH.
It can be seen that our DUT exhibits a resonance at high and another one at even higher
frequencies. Next, fig.4 shows the real part of the complex resistance (fig.2) versus
frequency. From this plot, both frequencies can be determined easily.
Considering first the resonance at 990MHz, we assume a capacitor Cp. in parallel to LP.
Furthermore, we neglect the second resonance for the moment. This allows to apply the
resonance frequency formula equ.3:
681
Solved for Cp., and introducing Res_freq=990MHz , we obtain for Cp.
with LP=120nH from fig.3.
Note
The formulas can be entered directly into IC-Cap's parameter list, allowing optimization with nested
parameters.
Since the S11 curve of fig.1 does not go exactly through S11=1 at the first resonance, we
assume a parallel resistor to LP and Cp. Its value can be obtained using equation (1) at
the resonance frequency of 990MHz. We obtain:
Rp=10 Ohm.
Now, we are ready for the modeling of the second resonance.
It has to be a circuit, which behaves neutral for low and high frequencies. This means a
series resonance circuit, in parallel to the so far developed schematic. Therefore, we add
the three more components RS, Cs and Ls to the schematic in fig.5. They represent the
resonance at 1.7GHz, see fig.4. From fig.1, we further assume a bandwidth of 150MHz.
Referring to table 1, the parameter values for Cs and RS can be calculated dependent on a
given Ls as follows (f0=1.7GHz, BW=150MHz):
Again, these formulas can be entered as 'values' into IC-CAP's parameter list. Therefore,
by tuning Ls or using the optimizer, we can improve the total fitting for fig.1 at the second
resonance (set XMIN and XMAX in the optimizer).
Of course, both resonance circuits overlay each other, and fine-tuning the second
resonance will distort the fitting of the first one and vice versa. Therefore, we introduce a
tuning parameter DELTA for Cp. as shown in equation (7) for a final optimization step.
This DELTA parameter is optimized with respect to the phase of S11.
Figure 5 shows the final result: the SPICE circuit deck and the schematic of our modeling
approach. Fig.6 depicts the IC-CAP parameter table with the used formulas and also the
final parameter values.
The curve fitting is given in fig.7. As can be seen, there is a very good agreement between
the measured and simulated curve.
IC-CAP parameter table including the formulas and the final parameter values is:
Model fit for the 1-port spiral inductor is:
682
2-port modeling of a spiral inductor

Note
IC_CAP file: 2SPIRAL_EXTRACT_MASTER.mdl
In this example, the spiral inductor has been measured between the ports of the network
analyzer. In a first approach, we will try to find the parameters of a typical, given circuit
for such type of inductors. In most cases, however, this will not lead to the desired
frequency independent model parameters. Therefore, in a second approach, we will
develop a suitable subcircuit out of the measured S-parameter data and calculate the
model parameter values for the lumped components. This method will follow two
approaches: a first approach which assumes a inner PI structure, possibly overlaid by
outer parasitics, and a second one which assumes a PI for then inner part of the
schematic.
Let's first go for the standard method, a given, predefined circuit.

Referring to the mentioned IC-CAP file, we consider
DUT spiral_ind_quicky
This is the underlying schematic for the standard interpretation of the Spiral S-
parameters:
The parameter values can be obtained as follows:

Interprete the Y-pars after a simple PI schematic
2. calculate Yxy = (Y12 + Y21) / 2

and then A = Y11 + Yxy
B = Y22 + Yxy
C = -Yxy
Z10 = 1/A
R10 = REAL(Z10)
C10 = (-IMAG(Z10)*2*PI*freq)^-1
Z20 = 1/B
R20 = REAL(Z20)
C20 = (-IMAG(Z20)*2*PI*freq)^-1
Z12 = 1/C
RMAIN = REAL(Z12) !RMAIN: typ. before resonance, at low frequ.
LMAIN = IMAG(Z12)//(2*PI*freq) !LMAIN: typ. before resonance, at low frequ.
C12 = -1/(2*PI*freq*IMAG(Z12)) !C12: typ. after resonance, at high frequ.
683
The underlying schematic is too simple for today's spirals up to 26GHz.
This is indicated by the frequency dependent component values. An other indication is that
the schematic isn't able to produce an inductor Sxy simulation trace around the zero. It is
rather tending towards the zero.. Such a trace is a hint for another overlying, more
complex schematic.
Therefore, the next DUTs in the mentioned IC-CAP file 2SPIRAL_EXTRACT_MASTER.mdl

refer to publications of more complex schematics.
Finally, the IC-CAP file features a more general passive components modeling tool. this is
described below:
DUT check_4_PI
The setups of this chapter develop an equivalent schematic based on a PI structure. Both,
the details of the schematic, and the parameter values are investigated. It is checked, if
there is an overlying TEE structure, and if yes, the model parameters of this overlying
structure are extracted as well.
A little bit of theory:

This setup is used to develop a PI based schematic out of the deembedded data. Usually,
the data are simply converted to Y parameters (see spiral_ind_quicky). In many cases,
however, this is too simple, and an overlying more complex circuit must be assumed.
Therefore, the input, the transfer and the output resistance of a PI structure (Y matrix)
are calculated and displayed, under the assumption of overlay series parasitics (TEE
structure, Z matrix), and a cross-over capacitor (another Y matrix).
See the header of transform S_2_Y_2_Z_2_Y.
Purpose of transform S_2_Y_2_Z_2_Y:

The de-embedded S-parameters are converted first to Y, and a possibly overlying cross-
capacitor C12_ is subtracted. Next, the resulting Y-parameters are converted to Z-
parameters, and possibly overlying series components (L1_, R1_, L2_, R2_, R3_) are
subtracted. After that, the remaining data are converted to Y parameter and finally
checked for the inner PI based schematic.
However, the parameter values of these overlying parasitics are not yet known. Transform
check_4_TEE_overlay includes TUNER function which calls transform S_2_Y_2_Z_2_Y.
Starting with some reasonable default values for the assumed overlay parasitics, (middle
position of the tuner), and including also parameter values which effectively switch-off the
effect of the corresponding parasitic (left position of the tuner), the tuner bars are
adjusted top-down in order to get as physical as possible remaining Y-parameters. In
order to check the physical meaning of these Y-parameters, they are converted to Rin,
Rtrans and Rout, and these complex resistances are plotted on a real/imaginary scale.
Therefore, the transforms Zin_from_Ypar, Ztrans_from_Ypar, and Zout_from_Ypar
calculate complex resistanc from the resulting S_2_Y_2_Z_2_Y Y-parameters. This is
displayed in the plots ZIN_Y, ZTRANS_Y, and ZOUT_Y.
How to proceed:
As a first step, transform check_4_TEE_overlay includes a tuner, which allows the tuning
of the assumed overlaying TEE parasitics. The goal is to tune these parasitics until we
obtain physical plots ZIN_Y, ZTRANS_Y, and ZOUT_Y. This is true if all curves turn clock-
wise with frequency, and the left half-plane of the complex resistance plane is not hit.
Note
If a resonance occurs (circle) it has to be symmetrical to the x axis.
After we obtain as physical as possible results for the plots ZIN_Y, ZTRANS_Y, and
ZOUT_Y, we have automatically found the parameter values of the overlying parasitics.
Therefore, we now copy the adjusted parameter values, stored temporarily in the setup
variable table, into the parameter list of the DUT test circuit.
Having obtained the input, the transfer and the output resistance of the inner, PI based
schematic, we have to interprete these plots in the next step. Transform
calc_L_or_C_utility supports this task.
This means, we middle-click a data point in one of these plots (beginning at the lowest
frequency), and can calculate the corresponding inductor. Then, these values are entered
into the parameter list. If required, the schematic is also enhanced for more components.
In order to further evaluate the schematic behind ZIN_Y, ZTRANS_Y, and ZOUT_Y, the
setups Zin_strip_RC, Ztrans_strip_RL and Zout_strip_RC are used to strip off step-by-step
the so far determined main components of the partial schematic and to determine the
second order components. However, this method is very sensitive, and it might be faster
to develop the schematic of the inner resistances by applying the calc_L_or_C_ulility to
the plots ZIN_Y, ZTRANS_Y, and ZOUT_Y.
After the ZIN, ZTRANS and ZOUT plots are modeled satisfyingly, we go back to fit the S-
parameter plots of this setup.
In the next step, all the parameters are finetuned, and optimized.
Note: in the tuner transforms, the tune bars are programmed so that their middle value is
684
the starting value. This allows you to find your way back if you are lost!
Finally, Program fit_qual calculates the vector error between the measured de-embedded
data and the simulation result
Program S11_1port calculates the 1-port S parameter out of the de-embedded data, and
program Q calculates out of it the quality factor Q.
Contents of Transform S_2_Y_2_Z_2_Y

This transform performs a de-embedding of the following parasitic structure:
s=S_deemb.M
y = TwoPort(s,"S","Y")
!de-embed possibly overlaying C12 cross-couple capacitor
i=0
WHILE i < SIZE(y)
y.M.11[i]=y.M.11[i] -j*2*PI*freq[i]*C12_
y.M.12[i]=y.M.12[i] +j*2*PI*freq[i]*C12_
y.M.21[i]=y.M.21[i] +j*2*PI*freq[i]*C12_
y.M.22[i]=y.M.22[i] -j*2*PI*freq[i]*C12_
i = i + 1
END WHILE
z = TwoPort(y,"Y","Z")
!de-embed possibly overlaying TEE R_series_L-structure
i=0
WHILE i < SIZE(z)
z.M.11[i]=z.M.11[i] -R1_-j*2*PI*freq[i]*L1_ -R3_
z.M.12[i]=z.M.12[i] -R3_
z.M.21[i]=z.M.21[i] -R3_
z.M.22[i]=z.M.22[i] -R2_-j*2*PI*freq[i]*L2_ -R3_
i = i + 1
END WHILE
RETURN TwoPort(z,"Z","Y")
DUT check_4_TEE
same idea as above, but now investigating in a TEE schematic, possibly overlaid by a PI
structure.
Contents of Transform S_2_Z_2_Y_2_Z

PRINT "running transform 'S_2_Z_2_Y_2_Z'..."
This transform performs a de-embedding of the following parasitic structure:
s=S_deemb.M
685
z = TwoPort(s,"S","Z")
de-embed possibly overlaying common ground resistor
i=0
WHILE i < SIZE (z)
z.M.11[i]=z.M.11[i] -R3_
z.M.12[i]=z.M.12[i] -R3_
z.M.21[i]=z.M.21[i] -R3_
z.M.22[i]=z.M.22[i] -R3_
i = i + 1
END WHILE
y = TwoPort(z,"Z","Y")
de-embed possibly overlaying PI structure (with R_series_C branches)
i=0
WHILE i < SIZE(y)
y.M.11[i]=y.M.11[i] -(R13_+(j*2*PI*freq[i]*C13_)^-1)^-1 -(R12_+(j*2*PI*freq[i]*C12_)^-1)^-1
y.M.12[i]=y.M.12[i] +(R12_+(j*2*PI*freq[i]*C12_)^-1)^-1
y.M.21[i]=y.M.21[i] +(R12_+(j*2*PI*freq[i]*C12_)^-1)^-1
y.M.22[i]=y.M.22[i] -(R23_+(j*2*PI*freq[i]*C23_)^-1)^-1 -(R12_+(j*2*PI*freq[i]*C12_)^-1)^-1
i = i + 1
END WHILE
RETURN TwoPort(y,"Y","Z")
Publications:
Y.Sun, J.L.Tauritz, R.G.F.Baets, Micromachined RF Passive Components ant Their

Applications in MMICs, Int'l Journal of RF and Microwave Computer-Aided Engineering,
vol.9, no.4, July 1999, Wiley&Sons, ISSN 1096-4290
Ali.M.Niknejad, Analysis, Design and Optimization of Spiral Inductors and Transformers for
Si RF ICs, master thesis at the University of Berkeley, California, is on the web at
http://kabuki.eecs.berkeley.edu/~niknejad/
Calculating the Q-factor of spiral inductors

The quality factor of a spiral inductor is represented by:
It cannot be calculated correctly out of the 2-Port S-Parameters because of the Z0=50 ? of
Port 2.
In this case, we convert the 2-port S-parameters to 1-port S-parameters and obtain:
From that, we apply the basic S11 <> R conversion (Smith chart),
and obtain for the input impedance at Port 1
This finally gives the requested 1-Port Q factor of the spiral inductor
Note
The same result is obtained when converting the 2-port S-parameters to
2-port Y-parameters, and then calculating
686
Modeling of a 3.5mm Connector
The connector is soldered to a board, made from FR4 material (up to 3GHz) or from
Rogers (for higher frequencies).
The TDR cable is connected with the connector-board combination. Since a TDR allows an
easy modeling 'from the left to the right', we can focus on the modeling of the connector.
On the TDR trace in the above figure, we see first an inductor effect, followed by a
capacitor effect. This gives the first netlist approach for the connector model. The final
parameter values of the connector capacitance cp1, the connector inductance ls1 as well
as the connector delay with the delay line tconn are fine-tuned either manually (TUNER)
or by an optimizer.
.subckt mdlconnector 101=tinp 105=toutp

v0 1 0 PWL(289p 1.375m 0.40404n 0)
...
* TDR step function into 50Ohm,
* plus delay line for the path TDR->Connector...
v99 100 99 PULSE(0 -0.625m 0.79596n 4.04p 4.04p 0)
ri 100 101 51.1
t1 101 0 102 0 TD=150p Z0=50
* now comes the connector
tconn 102 0 103 0 TD=38p Z0=50
ls1 103 104 460p
cp1 104 0 140f
* finally the line on the board (not to be modeled)
tmic1 104 0 105 0 TD=500p Z0=50
rend 105 0 1G
.ends
The above figure also gives the simulation result: the inductor models the overshoot, and
the capacitor the undershoot.
Delay time tconn.TD reflects the delay of the connector (33ps). Delay line tmic1.TD (for
the modeling of the delay time of the strip line on the board) is set to 500ps and is not
modeled here in this example.
687
Modeling Vias
This chapter is about the modeling of vias on ceramic substrates. The test structure is a
50 Ohm strip line between the NWA ports, with different vias to ground (backside of the
ceramic substrate) in the middle. This means, we will expect more Sxy for higher
frequencies, since the via will behave worse and worse.
The following examples are cited from literature /Bhm/.
For a single via:
For a serial twin-via:
688
Note
The higher phase shift stems from the broader reference planes!
And for a parallel twin-via:
689
Note
When vias are used to connect more than 2 layers, a special antenna effect may occur. For example, if a
via in a 12-layer board only interconnects layer1 with layer2, the remaining via down to the last 12. layer
may act as a resonance. This is depicted below.
Publications:
M.Bhm, Diploma Thesis: Metechnische und rechnerische Untersuchungen von
Verbindungsstrukturen fr elektronische Schaltungen in einem Frequenzbereich von
45MHz bis 25GHz, TU Ilmenau/Thringen, Germany, Oktober 1992
690
Package Modeling Workshop
Introduction
With today's clock rates,where chips become faster than the package performance, the
package cannot be neglected any more and needs to be modeled very accurately.
Limitations with Standard Models

Some problems with manufacturer's models such as:
Single pin model
Cross-coupling model
Step1. The Reference Plane of a Package Model has to be defined
Note
It is very important to define the beginning and the end of the package model.
Two Types of Characterization Measurements

Required
When compared to a conventional transistor modeling, packages represent a big
geometry. This implies a much more complex equivalent schematic. It contains Ls, Rs, Cs
and Delay Lines which are directly related to the geometries of the package.
Time Domain Reflectometer (TDR) measurements allow the identification of

schematic components related to geometries, but are poor in crosstalk resolution
between the package pins.
691
Vector Network Analyzer (VNA) measurements have excellent resolution related to
crosstalk, but S-parameter measurements are difficult to interpret for big delay
times.
Time Domain Reflectometer (TDR) basics

For Identifying the Lumped Components of a Single Line, Time Domain Reflectometer
(TDR) is commonly used. A TDR stimulates a very fast step voltage and measures its back
reflections. The measurement trace allows then to identify capacitors, inductors and delay
lines with respect to their physical location.
Interpreting TDR Traces
Time Domain Reflectometry Basics
TDR Applied to Package Modeling
TDR Calibration
692
Vector Network Analyzer (VNA) basics

VNA Usage for Package Modeling for accuracy and cross-talk
Vector Network Analyzer Calibration
After both, the TDR and VNA have been calibrated, we now need to model the test fixture.
Test Fixture for Package Modeling

Test fixture for TSOP44 for DRAMs
693
Before modeling the package, the test fixture has to be modeled standalone. Then, after
inserting the package, the additional model components belong to the package.
Test Fixture Time Domain Modeling Result
Network Analyzer test fixture model fine-tuning
Package Modeling
After the instruments are calibrated, and the fixture is modeled, Package modeling is
done.
Package Modeling Involves Two Steps:
The Basic Equivalent Schematic
694
TDR Modeling Result (incl.fixture)- Step1
TDR Modeling Result (incl.fixture)- Step2
The Total Model

After both, the test fixture and the package have been carefully modeled step-by-step, the
package model can be clearly identified out of the total schematic.
695
The Final Package S-Parameters
Package Modeling Conclusions

A Method to Develop an Equivalent SPICE Schematic for packages has been proposed. The
adaptive modeling procedure assures accurate and reliable models, and thus, successful
system designs.
696
Modeling of a SMT Resistor
In this case, the NWA has to be calibrated with a suitable method in order to shift the
calibration plane right up to the pins of the DUT. (see /Gronau 1992/ and also the chapter
'network analyzer measurements' for details)
Fig.2 shows the measurement result of a 47 Ohm resistor. The Smith chart shows a real
part of normalized 1.95*50 Ohm = 97,5 Ohm for low frequencies. I.e. the DUT plus the 50
Ohms of port 2 of the NWA are in series with each other. For higher frequencies, the
resistor shows a capacitive behavior.
The polar diagram of S21 shows power transmission from port1 to port2 with capacitive
effects at higher frequencies. This leads to the equivalent schematic below.
PUBLICATIONS about the NWA calibration:
G.Gronau, Scattering Parameter Measurement of Microstrip Devices, Microwave

Journal, Nov.1992.
G.Gronau, I.Wolff,: A Simple Broad-Band Device De-embedding Method Using an

Automatic Network Analyzer with Time-Domain Option, IEEE Trans.on Microwave
Theory and Tech., Vol. 37, No.3, March 1989.
M.Bhm, Diploma Thesis Technical University Ilmenau/Germany: Metechnische

und rechnerische Untersuchungen von Verbindungsstrukturen fr elektronische
Schaltungen in einem Frequenzbereich von 45MHZ bis 25GHz, 30.4.1993.
Rudolf Stassen, 'Einsatz eines Mikrowellen-Spitzenmeplatzes zur Charakterisierung von

GmbH,
available from Technische Informationsbibliothek, Hannover.
697
Developing Models
This topic explains developing models for Passive On-Wafer Components & Small Signal
Behavior of Transistors.
Introduction
The RF modeling of passive components is commonly based on a given equivalent
schematic. However, this may lead to frequency-dependent lumped components. The
following procedure proposes a new method which allows to develop both, the schematic
and the corresponding model parameters.
Step1. First of all, we need special dummy structures for accurate de-embedding,
especially for spiral inductors
It is recommended to have OPEN and THRU dummies available,and to model them first.
With the modeling, we check if there
are really no series components hidden in the OPEN dummy, no parallel components in
the THRU, and if we can de-embed
the DUT from the THRU by ABC matrix de-embedding.
Step2. After we have verified the dummy device performance,the device is carefully de-
embedded:
Note
it cannot be overemphasized enough, that an accurate, verified de-embedding is the prerequisite for
accurate passive device models!
Developing the Equivalent Schematic

Important schematics for free RF modeling
Measurement -> De-embedding -> Z-Matrix
698
Measurement -> De-embedding -> Y-Matrix
To develop the right subcircuit behind the matrix components

From the de-embedded S-parameters, we calculate the complex impedances of the
branches of the Y or Z matrix, and analyze
their locus curve on the complex impedance plane.
Locus curves of typical impedances
699
Selecting a PI or a TEE schematic

The de-embedded S-parameters are first converted to e.g. Y-parameters, and the locus of
the individual complex impedance of each PI structure branch is roughly analyzed and
checked if the locus curve has a physical meaning.
Also, a rough schematic of that branch is developed. Then, the same is done for the
opposite matrix, e.g. Z-matrix.
Spiral Inductor
Schematic developed from the locus curves based on a PI structure
Spiral Inductor
Checked against the locus curve of a TEE structure
The TEE structure seems to represent the actual device more physically than the PI
structure.
In the next step, we do the fine-work and develop accurately the branch schematic and
fine-tune the model parameters.
TEE structure schematic selected
700
If a TEE or a PI alone does not make it then the locus curve inspection method can also be
applied to overlying PI and TEE structures. For example, In the case of an assumed PI,
overlaid by a TEE, we set the TEE inductors and resistors first to 0, and then 'tune their
values in', until the locus curves of the inner PI components correspond to physical locus
curves. Then, the so found TEE components are stripped-off, and the remaining inner PI is
again modeled with the method given before.
Application of this method to other RF modeling problems

Besides spiral inductors, it can be applied to:
MIM capacitors,
varactor diodes
resistors
transistors small signal model, quasistatic and non-quasistatic
Small Signal Modeling Result
Modeling result for a SOI transistor up to 50GHz
701
Summary
A method has been proposed:
To adaptively develop an equivalent schematic for passive components

To precisely extract their parameter values
This method:
This method has been implemented into IC-CAP and shows a very good match
between measured and simulated data.
It can be combined with data management for parameter scaling
702
Modeling of a SOT-23 Package
IC-CAP file: demo_features/.../.../sot23.mdl
The SOT23 package

For the modeling of packages, two types of measurements are required: a SHORTed
package and an OPEN package. In most cases, an OPEN or 'dummy' package is available.
It will be used for the evaluation of the package capacitances. For the inductors and
resistors, the SHORTed package will be used. If it is not available, a 'cold' measurement of
the transistor may help. In this case, a strong current is applied to port1, while half of that
current is sunk out of port2. With such a bias, the transistor behaves like a good SHORT,
and we are able to measure the parasitic inductances and resistors. The following figures
show the results of these two measurements. The frequency range was from 30kHZ to
3GHz (HP8753D).
Since we assume an equivalent schematic with LC structures, we start with the modeling
of the SHORT measurements (inductors and resistors in series with the ports). As a first
attempt, the figure below sketches the starting point for our schematic.
Note
The resistors Rcoldxy represent the behavior of the shorted transistor in the 'cold' measurement mode.
Beginning the modeling with the lowest frequencies, the resistors R1 and R2 model the
ohmic losses of the connection between the test fixture and the package and also in the
package itself. They represent the 'non-zero Ohm' starting points of the Sxx curves. On
the other hand, Sxy is not '0' for low frequencies. Therefore, we have to model some
power crosstalk from Port1 to Port2, what is reflected by resistor R3. In order to evaluate
the correct values of these three resistors, the S-parameters are converted to Z, and the
real part of it is displayed against the frequency. The PEL program that performs that
transformation (and also for the imaginary parts for the modeling of the inductors later) is
printed out.
!PEL program in IC-CAP to calculate the resistors and inductors out of the S-parameters of the
'cold' measurement.
!PRINT "running transform 'Z_to_RLx' ..
tmpx=freq !import stimulus data
tmpy=TwoPort(S,"S","Z") !import measured/simulated data
i=0
WHILE i < SIZE(tmpy)
tmpxyM=(tmpy.M.12[i]+tmpy.M.21[i])//2 ! calculate the
tmpxyS=(tmpy.S.12[i]+tmpy.S.21[i])//2 ! values of the
tmpy.M.11[i]=(tmpy.M.11[i]-tmpxyM) ! underlying
tmpy.S.11[i]=(tmpy.S.11[i]-tmpxyS) ! TEE structure devices
tmpy.M.22[i]=(tmpy.M.22[i]-tmpxyM) ! Rx and j*2*PI*Lx
703
tmpy.S.22[i]=(tmpy.S.22[i]-tmpxyS)
!make the imag.parts contain the inductors only
tmpy.M.11[i]=REAL(tmpy.M.11[i])+j*IMAG(tmpy.M.11[i])*(2PI*tmpx[i])^-1
tmpy.S.11[i]=REAL(tmpy.S.11[i])+j*IMAG(tmpy.S.11[i])*(2PI*tmpx[i])^-1
i = i + 1
END WHILE
RETURN tmpy
Ideally, the resistors should show-up with frequency-independent values. In reality, as

can be seen above, this is not the case due to the fact that the supposed schematic (a
simple TEE structure) is too simple. Therefore, it would fit the S-parameter only if its
resistors would have exactly this frequency dependence!
Nevertheless, we want to stay with 'ordinary' frequency independent resistors. So we

extract their values from a frequency range where we are well above the lower frequency
limitations of the network analyzer and well below 2nd order effects at higher frequencies.
In our example, we use a maximum flat region of the curves and therefore the frequency
range between 500MHz and 700MHz. The simulation result is given in figure below. We
see a nice fitting for low frequencies, as expected. In the next step, we will discuss the
fitting for higher frequencies and the determination of the inductor values.
A second result from the little program above are the values of the inductors. They are
shown in the following figure. Unfortunately, these plots give no clear picture about where
to extract the values of Lx (there are no flat regions visible). Therefore, the extraction is
done manually, with respect to the S-parameters, but using the values of the plots in
figure below (for low frequencies) as a first guess.
After all, an optimizer run fine-tunes the values of the inductors Lx with respect to the S-
parameters. The result is given in the figure below:
The next step is to determine the values of the package capacitors. They show up with the
'open' (dummy package) measurement. Yet, that 'open' measurement is overlaid by the
already known series inductors and resistors. Therefore, the measured S-parameters have
704
to be de-embedded from these parasitic components first. This is done with a subtraction
of the corresponding Z-matrices.
We further assume a PI structure for the parasitic capacitors, and therefore, we have to
convert the resulting Z-parameters to Y-parameters. See the little program below which
contains printout of the de-embedding program that strips-off the dummy device data
from the already known parasitic inductors and resistors.
PRINT "running xform Ydeem ..."

! Z_open - Z_short
tmp = TwoPort(S_open,"S","Z") - TwoPort(S_short, "S","Z")
! transform to Y parameters
tmp = TwoPort(tmp,"Z","Y")
RETURN tmp
As already mentioned, the resulting Y matrix is then interpreted with respect to a PI

structure. This is done in transform 'Y_to_GCx', see the following code sample which
contains printout of the transformation of the Y-matrix elements into the capacitors and
parallel conductors:
tmpx=freq !import stimulus data

tmpy=Ydeem !import measured/simulated data
i=0
WHILE i < SIZE(tmpy)
tmpxyM=(tmpy.M.12[i]+tmpy.M.21[i])//2 ! calculate the
tmpxyS=(tmpy.S.12[i]+tmpy.S.21[i])//2 ! values of the
tmpy.M.11[i]=(tmpy.M.11[i]+tmpxyM) ! underlying
tmpy.S.11[i]=(tmpy.S.11[i]+tmpxyS) ! TEE structure devices
tmpy.M.22[i]=(tmpy.M.22[i]+tmpxyM) ! Rx and j*2*PI*Lx
tmpy.S.22[i]=(tmpy.S.22[i]+tmpxyS)
!change sign of Yxy
tmpy.M.12[i]=-tmpy.M.12[i]
tmpy.S.12[i]=-tmpy.S.12[i]
tmpy.M.21[i]=-tmpy.M.21[i]
tmpy.S.21[i]=-tmpy.S.21[i]
!make the imag.parts contain the inductors only
tmpy.M.11[i]=REAL(tmpy.M.11[i])+j*IMAG(tmpy.M.11[i])//(2PI*tmpx[i])
tmpy.S.11[i]=REAL(tmpy.S.11[i])+j*IMAG(tmpy.S.11[i])//(2PI*tmpx[i])
i = i + 1
END WHILE
RETURN tmpy
The REAL parts of the returned array of the program shown above represent the
conductors that are in parallel with the capacitors of the PI structure. This is depicted in
following figure. Also plotted are the extracted conductor values. Again, they represent
the mean value of a maximum flat region of these curves.
On the other hand, the IMAG parts of the program in above figure represent the
capacitors. This is displayed in the figure below. As we end-up with rather flat plots, it
seems that our assumption of a PI structure is quite correct. Therefore, the extraction of
the capacitor values is pretty straight-forward, and simply the mean value of a maximum
flat range of that plots. The resulting values are shown in the plots also.
This leads finally to the total equivalent schematic, as shown in the figure below:
705
The fitting of the final equivalent schematic of above figure to the SOT-23 OPEN
measurement is depicted in the figure below:
Another model for the SOT-23 package is given in the figure below. We want to check its
performance against the model from above, which was valid up to 3GHz.
This model represents more the actual physical facts in the package: The transistor is
fixed with its Collector to the package, while Emitter and Base are contacted by bond
wires. Therefore, the nodes 11 and 31 are now 'inside' the inductors L1 and L3:
The performance of the model of the above figure was simulated up to 26GHz.
The goal is to compare it with the model from above for the frequency range up to 3GHz,
see the markers in the figures below:
Result: The developed (simpler) schematic is very much applicable for frequencies up to
3GHz. The additional parasitic components are required for the performance above 5GHz.
706
707
S-Parameters of Strip Lines
Strip Lines Tutorial
A strip line exhibits a TEM field, i.e. a transverse electromagnetic mode field, where the
dielectric field is perpendicular to the magnetic field, and both are perpendicular to the
direction of wave propagation.
From this sketch, we can conclude the following basic properties of a microstrip line,
depending on the geometry factor W/L:
The characteristic impedance, Z0, ranges from about 20 to about 100. The limit of
100 exists for a very simple reason: the width is much less than the height, and such a
structure cannot be manufactured (under-etching etc).
This sketch allows to make some fundamental considerations:

As a matter of fact, a small microstrip line exhibits less capacitance than a wide one.
Inspecting the plot, this concludes that a lower capacitance in a microstrip line comes
along with a lower impedance Z0.
This is obvious, when recalling that
Referring to crosstalk between lines, we can learn from the sketch above that a low
impedance microstrip line is capacitive. I.e. the energy is rather between the metal
conductor and the ground. I.e. two low impedance striplines side-by-side, will exhibit less
cross-talk than two high impedance striplines. By the way, this is a key design rule for
packages and connectors.
As another important outcome, a shielding across a microstrip line will lead to the fact
that the impedance of the resulting strip line will be lower, because more of the electro-
magnetic field will now be present in the enlarged electric field consisting of the previous
field in the dielectric layer plus the additional space between the active metal layer and
the top cover. Therefore, a cover across a microstrip line reduces the resulting impedance,
and, thus, reduces cross-talk between adjacent striplines.
To further reduce cross-talk of adjacent lines, i.e. to reduce the impedance of each line
(increase the electric field, i.e. make the lines more capacitive), reduce the height of the
dielectric material.
However, on the other hand, a cover 'kills' the performance of filters designed from strip
lines based on electric field coupling!
Modeling a lossy strip line from S-parameters

Tutorial
The S-parameters of a strip line of length L and a characteristic impedance Z0, measured
with a network analyzer, can be considered like a transition of a 50 system into a Z0
system with a long delay.
Therefore, the Sxx parameters in a Smith chart start in the center, i.e. at 50, and turn
then to a curve around the value of Z0 of the strip line. For a lossless line, this curve is
represented by circles around the line's Z0. For a lossy line, this looks like a looping
towards Z0. Like with all S-parameters, this turning is always clock-wise.
708
For a Z0 < 50, the curve starts at 50, and turns with -90 clockwise, i.e.
downwards. If the line is lossless, we have a circling curve centered around Z0, and
touching again and again the center of the Smith chart, i.e. 50For a lossy line, we
have a looping around Z0 with the end point at Z0.
For the special case of a line with a characteristic impedance of Z0=50, lossless or
lossy, we have all S11 and S22 parameters in the center of the Smith chart: at 50.
For a Z0 > 50, the curve starts at 50, and turns with +90 straight upwards. If
the line is lossless, we have again a circling curve centered around the line's Z0, and
touching the center of the Smith chart, i.e. 50 again and again (if it is long
enough!)? If the curve is lossy, we will again have a looping towards the end point
Z0.
The Sxy parameters in the polar plot start at '+1', and also turn clock-wise.
For a lossless line of Z0=50, we have circles with magnitude '1' . For a lossy 50
line, the circles are replaced by a looping towards '0', because for a long, lossy line,
there is no signal reflected back any more.
If the Z0 of the line is <>50, we have a change in magnitude (see further below in
the next chapter), which is represented in the Sxy plot as ellipses for a lossless line,
or as an elliptic looping towards '0' for a lossy line.
For the modeling

we commence with setting all line parameters to default, and the length to infinite.
The loss parameter is set to '0'.
In this case, we can determine Z0 from fitting the magnitude of the simulated curve to the
trace of the measured data in the Smith chart. We ignore the phase.
For Z0 = 50 we have this situation:

Step1. Set length to infinite and adjust Z0 in smith chart
And for Z0 less or bigger than 50 we have:
Note
For Sxx, all curves start for freq=0 in the center of the smith chart i.e. at 50. Their end point at infinite
length is their Z0!
After having determined Z0, we set the physical length of the strip line to its real, physical
value and adjust the strip line's delay by fitting Sxy:
Step2. Set length to its physical value and adjust delay (beta) in Sxy
Then, we adjust the loss, again by fitting Sxy:

Step3. Adjust loss (alpha) in Sxy
709
If the starting point of Sxx is not at 50, and the starting point of the Sxy is not at '1', we
need to consider a series resistor (contact resistance) in series with the delay line model.
This is, last not least, the last remaining parameter of typical strip lines on the wafer:
Step4. Adjust Series Resistor Rs
Note
If a shift to the left is required in Sxx, split the delay line into 2 and insert a resistor to ground between
them!
Modeling a Strip Line Transition
When the width of a strip line changes, current crowding occurs. This is modeled by the L,
while the change in the step contour is modeled by the capacitance C.
710
Testing S-parameters For Typical
Passive Circuits
IC_CAP file: pass_RLC_TEE_or_PI.mdl
This chapter is intended to present some matrix manipulations for the interpretation of S-
parameters of typical passive LRC networks. This is done by transforming the S-
parameters to Y or Z-parameters. For these types of matrices exist simple interpretation
capabilities, provided their underlying circuits have a structure like in the following figure.
711
Test Fixture Modeling in Time Domain
When measuring a SMT (surface mount technology) or packaged device, the test fixture
contributes to the total measurement. Sometimes it is not possible to use dummy devices
like SHORT, OPEN, THRU with known cal-kit data.
Hint
a NWA calibration will only work correctly, if the 'non-idealities' of the cal-set components are known. This
is the so-called cal-kit data. This data has to be entered into the NWA before the instrument is calibrated
and the calibration result is stored in a cal-set of the NWA.
A nice work-around is to use a TDR (time domain reflectometer) in order to characterize

the effects of the test fixture in the time domain and to perform a de-embedding later in
the frequency domain. What has been a full 2-port calibration for a NWA, is now:
first the modeling of the TDR's step function in a THROUGH measurement,

second the modeling of the TDR cable length, an OPEN measurement.
and finally the modeling of the structure on the test fixture itself.
WHAT CAN BE MODELED USING A TDR?

Resistors: the back-reflected curve changes its level.
Inductors: the back-reflected curve overshoots.
Capacitors: the back-reflected curve undershoots.
Delay lines: the separation in time between reflections.
The starting point is again the S-parameter measurement of the parasitic structure
alone, see fig.3. As our test fixture contains a strip line, we have a lot of phase shift.
Our task is to 'turn back' this phase shift and to obtain the S-parameters of the open.
To keep things simple, we only consider S11.
712
Setup 'define_step_fct' of the model file in fig.2 is used to model the TDR step
function. Therefore, we connect the channel CH1 of the TDR (the very TDR channel)
with an TDR scope input (e.g. CH2), see fig.1. The measurement result is given
below:
The model corresponding to the measurement of fig. 4 is defined in the test circuit of
this setup: several SPICE pulse functions are used to model the step function, plus a
delay line. It looks like this:
.subckt mdltdr 101=tinp 105=toutp

* a series of step functions in the time*
* domain *
****************************************
v0 1 0 PWL(289p 1.375m 0.40404n 0)
v1 2 1 PULSE(0 1.5625m 0.4n 4.04p 4.04p 0)
v2 3 2 PULSE(0 -0.1875m 0.40404n 4.04p 4.04p 0)
v3 4 3 PULSE(0 0.375m 0.40808n 4.04p 4.04p 0)
...
...
v99 101 99 PULSE(0 -0.625m 0.79596n 4.04p 4.04p 0)
ri 101 100 50
t1 100 0 105 0 TD=1n Z0=50
ra 105 0 50
.ends
The delay time t1.TD=1n reflects the length of the actual connection cable.
In the next step, we disconnect the cable from CH2, and use another cable which we
will use later to connect CH1 with the test fixture. Then, we measure the back-
reflection on CH1 of the open cable. Setup 'define_tdr_zero' holds the result, what is
depicted in fig.4:
The SPICE circuit of this setup is the one from above, but now enhanced with the
correct cable delay time and components to model the open end.

v0 1 0 PWL(289p 1.375m 0.40404n 0)
...
...
v99 100 99 PULSE(0 -0.625m 0.79596n 4.04p 4.04p 0)
ri 100 101 50
r1 101 102 2
t1 102 0 103 0 TD=150p Z0=50
ra 103 0 43k
c1 103 0 1f
.ends
713
After these two pre-measurements, we are ready to go after the modeling of the test
fixture itself.
So, the cable is connected with the test fixture. As we model 'from the left to the right',
we first concentrate on the modeling of the connector of the test fixture. Setup
'define_conn' holds the result, what is given in fig. 6.
And again, the test circuit of this setup is enhanced to fit the measured data:

v0 1 0 PWL(289p 1.375m 0.40404n 0)
...
...
v99 100 99 PULSE(0 -0.625m 0.79596n 4.04p 4.04p 0)
ri 100 101 50
r1 101 102 1.1
t1 102 0 103 0 TD=188p Z0=50
ls1 103 104 460p
cp1 104 0 140f
r2 104 105 .01
tmic1 105 0 106 0 TD=500p Z0=50
rend 106 0 1G
.ends
Fig.6 also gives the simulation result: the inductor models the overshoot, and the
capacitor the undershoot.
Delay time t1.TD reflects now the delay of the cable (150ps) plus the connector (33ps).
Delay line tmic1.TD (for the modeling of the delay time of the strip line on the test fixture)
is set to 500ps and will be modeled in detail in the next step.
Finally, we consider the strip line and its open end on the test fixture. The TDR
measurement is kept in setup 'define_line' and is given in fig 7:
The test circuit of this setup is further enhanced and looks finally like:
714
v0 1 0 PWL(289p 1.375m 0.40404n 0)
...
...
v99 100 99 PULSE(0 -0.625m 0.79596n 4.04p 4.04p 0)
ri 100 101 50
r1 101 102 1.1
t1 102 0 103 0 TD=188p Z0=50
ls1 103 104 460p
cp1 104 0 140f
r2 104 105 .01
tmic1 105 0 106 0 TD=366p Z0=50
lend 106 107 700p
cend 107 0 120f
rend 107 0 4.5k
.ends
After all, we have the following model:
The final SPICE schematic is:
Now, we are done and can re-connect our test fixture with the NWA to perform a
measurement. With the S-parameters, we perform a de-embedding using the equivalent
schematic of the test fixture as developed in the time domain.
Therefore, the de-embedding follows this scheme:
transform S_to_Z: de-embed rl

transform Z_to_S: de-embed the delay of the connector (33ps)
transform S_to_Z: de-embed ls1
transform Z_to_Y: de-embed cp1
transform Y_to_Z: de-embed r2
transform Z_to_S: de-embed the delay of the strip line tmic1 (366ps)
transform S_to_Z: de-embed lend
transform Z_to_Y: de-embed cend
transform Y_to_S: to obtain the de-embedded S-parameters
To check the quality of this procedure, we apply the de-embedding to the measurement of
fig.3, i.e. the S11 parameters of the strip line at port 1 of the NWA.
Fig.10 shows the de-embedded result: the OPEN of the strip line on the test fixture.
715
For frequencies up to 3GHz, the de-embedded curve represents an ideal OPEN, while for
higher frequencies we might include in our SPICE schematic a special circuit the is able to
reflect dispersion effects. See chapter 'Modeling the skin effect' in the appendix.
716
Modeling of a TSOP44 Packagae
A step-by-step tutorial, based on the IC-CAP model file IC-CAP file: tsop44.mdl
The task is to model a TSOP44 package, using the following test fixture:
The test board, on the right side, shows that 3 pins of the TSOP44 will me modeled, while
the other pins are connected to ground.
TDR-related issues
Before we can start measuring the package with the TDR, we have to model the step
function of the TDR.
In this plot, we can see that the signal splitter in the TDR detects the emitted TDR step
function after 6ns. Until roughly 19ns, we have 200mV, what corresponds to a lossless
50 delay line. At 19ns, we see an open end and some power radiation (the 'creeping'
slope at the end reflection).
Note: This 'creeping' slope does not need to be modeled for our project, since the
connector at the end of the cable will be connected later to the test fixture, i.e. it will no
longer radiate a signal. However, for didactical reasons, Copen and Ropen are added to
the SPICE circuit of the section 'charact_TDR_pulse', and removed in all other sections of
the model file.
I.e. we can start with this subcircuit model:
As a SPICE circuit for the TDR pulse, with default parameter values, we define:
.subckt tdr_pulse 95
*
Tcable 95 0 96 0 Z0=50 TD=500p
Tconn_half 96 0 97 0 Z0=50 TD=22p
Lconn 97 100 1f
Cconn 100 0 1a
Copen 100 102 1f
Ropen 102 0 1G
*
.ends
Note
22psec for the delay of half a 3.5mm connector is a typical value.
717
Note
For the data shown in this tutorial, i.e. the data in file TSOP44.mdl, Lconn and Cconn are not required due
to the good performance of the connector.
For our example, the extracted parameters are:

Tcable.Z0 = 50
Tcable.TD = 542p
Tconn_half = 22p
with a pulse risetime = 44p
Note
Copen = 75f and Ropen = 780 for the intermediate, didactical modeling of the open cable radiation.
Here is the fitted result, referring to the cable end, i.e. the step from 200mV -> 400mV :
In the next step, we have to model the test fixture. It's layout is given below:
The TDR cable is connected to line 1 of the test fixture and we obtain the following
measurement result:
For the modeling, we will model all three lines, one after the other.
As can be seen below, the first SPICE model in this test fixture subcircuit is a delay line for
modeling the half of the connector. For the strip line on the test fixture, since there is a
knee in line 1, we will use two strip lines in series in the SPICE deck. Finally, a special
SPICE circuit will be used to model frequency-dependent effects which are not covered by
the ideal SPICE delay line model.
718
Let's commence with modeling the half of the connector, associated with the test fixture.
This is depicted below:
In the next step, we need to model the strip line on the test fixture. We will apply one
ideal SPICE delay line before the knee on the layout, and another one after the knee. The
modeled measurement is given in the next figure
There is still a mismatch for the slope of the OPEN end as well as for the overshoot. This
can be modeled with frequency-dependent losses of the strip lines. However, SPICE only
offers lossless strip lines. However, using the following subcircuit /after Katzier/, we can
also fit this region.
Applied to modeling the test fixture line 1 in the time domain, we finally obtain the
following modeled TDR performance:
What is left is the modeling of the crosstalk between the lines on the test fixture. This is
best measured using a network analyzer. Compared to TDRs (which are broadband
measurement instruments), a NWA has a better resolution, and also allows an easy
separation of crosstalk (Sxy) and main strip lines (~Sxx). On the other hand, using a TDR
offers the advantage to have a clear picture of the geometric sequence of the SPICE
elements, while in NWA plots, i.e. the Smich chart or the polar plot, we only see turning
curves.
Therefore, before modeling the test fixture lines cross talk, we verify the fitting of the Sxx
plot of the NWA. This is given in the next slide (Test Fixture Pin1: modeled with 2 strip
719
lines and freq. dependent losses)
We are now ready to model the crosstalk, referring to plot S21 (crosstalk of Pin1 (NWA
Port1) to Pin2 (NWA Port2)):
Test Fixture Pin1: modeled crosstalk to line2
In the same way, lines 2 and 3 of the test fixture are modeled. Here the result:
Please note that the test fixture modeling has only to be performed once for a given test
fixture!
After the accurate modeling of all the lines of the test fixture, we are finally ready to cover
the modeling of the TSOP44 package.
Since a package model, like a connector model, is represented by a big sub-circuit, it is

essential to agree first on some standardized SPICE circuit node names.
The following figure depicts a segment of a package model:
Such a segment is usually able to model a package performance up to ~500MHz. For

higher frequencies, more segments have to be cascaded. For package models up to 3GHz,
typically 3 segments are cascaded.
Standardization of SPICE node names:
we use 3-digit node names:

package pin 1: the outer package node is 100,
720
the first segment is from 101 to 110
the next from 110 to 120
the last from iii to 199
package pin 2: the outer package node is 200,
the first segment is from 201 to 210
etc.
This leads to the following sub-models for the TSOP44 package:
Modeling considerations for our example:
We will only model pin4

For simplicity of this demo, we assume that all pins of the TSOP44 package have an
identical model
We commence with modeling the SPICE components along the pin. The start values of the
model parameters could either come form 3-D field simulations, or, like in this example,
from some raw hand calculation formulae.
Line capacitance towards ground plane and line inductance:
721
Taking 2 lines inter considerations, we get for the capacitance between them:
Equations from : Nhrmann, Das grosse Werkbuch Elektronik, Franzis-Verlag, Munich,
1984, ISBN 3-7723-6544-2
This gives the first modeling result for the TDR measurement of the package, pressed with
an electrically neutral stamp onto our test fixture:
Equations from : Nhrmann, Das grosse Werkbuch Elektronik, Franzis-Verlag, Munich,

1984, ISBN 3-7723-6544-2
This gives the first modeling result for the TDR measurement of the package, pressed with
an electrically neutral stamp onto our test fixture:
Fine-tuning the line capacitances towards ground and the inductance along the package
pin, we obtain a tuned result for this TDR measurement like in the plot below:
The remaining fitting problems are assumed to be related to cross-coupling between the
lines. For resolution reasons, as mentioned already further above, we therefore switch to
NWA measurements.
With the NWA calibrated to the ends of its two cables, and with the test fixture already
modeled, as well as with a first estimation of the SPICE components along the package
pin, we can fine-tune the crosstalk PICE components between our package pin and its two
direct neighbor pins, and obtain:
722
In a second step, we now also tune the SPICE parameters representing the cross-talk to
the next-but-one neighbor pins of the package. The plot below shows the result:
With this fitting result, we go back to the TDR measurement and check the fitting there. If
required, a small parameter fine-tuning is performed here and back in the S-parameter
domain as well.
The plot below depicts the fitting result for the TDR measurement:
Publications
Modeling and Simulation of high pincount connector systems, T. Gneiting, H. Khakzar,

Proc. 4. European IC-CAP User Meeting, Berlin 1997
H.Wolter, H.Katzier, O.Podebrad und T.Weiland: Field-Based Analysis of a High Pincount

Board Connector; IEEE Transaction on Components, Packaging, and Manufacturing
Technology, Part B. Advanced Packaging, Feb. 1996, Vol.: 19, No. 1, 23-31.
H.Katzier, R.Reischl und P.Pagnin: SPICE-Models for High Pincount Board Connectors;
IEEE Transaction on Components, Packaging, and Manufacturing Technology, Part B.
Advanced Packaging, Feb. 1996, Vol.: 19, No. 1, 3-6.
H.Katzier, R.Reischl: Elektrische Simulationsmodelle fr hochpolige Leiterplattenverbinder;
723
Elektronik, Nov. 1997, No. 24, 104-112.
Mark S. Mirotznik and Dennis Prather: How to choose EM software. IEEE Spectrum, 12,
1997, 53-58.
H.Katzier, R.Reischl, Siemens AG, Munich, Reliable SPICE Models for Lossy Coupled
Transmission Lines, , 2000 High-Performance System Design Conference
R.Reischl; Modeling of BGA-IC-Packages; European ICCAP-User-Meeting, 17-18. June

1999, Marseile, France.
724
Low Frequency (1/f) Noise Modeling for
Semiconductors
Contents
Introduction (iccapmhb)
Types of Noise in Semiconductor (iccapmhb)
Noise Models in Semiconductors (iccapmhb)
Low Noise Measurement Setup (iccapmhb)
Low Noise Measurements and Modeling (iccapmhb)
Noise Parameter Extraction and Verification (iccapmhb)
Appendix (iccapmhb)
725
Introduction
Phase noise of oscillator circuits is among the key parameters of today's communication
systems. It limits the modulation quality of the information signal, and the cross-talk to
adjacent channels. In the receiver, on the other hand, it can reduce the selectivity and the
demodulation quality. With today's trend to even more complex modulation schemes, low
phase noise becomes even more critical. Therefore, minimizing noise with the design of
high frequency communications circuits is a must for the cost-effective exploitation of
limited bandwidth, and to obtain low bit error rates.
In order to predict the noise behavior of such systems correctly, accurate noise models
are required. Without them, the design and optimization of an amplifier's noise figure or
the phase noise of oscillators cannot be successful.
A key element is the transistor. While for linear circuits the modeling of the noise at the
operating frequency f0 is sufficient, non-linear circuits do also convert the low-frequency
noise up to the operating frequency range. For oscillators as an example, the 1/f low
frequency noise at fm will be mixed upwards to contribute to the phase noise of the total
circuit at the frequencies f0_fm and f0-fm . Therefore, noise modeling of transistors can
be split into a high and a low frequency segment. The graph below depicts the Collector
low-frequency noise spectral density of a bipolar transistor.

It can be seen that the resolution of the toolkit measurement setup is at about .
Broadband noise in bipolar or FET models is determined essentially by thermal noise and
shot noise. These noise sources are automatically determined within the model from the
large signal model parameters. An important prerequisite for this is, however, that the
noise determining parameters, e.g. the Base resistor of bipolar transistors, has been
determined carefully and that their values have a physical meaning. The low frequency
noise, on the other hand, dominated by the 1/f noise, is modeled by some specific model
parameters. In general, these are the AF and KF and sometimes BF/EF parameters,
covered in this chapter. With some models, like the BSIM3v3, an alternate 1/f noise model
is available too [10,15].
Noise Term Definitions

In order to keeps things simple, we consider white noise of a resistor. This frequency-
independent and bias-independent noise is described in terms of a voltage by the well-
known formula
where,
Since the term 4kTf represents a power, and if we normalize this power to f, we end up
with a power density. Usually, this normalization is done for f=1Hz.
This result can be plotted against frequency. In case of our resistor, it is a constant value.
For semiconductors, such a plot can also exhibit frequency dependencies, e.g. 1/f noise.
In any case, such a plot shows a spectrum.
Therefore, this term is called power noise spectral density.

However, this is not exact, referring to the dimensions. A more appropriate term is
therefore voltage noise spectral density. Its symbol is usually Snv, and its dimension is
726
In some cases, the square root of Snv is used, called the equivalent noise voltage,
Note: Using Env instead of Snv does not require to calculate noise transmissions/amplifications using
squares of R, amplification etc.
When expressing noise in terms of currents, e.g. ,the same terminology
applies.
Note: When performing circuit calculations based on noise spectral densities, keep in mind that you have
to calculate with squares instead of the commonly used expressions.
727
Types of Noise
This chapter deals with the description of the different noise mechanisms and their noise
density spectrum in semiconductor components. It is basically aimed to highlight the main
fundamentals rather than to explain the individual physical effects. We start with the
thermal noise of the ohmic parasitics, and lead then over to the semiconductor-specific
noise sources.
Thermal Noise
Related to the thermal oscillation of electrons in a resistor, we can measure a current at
its contacts, without applying an external voltage. This is the thermal noise current. The
effective current is described by
An equivalent schematic for this condition is a noise-free resistor with a current noise
source in parallel.
In analogy to this circuit, a noise-free resistor and a serial noise voltage source can be
used as well. In this case, the effective voltage is given by
Equivalent schematic of a noisy resistor
Note: since noise sources have a mean value of '0', but an effective value of usually non-zero, effective
voltages and currents are used to calculate the circuit performance with respect to the noise sources. This
means that such an effective voltage is amplified by the square of the amplifier's gain. As another
example, two parallel current noise sources can be represented by a single noise current source with a
value of
Also, the ohmic law is now referring to the square of voltages and currents, see
Equation(2).
Back to the thermal resistor noise, the equivalent noise power density spectra are
These frequency independent noise spectra represent a simplification. An accurate

calculation based on a quantum mechanic model gives
so that equations (1) to (4) are basically only valid for h*f<<k*T , i.e. for 'low'
frequencies and high temperatures. However, the quantum noise for h*f>>k*T has to be
considered basically only for frequencies very much higher than in RF and microwave
applications, i.e. the >1013Hz range.
Shot Noise
The current through the space charge area of a diode is composed of many individual
current impulses, due to the transport of individual charge carriers. Since this motion of
electrons/wholes is statistical, we always have, besides the expected DC current, also a
noise component. With the assumption of individual, rectangular current impulses of the
width for every charge component, we can calculate a power density spectrum after
[1,2]
For low frequencies, i.e. small values of * f , there is sin(x) /x ~ 1 and we get the
commonly used equation
for the shot noise. The figure below shows the equivalent schematic of an ideal, but noisy
728
diode. The noise current is described by a current source in parallel with the small signal
conductance.
Generation-Recombination Noise
In semiconductors, there is a static charge carrier generation/recombination process. This
refers to a statistic change of the carrier states. Referring to the energy band model,
transitions between conduction and valence band as well as various trap levels are
possible, leading to a noise spectrum following
The current dependency of the combination/recombination noise, also called burst-noise

or popcorn-noise [3], is generally reflected by a modeling equation like
with the modeling parameters KB, AB and FB.
1/f Noise
A pretty often measurable phenomenon is noise with a spectrum proportional to 1/f. This
leads to the name 1/f noise. Another name is flicker noise. It is caused essentially by
recombination effects at defects in the semiconductor volume, the borders of diffusion
areas or the material surface.
An empirical description after Hooge [4, 5] is a spectrum with
where Ntot means the total number of moving charges in the device. The Hooge-
Parameter is a material characteristic.
In most simulation programs, the current dependency of the 1/f noise is covered in
analogy to the generation/recombination noise by an exponential form like
with the model parameters AF, KF and B. B is commonly set to '1'.
729
Noise Models in Simulation Programs
When simulating the noise behavior of circuits, all circuit components need to have noise
models included. All lossy components will exhibit thermal noise, corresponding to the
simulation temperature TEMP. Semiconductor devices will additionally also exhibit 1/f
noise.
Resistors
Their noise distribution is modeled by white noise, as described above i.e., the effective
current is described by
representing a noise-free resistor with a current noise source in parallel.

Alternatively, a noise-free resistor and a serial noise voltage source can be used as well.
Inductors, Capacitors, and Lines

Inductors, Capacitors, and Lines are considered noise-free. This is valid for ideal
components. For real RF components, including parasitic losses (SPICE sub-circuits), the
noise contributions stem basically from the parasitic resistors.
Diodes
The main noise source of diodes is shot noise. Besides this, 1/f noise as well as thermal
noise due to the parasitic resistor is observed. The capacitance CDiode, covering the space
charge and the diffusion capacitance, does not contribute to the noise.
This leads to a small signal schematic with the noise sources.
Since the shot noise formula refers directly to the diode current Id and the thermal noise
to the parasitic resistor Rpar , it is only the parameters AF and KF which have to be
modeled additionally.
Bipolar Transistors
SPICE and similar simulators feature a noise schematic for bipolar transistors.
Associated with each physical resistor (Rb at the Base, Rc at the Collector and Re at the
Emitter), is a thermal noise source
730
Base and Collector currents are considered to be independent. Therefore, the shot noise
can be described by a source
at the Base and
at the Collector. The total 1/f noise of the transistors is described by a noise source in
parallel with the Base-Emitter contact. This leads to the noise source at the Base following
with, again, the noise model parameters AF and KF.

Agilent's Advanced Design System (ADS) include with the BJT transistor model
additionally a generation/recombination noise effect in the noise spectrum of the Base.
Referring to equ. (17), this means
with the new model parameters KB, AB and FB.

With PSPICE, the Collector current ic includes an additional 1/f noise source. This gives
using the same parameters KF and AF of the Base contact.

The VBIC model for bipolar transistors [6] includes, similar to the UCB SPICE BJT model,
thermal noise sources as well as shot noise for the Base-Emitter current and the transport
current. The 1/f noise is described again by a noise source between Base and Emitter with
However, the frequency dependence is additionally modeled by the parameter BF. Besides
the main NPN transistor, VBIC includes also a parasitic PNP transistor. Its noise
contribution is modeled in analogy to the NPN, sharing the same model parameters for the
1/f noise. More details can be found in [7,8].
Noise in the MEXTRAM model is described like with the SPICE BJT model [9]. Besides
thermal noise, there are shot noise sources associated with the Base-Emitter current as
well as the transport current. 1/f noise is described after (17).
Junction FETs
The noise schematic of field effect transistors is most often as shown below
The noise behavior is described by four noise sources. The resistors RD and RS are
associated with thermal noise after
The amplifier effect of JFETs is based on changes in the channel resistor. This leads to
describing the noise of the channel current ID also by thermal noise
Additionally, an 1/f noise source is assumed with the channel after
with the model parameters AF and KF. The shot noise of the Gate current is usually
neglected.
This model is valid for the simulators SPICE2, SPICE3, PSPICE and ADS. With HSPICE, an
additional series resistor at the Gate is assumed with corresponding thermal noise. For the
731
thermal noise of the channel, there is a choice to replace equ.(23) by a more accurate
model [3].
it differentiates between linear and saturated region. And, with GDSNOI , an additional
noise parameter is introduced.
MOSFETs
The noise formulation of MOSFET models in most of the commonly used simulation
programs is based on the SPICE2 model of the UCB University of California, Berkeley,
described in [3]. It is known, however, that the thermal channel noise included in this
model is essentially valid only in the saturated region of the output characteristics [11].
Therefore, the latest UCB model, the BSIM3v3 model [10], models the noise differently,
see further below.
The figure depicts the noise equivalent schematic of the SPICE2 MOS2 and MOS3 model.
Like with the JFET transistor, the resistors RD and RS are associated with thermal noise
after
The channel noise is described by
using again the parameters AF and KF for the flicker noise. Cox and Leff are calculated
inside the model from the other model parameters.
In PSPICE, an additional thermal noise of the Gate and of the Bulk resistor is included.
With ADS, the models MOSFET Level 1-3 and EEMOS1 are also based on figure above.
With EEMOS1, however, no 1/f noise is included.. With HSPICE, like with the mentioned
JFETs from above, the user has the choice between this and an alternate noise
description. The thermal channel noise is described again after equ.(25), with the
exception
With the BSIM3v3 model of UCB SPICE3, the user can select for both, the flicker noise as
well as for the thermal channel noise, one of the following models. This is done by setting
the model parameter noimod accordingly. Either, a slightly modified SPICE2 model
describes the 1/f noise by
Here, with the same functionality like in the VBIC bipolar model, an additional parameter
EF has been introduced, modeling the frequency dependence of the noise spectrum. The
thermal channel noise is calculated in BSIM3v3 after
Or, depending on the parameter NOIMOD, the BSIM3v3 noise model calculates the 1/f
noise after
using a relatively complex formula, and involving the five new noise parameters NOIA,
NOIB, NOIC, EF and EM as well as large signal model parameters [34].
For the channel noise, it is
732
The procedure described further below in this paper, covering the 1/f noise parameter
extraction, can therefore be applied to the BSIM3v3 model only if the SPICE2 model for
the 1/f noise has been selected specifically by the parameter NOIMOD. For extracting the
parameters NOIA, NOIB, NOIC, EF and EM, see the BSIM3v3 toolkit for IC-CAP.
MESFETS
The noise behavior of MESFETs is modeled in the simulation programs PSPICE and HSPICE
essentially analogous to the corresponding JFET -models [3]. See [12] for more details.
As a final remark, the noise bandwidth f is set to f=1Hz in most simulators.
733
Low (1/f) Noise Measurement Setup
When measuring noise of high frequency transistors, we continue to distinguish between a
low frequency range up to several MHz, followed then by a second range up to the
maximum operating frequency of the device (GHz). In the first range, the noise power
density spectrum is measured in order to characterize the 1/f noise and the
generation/recombination noise. In the upper range, the noise figure Fmin, and the
parameters Rn , Yopt are measured.
We will now discuss a possible measurement setup for the first range, the 1/f noise
characteristics. It is depicted in figure below. It consists of an HP4142 or HP415x SMU,
which, however, is operated in 'power supply' mode. This means that it is triggered to
output either a current or voltage (depending on the type of transistor), and to keep this
current or voltage switched on, until the 1/f noise measurement is over and a new 'stop'
command is sent to the SMU. The SMU is connected to the Gate or Base of the transistor
by a 1Hz filter. The SMU is connected to the Gate or Base of the transistor by a 1Hz filter.
Its output resistance is switchable. For MOS transistors, a low output impedance of the
filter is required (e.g. 50 ), while a high output impedance (e.g. 500k ) is used for
bipolar transistors.
Note: for a bipolar transistor, the 1/f noise is generated in the Base region. If the 1Hz
filter had a low output impedance, it would shorten this noise source and we would
measure/simulate a too low 1/f noise characteristic. The higher the 1Hz filter's output
impedance is, the less shortening occurs, see also [14]. In the toolkit, a value of ~330k
has been found to work best.
The output of the transistor is connected to a special low-noise current amplifier. Besides
amplifying the 1/f noise, this amplifier also serves to bias the Drain or Collector of the
transistor, and also to compensate the DC Drain or Collector current with an ultra-low
noise DC current source. This allows to amplify only the noise current of the transistor:
With this setup, the cabling is minimized. Only 2 short coax cables are used to bias the
transistor and to measure the noise. Therefore, extremely reliable and extremely clean
noise measurements are obtained.
The output of the noise amplifier is further connected to an HP35670 dynamic signal
analyzer, which then converts the 1/f noise time signals into a frequency spectrum.
The whole setup is controlled by special IC-CAP macros for bipolar transistors, and for
MOS transistors. Ready-to-use extraction routines for the noise parameters AF, KF and EF
(for MOS) are included as well.
Note: For on-wafer measurements, special attention has to be paid regarding shielding. It is
recommended to use a prober with a special 1/f shielding. However, most important, it is recommended to
use the G-S-G RF probes instead of using DC probes
1Hz Low Pass Filter
734
Measurement Setup
The 1/f noise modeling toolkit contains all required IC-CAP measurement setups and
driver macros for measuring the bipolar and MOS transistor 1/f noise and to extract the
model parameters. Figure below shows its structure, followed by the macros for
measurements and parameter extractions.
Measurement resolution of the setup:

The current noise spectral density resolution obtainable with this system has been
measured as ~5E-21 .
735
1/f Noise Measurements and Modeling
This presentation is about the bipolar and MOS transistor 1/f noise modeling toolkit before its
implementation into IC-CAP as Agilent product number 85195B.
Introduction
Design Goals For Minimum Noise Circuits in RF Applications
Low frequency noise:'classical' model parameters AF, KF

High frequency noise,above 1/f noise:noise figure Fmin
Low Frequency 1/f Noise

As shown in the figure below an effect of low frequencies, this type of noise is upconverted
in mixers and VCOs, and can affect the overall circuit performance considerably.
Noise dependencies in Semiconductors

1/f Noise: Operating Point and Frequency Dependent
White Noise: Operating Point Dependent
736
Description of Noise Signals

Noise signals exhibit power, and therefore, their quadratic mean value is non-zero
This means that noise signals can be described with effective values
However, since the time function of noise is unknown, the conventional formula for
effective signals in the time domain cannot be applied.
The Wiener-Kintchine theorem permits to associate the auto-correlation function
with the spectral power distribution
This means, noise is best described by its spectral distribution function

Therefore, we measure the noise spectral density per 1Hz:
Voltage: Unit:
Current: Unit:
The resistor effective noise voltage (root means square value), in , is
The term 4kTf represents a power. We normalize this power to f, what gives a power
density.
Usually, this normalization is done for f=1Hz.
This result can be plotted against frequency. In case of our resistor,it is a constant value.
For semiconductors, such a plot can also exhibit frequency dependencies, e.g. 1/f noise.
In any case, it shows a spectrum.
Therefore, this term is called voltage noise spectral density. Its symbol is Snv, and its
dimension is
Circuit Calculations with Noise Spectral Densities
737
When performing circuit calculationson noise spectral densities, or effective noise voltages/currents, keep
in mind the unit is a square: , , , or etc.
Therefore:
Noise Mechanism
Thermal Noise
Thermal Noise is current through a resistor
Lossy passive components are known to produce temperature dependent and bias
independent white noise due to the thermal induced motion of charge carriers.
Shot Noise
Shot Noise is current through a semiconductor junction
Semiconductor junctions are known to generate noise caused by the quantized and
random current flow across them. This (typically) white noise is bias-dependent but
independent of temperature
1/f Flicker Noise

This low-frequency, bias-dependent noise effect is typically attributed to traps associated
with contamination and crystal defects.
738
Generation/Recombination Noise (also called: burst or popcorn

noise)
This noise is due to generation/recombination effects and trapping and is observed in
some semiconductor devices when other noise effects are minimal. It is bias and
frequency dependent.
The Shape of White Noise and Flicker Noise
Noise Models
739
Diodes
1/f noise contributions
Thermal Noise: Rpar
Shot Noise and 1/f Noise: Id
Noise Models in Simulators
Bipolar Transistors
Thermal Noise: RB, RC, RE
Shot Noise and 1/f Noise: Ibe
Shot Noise: Ice
Junction FETs
Thermal Noise: RD, RS
Thermal Channel Noise and 1/f Noise: ID
740
MOSFETs
Thermal Noise: RD, RS
Thermal Channel Noise and 1/f Noise: ID
MOSFETs: BSIM3v3.2
1/f Noise and Channel Noise:
conventional, like in MOS3:
BSIM3v3 specific:
Model Choice:
Parameter noimod selects a combination of the four modeling possibilities:
741
The 1/f Noise and its Model Parameters

The thermal and shot noise is determined inside the model from the other model variable
values.
NOTE: This implies that only physical meaningful values for the resistor and DC large signal parameters
will finally give an accurate modeling of these types of noise sources.
This means, only the 1/f noise can be modeled individually.

The modeling of the 1/f Noise follows for:
All Diode Models,

Bipolar Models like the Gummel-Poon, VBIC or MEXTRAM,
JFET Models,
MOSFET Models
BSIM3v3/BSIM4 when selecting the conventional Noise 1/f-Model
Extraction Method for the Parameters AF, KF and B is required.
1/f Noise Measurement Toolkit in IC-CAP

Applied 1/f Measurement Setup
1Hz Lowpass Filter

The filter is required for biasing the transistor input.
It offers -60dB attenuation at 50Hz.
Required Output Impedance Rout (to avoid shortening the 1/f transistor noise):- Rout
=330 k for bipolar (in any case Rout > Rbe_noise_bias),- 50 for MOS transistors
Agilent 4142B or 415xB

Operated as a power supply with two special macros in IC-CAP:
Voltage or current ON
Performing the 1/f measurement, typ. 30 sec., with dyn. signal analyzer
Voltage or current OFF
742
Stanford Research SR570 Current Preamplifier

1 pA/V maximum gain
Adjustable bias voltage with test point +- 5V max
Variable input offset current +-1pA .. +-5mA
Low Noise, High BW, and Low Drift modes 1 MHz maximum bandwidth
Agilent 35670A
The 35670A is a portable two- or four-channel dynamic signal analyzer
102.4 kHz Bandwidth

Measurements: Linear Spectrum, Cross Spectrum, Power Spectrum, Coherence,
Power, Spectral Density, Frequency Response, Time Waveform, Auto-correlation,
Cross-Correlation, Histogram etc.
ON-WAFER MEAS. SETUP
IC-CAP Toolkit for Bipolar and MOS
743
The slide above and the following IC-CAP screenshot slides describes the 1/f bipolar and
MOS transistor toolkit before its implementation into IC-CAP as Agilent product number
85195B.
Ready2go, open measurements and extraction routines (user modifiable)
Intuitive User Interface
By default, 5 DC bias points at mean_vd are measured for 1/f noise, plus 3 a low and 3 at
high vd for verification.
(can be changed by user if required).
1/f Parameter Extraction

Dominant 1/f noise source location:
BIPOLAR input current Ib
744
MOS output current Id
Bipolar Transistors vs. MOS Transistors

Since the 1/f modeling formula refers to ib for bipolar transistors, a separate
measurement of = Ic / Ib is required in order to transform the noise, measured at the
Collector, to the Base.
MOS
Bipolar
Examples
MOS Transistor
1st DC condition
All DC conditions
745
Analyzing the slopes
Measured and modeled 1Hz points
Fitted model
746
1/f modeling at a glance: multiply Noise Spectrum by freq
Bipolar transistor
1st DC condition
All DC conditions
747
Analyzing slopes
Fitted model
Bipolar measurement result with 50 averages (instead only 10) of the dyn.signal
analyzer
748
Verification Issues
A note on data consistency checks for MOS transistors.
For MOS-Transistors, vg_noise is independent of the applied bias. Therefore, calculating
vg_noise=id_noise/gm^2 should give overlying curves .
*See: M.T.Yang et.al., Characterization and Model of On-Chip Flicker Noise With Deep N-
Well Isolation for 130nm and Beyond,ICMTS 2005, Leuven, Conference Proceedings
Noise measurements of different metal film resistors (low 1/f contribution), were executed
and compared with noise calculations. Measurements agreed very well with calculations.
Conclusions
Modeling of 1/f noise is important for accurate design of tomorrow's communication
systems.
A precise modeling strategy for most transistor models, using AF and KF, has been
demonstrated.
A ready2go IC-CAP toolkit is available.
Toolkit Parts incl. TestFixture for verification with packaged standard device
749
750
AF, KF and BF/EF Noise Parameter
Extraction and Verification
Considering the parameter extraction, we refer to the method introduced in [13, see also
14]. Following this approach, there is no need to exactly determine the corner frequency
fc like with other proposed methods. For the corner frequency fc, the 1/f noise equals the
white noise. For frequencies below that, the 1/f noise is dominant, allowing to neglect all
other noise sources for the further analysis. The current noise source of the transistor can
now be calculated out of the measured noise at the output of the trans-resistance
amplifier.
Bipolar Transistors:
For the bipolar transistor models, the origin of the 1/f noise is the Base region, see
equation (17). However, the effective 1/f current noise spectral density is
measured at the Collector of the transistor. Therefore, the 1/f noise at the Base has to be
calculated first after
To begin with the parameter extraction, we first repeat the formula of the 1/f effective
noise current generated at the Base (equ.17)
Note: the VBIC model features an additional 1/f noise parameter, BF, see equ.(20). It acts like the EF
parameter in the BSIM3v3 model: to fit the -10dB/decade slope of the measured 1/f noise. For details
about its extraction, please refer to the next section on MOS transistors.
In order to match equ.(37) to equ.(36), we normalize to f and set f = 1Hz. This gives
the Base current noise spectral density.
Since the 1/f slope is a 'given' for our actual modeling problem, our next step is to get rid
of it by multiplying the measured curve with the frequency points 'f'. This results in a flat
trace where we had the 1/f slope before.
The advantage of this method is that we are now easily able to identify the value of the
1/f noise at 1Hz, what will be used in the next step. The 1Hz noise value is simply
calculated as the mean value of a maximum flat sub-range of these so transformed data.
Considering the extrapolated measurement result at 1Hz, the above formula simplifies to
This means, we are now ready to obtain an 1Hz value of our 1/f noise for each bias
condition .
In the next step, we draw these values against the bias current. We apply a logarithmic
conversion to the above formula and obtain
what can be interpreted as a linear function like
where
and the DC bias current at the Base is transformed by

A linear regression is applied, which returns the y-intersect 'a' and the slope 'b' of a best
fitting line for equ.(42).
The noise parameters AF and KF are then calculated after
and
Step-by-step modeling procedure for bipolar transistors

1. For a first DC bias condition (e.g. iB_DC = 1uA, vCE = 2V), the 1/f noise spectral
density is measured. As stated above, for the bipolar transistor, the 1Hz Base filter's
output impedance is set to a high value, e.g. 330k
751
NOTE: This value must be considerably bigger than the input resistance
Otherwise, it would shorten the 1/f noise current source at the Base.
The following plot shows the measured noise current at the Collector:
2. This is repeated for e.g. 5 more different Base currents, but the same vCE.
And we get five 1/f noise curves for each iB_DC bias condition:
3. We have so far considered the noise current data measured at the Collector of the
transistor. The 1/f noise source of a bipolar transistor is, however, located and
modeled in the Base region. Therefore, we have to divide the above obtained
Collector current noise spectral density SiC by beta2, and obtain SiB, see equ.(36)
above.
4. After we have obtained SiB at the Base, we multiply it by 'freq' , see (39), and
obtain:
Thus, the 1/f noise appears as flat traces and can easily be identified.
5. We calculate the mean value of each transformed 1/f curve for each Base bias
752
condition from the max. flat traces. These mean values represent the 1Hz values of
the 1/f current noise spectral density, as a function of the iB_DC bias.
6. Finally, we are ready to draw the 1Hz Base noise data points against the DC bias i
B_DC, and fit a line to these data points, see equations (41) and (42).
From the y-intersection and the slope of the fitted curve, we calculate AF and KF
after equations (43) and (44).
7. After the model parameters have been obtained, the simulation result of the Collector
current noise spectral density is compared with the original measured data, and the
AF and KF model parameters are fine-tuned.
Measured and simulated Collector noise spectral density for the bipolar
transistor.
MOS Transistors
The 1/f noise description for MOS transistors in this chapter refers to the AF, KF, EF
formulation. For details about how to model the 1/f noise in BSIM3v3 using the
parameters NOIA, NOIB etc., see publication [15].
What makes the modeling easier compared to the bipolar case, is that the dominant 1/f
noise source is located at the output of the transistor, i.e. the Drain-Source channel. On
the other hand, the transistor Gate impedance is high. Related to the measurement setup
and our 1Hz filter at the Gate, this corresponds to a low output impedance. We choose
50.
On the other hand, a 3rd model parameter is used to reflect slight differences from the
theoretical 1/f slope: the parameter EF.
Also, different to the bipolar case, the 1/f noise formulation of the BSIM3v3 model
includes some geometry parameters.
As with the bipolar case, we first refer to the formula for the 1/f Drain-Source effective
753
noise current as
with
or simplified:
Again, we normalize to f, then set f = 1Hz and obtain the Drain-Source current noise
spectral density in .
As the first modeling step, we extract the parameter EF, the 1/f slope correction. A log
conversion of equation (45) gives:
i.e. we convert both the measured noise data and the frequency points logarithmically and
apply a regression curve fitting. The parameter EF is then the '-slope' of the fitted line.
Since the slope is now already modeled, we can get rid of it by multiplying the
measured curve with the frequency points . This results in a flat trace where we had
the slope in the measurements before.
After this step, we are again easily able to identify the value of the noise at 1Hz: it is
simply calculated as the mean value of a maximum flat sub-range of these so transformed
data.
Considering the extrapolated measurement result at 1Hz, the above formula simplifies to
This means, we are now ready to obtain an 1Hz value of our noise for each bias
condition i D_DC.
In the next step, we draw these values against the i D_DC bias current. We apply a
logarithmic conversion to the above formula and obtain
what can be interpreted as a linear function like
where
and the DC bias current at the Base is transformed by

A linear regression is applied, which returns the y-intersect 'a' and the slope 'b' of a best
fitting line.
The noise parameters AF and KF are then calculated after
and
Step-by-step modeling procedure for BSIM3v3 CMOS transistors
1. For a first DC bias condition (e.g. vG = 0.6V, vDS = 1V), the 1/f Drain current noise
spectral density is measured. The 1Hz Base filter's output impedance is set
to 50. The following plot shows the measurement result:
754
2. This is repeated for e.g. 5 more different Gate voltages, but the same vDS.
And we get five 1/f noise curves for each vG DC bias condition:
3. In the next step, we extract the EF parameter after (47), and check if the slopes of
the intermediate simulations match.
Note: the AF and KF have not yet been extracted, therefore only the slopes are important to
compare.
4. Now, we are ready to multiply by in order to easier extract the 1Hz value of the
noise.
5. We calculate the mean value of each transformed 1/f curve for each vG bias
condition from the max. flat regions. These mean values represent the 1Hz values of
the 1/f Drain current noise spectral density, as a function of the iD_DC bias.
755
6. Finally, we are ready to draw the 1Hz Drain current noise data points against the DC
bias current iD_DC, and to fit a line to these data points, see equations (50) and
(51).
From the y-intersection and the slope of the fitted curve, we calculate AF and KF
after equations (52) and (53)
7. After the model parameters have been obtained, the simulation result is compared
with the original measured data, and the AF, KF and EF model parameters are fine-
tuned.
Measured and simulated Drain current noise spectral density for the
MOS transistor
The noise data returned from hspice and spice3 differs from the noise data returned from
hpeesofsim, mns, and spectre.
hspice and spice3 have units of V^2/Hz and the other simulators have V/Sqrt(Hz).
756
Noise floor information on the toolkit
measurement setup
The figure above shows the current noise density of the system measured at the device
output (at the low noise current amplifier LNA).
The noise is expressed in A^2/Hz and varies with the LNA sensitivity. The figure shows
the noise floor for the 4 most commonly used values of the LNA sensitivity: 20 uA/V, 50
uA/V, 100 uA/V and 200 uA/V.
Comments:
The 1/f noise observed at the beginning of the trace is due to the internal
1/f noise of the LNA.
The noise drop observed at low sensitivity values (high gain) is due to the
bandwidth limitation of the LNA. Note that this is not in the frequency band
used for the extraction (typically between 10 Hz and 1 kHz). Using a
sensitivity of 200 uA/V or greater does not have this limitation but on the
other hand, increases the noise floor.
This noise floor should be compared to the output current noise density of
the DUT . If the DUT is a CMOS, this would be the drain current noise
density. If the DUT is a bipolar, this should be the collector current noise
density. Note that in the bipolar case, the current noise source is actually
modeled at the input (base current). The Ib noise density is determined by
dividing the Ic noise density by the DUT current gain (squared).
Parts list for 1/f noise toolkit

Instruments:
1 Wafer Probe Station with GSG probes (alternatively shielded DC probes)

1 Agilent DC Analyzer 414x or 415x (input biasing for transistor)
1 Custom made Lowpass Filter (1Hz), input triax, output coax or 3.5mm connector
following the recommended circuit description
1 Stanford Research Amp SR570
1 35670A Dynamic Signal Analyzer
1 Multimeter for controlling the DC bias provided by the SR570
Cables:
1 Cable: filter output (coax or 3.5mm) to GSG probe (3.5mm)

1 Cable: GSG probe (3.5mm) to SR570 (coax)
1 Cable: SR570 (coax) to 35670A (coax)
IC-CAP:
85199A (framework), 85199B(analysis module: simulator interface)

85199D (DC analyzer driver)
85199G (35670 driver),
85195B (1/f noise toolkit for AF, KF, EF parameter extraction for NPN-BJTs and MOS)
For PNP/PMOS transistors, the user has to make the changes related to the inverse voltages by himself.
For bipolar transistors, the noise measured at the collector has to be transformed to the
base: i.e. dividing by .
This , however, is the AC-beta d(ic)/d(ib), and not the DC-beta ic/ib.
757
Explanation of Spice circuit deck for 1/f noise

simulations
SPICE Circuit Deck in the DUT Test Circuit:
.subckt _noise 1=C 2=B 5=L 99=AUX

*satisfy SPICE for a noise simulation
Raux 99 0 1
*emulate the bias lowpass filter at the Base/Gate
RB1 2 21 10k
CB1 21 0 100u
RB2 21 22 10k
CB2 22 0 100u
*here is the filter output resistor
* -high value for bipolar transistors: bigger than biggest rBE
* - 50 Ohm for MOS
RB3 22 23 330k
*here comes the call to the transistor subcircuit
X1 11 23 0 noise_1_f_bip
*add a dummy 0V source to sense the transistor noise current
V1 1 11 0
*convert the Collector noise current into an identical voltage
H1 5 0 V1 1
Rspice 5 0 1
.ends
The netlist implements a workaround for probing the current noise density flowing through
the collector node because the simulator cannot output current noise densities directly.
Here is the detailed explanation of the circuit:
The external nodes of the circuit are 1=C, 2=B, 5=L , 99=AUX
The AUX=99 node is just an auxiliary nodes used only to run an AC simulation. In the
IC-CAP setup verify/simulate, look at the input sweeps: the node AUX is connected to
an AC generator. We need to run an AC simulation since we want to get noise data of
the device, on the other hand we do not want to apply any AC signal to the device so
we use this dummy node connected to a 1 Ohm resistor. Note that the value of the
resistor is not relevant.
The line: V1 1 11 0 connects a voltage generator between node 1 and 11 with zero
voltage. In the spice syntax this acts as a current probe. Note that we want the noise
contribution at the Collector as a current.
In order to get the noise part of the current through V1, we setup another
independent sub-circuit.
The line: H1 5 0 V1 1 represents a current-controlled-voltage-generator. Therefore
the voltage between node 5 and 0 is controlled by the current through V1, multiplied
by '1'.
The resistance called Rspice is connected between node 5 and 0. Note that the value
of this resistance is not relevant.
Finally, in the IC-CAP setup, we output the noise at node 5, which in turns represents
the noise of the current flowing in the collector.
The syntax for ADS:
define ads_noise (C B AUX)

; note: ADS circuit syntax is case sensitive!!!!!
R:RB1 B 21 R=10k Noise=no
C:CB1 21 0 C=100u
R:RB2 21 22 R=10k Noise=no
C:CB2 22 0 C=100u
R:RB3 22 23 R=1MEG Noise=no
;here comes the transistor
noise_1_f_bip:X1 11 23 0
;add a dummy resistor to sense the noise current (current-controlled voltage source in MDS cannot
be controlled by a V=0 source)
R:Rdummy C 11 R=1m Noise=no
;convert the Collector noise current into an identical voltage
cdvs:Haux 11 C AUX 0 M=1 ANG=0 T=0
R:Raux AUX 0 R=1k Noise=no
end ads_noise
;define a current dependent voltage source
758
#echo define cdvs(senssnk senssrc pos neg)
#echo parameters M=1 ANG=0 T=0
#echo gain=M*polar(1,-ANG)*exp(-j*omega*T)
#echo Z_Port:v senssnk senssrc pos neg Z[2,1]=gain
#echo end
Please note that you have to calculate the square of the ADS noise in order to get the
spectral density as calculated by SPICE. In the demo_features directory of IC-CAP, see the
file
6_SIMULATORS\0_general\noise_1_f_using_Spice_MDS_ADS.mdl
For the Gummel-Poon model, there is a special GP BJT model parameter 'rbnoi' ,
describing the effective base noise resistance. This parameter is available in Spectre, and
also available in ADS. Below are the governing (thermal) Rb noise equations from the
Spectre (4.4.6) and ADS (2002C) documentation showing identical implementation
(Default for ADS Rbnoi=Rb)
Spectre:
if RBNOI is specified:
otherwise:
ADS:
Rbnoi was added only to the Gummel-Poon model by Cadence, so it was added to ADS's
Gummel-Poon model too for compatibility. VBIC, Hicum and Mextram have more
sophisticated models for the base resistance, so this parameter wasn't added to these
other models by Cadence, Agilent or anyone else.
Acknowledgments and Publications

I would like to thank specially F. X. Sinnesbichler of the Technical University Munich,
Institute of High Frequency Techniques, for having laid the base for this manual. A big
part of the chapters stem from this work. I would also especially acknowledge
Mr.Sinnesbichler's IC-CAP workshop about Noise modeling in Munich, July 1998.
In the same way, I have to acknowledge Mr.Knoblinger of Infineon Technologies AG in

Munich, for his hint on using the special low-noise current amplifier as an excellent and
straight-forward way to measure 1/f noise accurately and without big influences from
power line frequency harmonics.
For detailed verification tests and measurements of the proposed measurement setup, I'd
like to mention A.Blaum, O.Pilloud, G.Scalea, J.Victory, Motorola, Geneva
For important discussions on the noise terminology, I'd like to mention Mr. Berkner, also
Infineon Technologies AG in Munich.
For detailed verification tests and measurements of the proposed measurement setup I
am grateful to A.Blaum, O.Pilloud, G.Scalea, J.Victory, Motorola, Geneva
Web Info
Stanford Research Web Page: http://www.thinksrs.com//
Publications
[1] R.Mller, Rauschen", Springer-Verlag, Berlin, 1990.
[2] B.Schiek, H.-J.Siweris, Rauschen in Hochfrequenzschaltungen", Hthig Buch Verlag,
Heidelberg, 1990.
[3] G. Massobrio, P. Antognietti, Semiconductor Device Modeling with SPICE", McGraw-
Hill, New York, 1993.
[4] F. N. Hooge, 1/f Noise is No Surface Effect", Physics Letters, Vol. 29A, Nr. 3, pp. 139-
194, April 1969.
[5] F. N. Hooge, 1/f Noise Sources", IEEE Transactions on Electron Devices, Vol. 41,
Nr.11, pp. 1926-1935, Nov. 1994.
[6] C. McAndrew at al, VBIC95: An Improved Vertical, IC Bipolar Transistor Model",
Proceedings of the 1995 BiCMOS Circuits and Technology Meeting, , Minneapolis, pp. 170-
177, 1995
[7] C. McAndrew et al, VBIC95, The Vertical Bipolar Inter-Company Model", IEEE Journal
of Solid-State Circuits, vol. 31, No. 10, pp. 1476-1483, October 1996.
[8] F. X. Sinnesbichler, G. R. Olbrich, VBIC - The Vertical Bipolar Inter-Company Model.
Ein berblick ber das Modell und ber zugehrige Parameterextraktionen", Hewlett-
Packard IC-CAP Workshop-Reihe 1997/98, Mnchen, Januar 1998.
[9] H. C. deGraaff, F. M. Klaassen, Compact Transistor Modelling for Circuit Design",
Springer-Verlag, Wien, 1990.
[10] Y. Cheng et al, BSIM3v3 Manual", University of California, Berkeley, 1996.
[11] C.C. McAndrew, Practical Modeling for Circuit Simulation", IEEE Journal of Solid-
State Circuits, vol. 33, no. 3, pp. 439-448, March 1998.
[12] A. Cappy, Noise Modeling and Measurement Techniques", IEEE Transactions on
Microwave Theory and Techniques, vol. 36, no. 1pp., 1-10, Jan. 1988.
[13] F. X. Sinnesbichler, M. Fischer, G. R. Olbrich, Accurate Extraction Method for 1/f-
Noise Parameters Used in Gummel-Poon Type Bipolar Junction Transistor Models", IEEE
MTT-S Symposium, Baltimore, 1998, pp. 1345-1348.
[14] J. C. Costa, D. Ngo, R. Jackson, N. Camilleri, J. Jaffee, Extraction of 1/f Noise
759
Coefficients for BJTs", IEEE Transactions on Electron Devices, vol. 40, no. 11, pp.1992-
1999, Nov. 1994.
[15] J.C.Vildeuil, M.Valenza, D.Rigaud, Universite de MontpellierII, France, CMOS 1/f
Noise Modeling and Extraction of BSIM3 Parameters Using a New Extraction Procedure
ICMTS1999 Conference Gothenburg, March 15-18,1999, ISBN 0-7803-5270-X
[16] C.G.Jakobson, I.Bloom, Y.Nemirovsky, 1/f Noise in CMOS Transistors for Analog
Applications, 19. Convention of Electrical and Electronics Engineers in Israel, 1996,
pp.557-560
[17] A.Blaum, O.Pilloud, G.Scalea, J.Victory, F.Sischka, A New Robust On-Wafer 1/f Noise
Measurement and Characterization System, ICMTS conference 2001,
Kobe, Japan, March 19-22,2001
The use of the SR560 low-noise voltage amplifier (lower system noise, but no internal DC
bias available) is published in:
Hardy et al., Low-Frequency Noise in Proton Damaged LDD MOSFET's,
IEEE Trans. on el. devices, vol.46, no.7, July 1999, P.1341
760
Thermal Modeling
Note: See also the chapter on Pulsed Measurements in the Measurement chapter.
When specifying the thermal model parameter RTH, make sure to set CTH to a non-zero value. Otherwise,
your device is 'thermally faster' than electrically i.e. if you simulate pulses, you do not get the expected
fast risetime of the electrical performance, followed by the slower thermal time constant.
If you are not sure about your CTH value, set it to a big value, e.g. several milli-Farad.
When simulating the temperature dependence of circuits, two global simulator variables
come into play:
TEMP: the temperature for the simulation result
TNOM: the temperature at which the device was measured and modeled.
This assumes that there is a constant simulation temperature, and no dependence of the
simulation result from self-heating effects.
This can cause a major problem to modeling, because as a general rule, we have to
account for self-heating effects at currents above ~10mA (for typical vmax = 5V). While
this is true when measuring packaged devices, for on-chip measurements using a
thermochuck, the critical current is about 50mA.
Therefore, for modeling, we can consider three cases:
Stay well below a few mA for all device characterization measurements

For currents above some mA, measure isothermally (pulsed), and avoid self-heating,
Measure both, pulsed and CW at the higher currents, change the device models and
enhance them for self-heating.
Modeling Procedure
For the last case, the modeling procedure is as follows:
1. Perform a full device measurement characterization at e.g. 25C in pulsed mode

2. Extract all temperature-independent model parameters (conventionally)
3. Perform another pulsed measurement at the other temperatures, e.g. -25C and
+75C
4. Extract now the temperature parameters (e.g. XTI and XTB for the Gummel-Poon
model)
5. Perform a conventional, non-pulsed measurement. The deviation from the pulsed
measurement of above is then due to self-heating, modeled by the extra thermal
network of the model (see fig. 1). This means, the fitting is obtained by tuning the
thermal network model parameters.
Electro Thermal Analogy
A simple thermal model is depicted above. A current, called i_dissip_power, is calculated

(inside the custom thermal model) from the dissipated power inside the model.
In the thermal model, this current feeds the thermal RC equivalent schematic. Here,
C_chip models the thermal memory of the inner, active part of the chip.
R_insidechip_to_surface models the thermal resistor from this inner chip region to its
surface. The thermal storage capabilities of the chuck is represented by capacitor
C_chuck. Finally, R_chuck models the thermal resistance of the chuck.
There are two voltages available with this simple thermal model: V_chip_overtemp models
the inner chip overtemperature, and V_chuck_overtemp represents the overtemperature
of the chuck.
Back to the custom thermal model of the DUT itself, it senses the voltage
V_chip_overtemp and adds it, interpreted as a temperature, to the actual value of the
parameter TNOM (or the model specific temperature parameters).
Static and Dynamic Thermal Modeling

For the thermal modeling, we have to distinguish between static and dynamic thermal
modeling. Static dynamic modeling means that the steady-state overtemperature was
reached for every individual measurement point. For the corresponding measurements,
this means that the DC analyzer for example is set to the absolutely slowest measurement
mode possible
In the dynamic case, the end temperature may not have been reached for every
measurement point, and also the actual temperature of the current point may be an
761
overlay of the self-heating right before and the current one. In this case, for a DC
measurement using an HP 414x or 415x instrument, we will measure the time required for
that measurement, calculate the corresponding time for each measurement step, and
replace later the pure DC simulation by a time domain simulation with the same number
of measurement points like the measurements, and each measurement point lasting as
long as in the measurement. This means, the kind of the simulation is a copy of the
measurement process.
Example: Static and Dynamic Thermal Modeling of the DC characteristic of a

Temperature Dependent Resistor
IC_CAP file: diode_selfheat_static.mdl

This example for modeling the static thermal DC characteristics of a diode is
after:
Patrick Halloran and Lu Ke, University College Dublin, Ireland, 'Using the
Symbolically Defined Device to Simulate Diode Self-Heating', published in the
Hewlett-Packard High-Frequency Horizons Journal June 1993.
This model consists of a diode DC model, enhanced for static thermal self-heating. By
varying the thermal resistor Rth, the dependence of the diode characteristic in the ohmic
range can be studied to also depend on self-heating, described by this parameter Rth. It
can be seen that for low Rth, i.e. self-heating, the trace in the ohmic region is below the
one without self-heating.
Self-heating reduces the measured diode DC currents and thus affects the measured
effective ohmic resistance of the diode.
The user-defined sdd circuit description is given below:
define SelfHeatingStaticDiode (A C)
is= 1e-17
n = 1
rs= 5
Pt=3
Eg=1.11
Tamb=300
Rth= 1k
k=BOLTZMANN
q=QELECTRON
Tjunct = (Tamb + deltaT)
exparg = (-q * Eg/(n*k*Tjunct) * (1.0 - Tjunct/Tamb))
isT = (is * (Tjunct/Tamb)**(Pt/n)*exp(exparg))
id = (isT*(exp(q*vd/(n*k*Tjunct))-1.0))
deltaT = (_v1 * _i1 * Rth)
vd=(_v1-(_i1)*rs)
sdd:selfheating_diode A C f[1,0]=(_i1-id)
end SelfHeatingDiode
Example: Static and Dynamic Thermal Modeling of the DC charcteristic of a
Temperature Dependent Resistor
IC_CAP file: resistor_selfheat_static_n_dynamic
762
763
ADS Netlist
;ADS simul input file for temperature-dependent resistor

define selfheating_resistor (1 2 CHIP_OVERTEMP)
;define some variables
temp=25
tnom=25
teffK=((temp+273.15)+_v3)
tnomK=(tnom+273.15)
R0=5
a=10m
;calc. temp.dependent resistor current.
; _v1 and _v2 are the potentials at the sdd ports 1 and 2
; HINT: the 3rd sdd port represents the CHIP_OVERTEMP
i_res= ((_v1 - _v2) / (R0 * (1 + a * teffK/tnomK)))
;calc. dissipated power
p_res= (i_resistor*_v1)
SDD:resistor 1 0 2 0 CHIP_OVERTEMP 0 I[1,0]=i_res I[2,0]=-i_res I[3,0]=-p_res
;-> this the thermal capacitor of the chip region around the device
C:C_chip CHIP_OVERTEMP 0 c=.01m
;-> this is the thermal resistor from inside the chip to its surface
R:R_inside_chip_to_surface CHIP_OVERTEMP CHUCK_OVERTEMP r=10
;-> this the thermal capacitor of the chip surface and the chuck
C:C_chuck CHUCK_OVERTEMP 0 c=1m
;-> this the thermal resistor between the chip surface and the chuck
R:R_chuck CHUCK_OVERTEMP 0 r=100
end sdd_selfheating_resistor
764
HINTS:
If device models exhibit a thermal node, they include usually a thermal resistor and a thermal capacitor.
The default values are usually:
RTH = 0
CTH = 0
When specifying RTH, make sure to set CTH to a non-zero value. Otherwise, your device is 'thermally
faster' than electrically! I.e. if you simulate pulses, you do not get the expected fast risetime of the
electrical performance, followed by the slower thermal time constant.
If you are not sure about your CTH value, set it to a big value, e.g. several ms.
If a model includes temperature effects, proceed as follows:
1. Perform a full device measurement characterization at e.g. 27C

2. Set TNOM=TEMP=27, and extract all temperature-independent model parameters
(conventionally).
For a diode, extract e.g. IS N, RS, CJO etc.
Then:
1. Perform another measurement at the other temperatures, e.g. -25C and +75C
2. Keep the previously extracted parameters fixed,
Keep TNOM=27 (this is the parameter extraction temperature),

Set TEMP ( the simulation 'prediction' temperature to the one of the measurement,
and extract/optimize now (only)the temperature parameters (e.g. XTI for the
classical SPICE diode model).
However,if XTI cannot fit the thermal behavior of your diode, you need to extract a full set
of model parameters at each measured temperature (TNOM=TEMP=meas.temp).
And you obtain N full parameter sets for N measurement temperatures
NOTE: set the temperature parameters to default, i.e. switch-off their effect.
For the diode, set XTI=0 i.e., you will end up with N individual IC-CAP .mps parameter
files (In the ModelEile, perform a 'Save Model Parameters' for each modeled temperature).
In the next step, we need to interpolate these N parameter values vs. the temperature i.e.
to find a fitting curve for each parameter and its N temperature-dependent values.
Procedure
From demo_features/DEPOTS.mdl, apply Transform RUN_ALL in
GUI_DEPOTS/PARAMS_TO_MDM.
This will convert the N .mps files into a single .mdm file, which has the temperature as the
Input (sweep) with N temperature values, and as many Outputs as you had modeling
parameters in your .mps files before.
Then, ImportCreate the .mdm file into a new Modelfile Setup, and display each model
parameter (Output) vs. the temperature (Input).
Write little PEL Programs for finding the fitting functions, e.g. Program 'IS_fit':
RETURN IS_0 + IS_1*TEMP + IS_2*TEMP^2
or any other fitting function.
NOTE: the Variables IS_0, IS_1 etc. will be SetupVariables.
Display the result of IS_fit in the same plot as the IS (from Output) vs. temperature.
Apply a PlotOptimizer to tune/optimize your fitting parameters (IS_0, IS_1 etc.).
Last not least, combine all your diode DC curves, CV curves etc. into a single, big .mdm
file (applying GUI_UTILITIES/COMBINE_MDMS from demo_features/DEPOTS.mdl), and
load these big files into a new ModelFile. Copy-Paste your Circuit from the individual-
temperature-modeling Modelfile into the new all-temperature Modelfile, replace the model
parameters by the previously identified temperature fitting equations either in the
ModelParameter Tab list of the Modelfile, or if your simulator (e.g. ADS, Spectre, Eldo)
permits into the Circuit Tab of the Modelfile.
Simulate.
Apply a final PlotOptimizer for fine-tuning of
IS_0, IS_1, IS_2 and N_0, N_1, N_2 etc. etc.
765
NOTE: A demo about scaled parameter modeling can be found under

demo_features_2008xxxx\1_BASIC_MDLG_EXAMPLES\24_Scaled_Parameters_Mdlg\scaled_parameterized_modeling.mdl
Although it is about geometry scaled modeling, the example given there can be applied to
temperature modeling as well.
Publications
R.Bouchakour, C.Lallement, T.Maurel: Thermal and Electrical Modeling and
Characterization of Bipolar and Power MOSFET Transistors With IC-CAP and Saber,
Hewlett-Packard Characterization Solutions Revue, Winter/Spring 1994, S.6-13
and also published in the Proceedings of the 1993 European IC-CAP User Meeting,
Esslingen, Germany
T.Maurel, R.Bouchakour, C.Lallement, Modeling and Simulation of the Electrothermal

Behavior of the Power Bipolar Transistor, Proceedings of the 1994 European IC-CAP User
Meeting, Colmar, France
M.Dunn, B.Schaefer, Link measurements to nonlinear bipolar device modeling, published

in Microwaves&RF, Feb 1996, reprint as HP publication number 5964-9170E
Patrick Halloran and Lu Ke, University College Dublin, Ireland, 'Using the Symbolically
Defined Device to Simulate Diode Self-Heating', published in the Hewlett-Packard High-
Frequency Horizons Journal June 1993
K.Lu, P.Halloran, T.J.Brazil, Simple Method to Simulate Diode Selfheating Using SPICE,
Electronics Letters, August 13, 1992, Vol.28, No.17
766
On Target Modeling
Contents
Using PCM Data for Device Modeling (iccapmhb)

Working of On Target Modeling (iccapmhb)
Tutorial to use On Target Modeling (iccapmhb)
767
Using PCM Data for Device Modeling
Introduction
The key feature to use PCM Data for Device Modeling is to import PCM Data from an
EXCEL sheet to IC-CAP as shown below
PCM Data Representation
Application Scenarios
Foundry
Production measures and publishes key parameters like (Vtlin, Vtsat, ...).
Applying the On-Target Tool, an existing model can be compared to these PCM data.
If required, the model is adjusted to changes in the process.
Process Development
PCM Data are available from TCAD Simulations. Maybe, additionally a few I-V Curves, with
a few data points.
The On-Target Tool allows to establish a first Model, without having Silicon available.
Modeling and Model Centering

A 'classic' modeling engineer extracts model parameters based on I-V curves.
Provided there are PCM data available, the model can be verified against the
statistical spread of the PCM data.
Instead of measuring several wafers with complete measurement sweeps.
Data to be used for Application
768
769
On Target Modeling working
Three fundamental steps are as:
1. Data Import
2. Data Organization
3. Model Validation
IC-CAP Structure
Control Module
IC-CAP .mdl file with GUI

Manages the Data Module .mdl file
Handles templates for PCM simulation, plots, reports, etc.
Data Module
Implemented as a standard IC-CAP .mdl file

770
Holds the data for the On-Target modeling tasks
Setups with meas. and simul. data
User specific extensions
The OnTarget Toolkit consists of a Control Module 'Target_Modeling', and a Data Module 'OTM_xxxx'.
The 'Target_Modeling' is the master, the 'ITM_xxx' the slave.
IC-CAP has defined a Simple .csv Data Format for storing
Device ID and name,

Instance Parameter data,
PCM data (e.g. ION, ILIN, etc.)
For each PCM value, the spreadsheet above may include a spread (Mean, Min and Max
values),and also different temperature values.
THE IC-CAP .csv DATA FORMAT IN A NUTSHELL:

The Header keywords 'TYPE', 'NAME', 'SUBSET' and 'TEMP' are a MUST in the .csv file.
Also, in the row 'TYPE', the keywords 'ID', 'IP' and 'PCM' are a MUST.(quote}
Any PCM data names are accepted.
NOTE:
If you don't have min and max values, neither mean values,
but only pure measured data, you can delete the columns 'min' and 'max' in the PCM columns
section.Enter your meas. data to the column 'mean'. Important: the keyword 'mean' in the row SUBSET
has to remain.
Further info from the IC-CAP Manual:

The first n rows of the table are the KEYWORD rows.
Keywords are:
TYPE with
ID=Identification
IP=Instance parameters
PCM=PCM parameters
NAME with
DeviceID
Name
followed by any number and combination of Instance Parameters (IP) and PCM
values (PCM)
SUBSET: specifies in detail the PCM values: Allowed names are mean, min, max.
These identifiers are used to add statistical data.
TEMP: Temperature of the PCM value
In the file, each column is separated by a comma (,). This enables you to import the file
into table calculation programs like MS Excel.
A further rule for the rows NAME and TEMP is that if a cell does not contain a name,
the value of the previous one is used. For example, T=25C is used as a temperature
for all PCM values as long as no other value is specified.
An empty cell inside the NAME row leads to an error message.
For example, if the cell following AD is empty, you will get the error message
Instance Parameter Name AD is defined twice. Reason: In this case, the name of the
previous cell is used and double definitions are not allowed.
An empty data value cell (PCM data) is assumed to have the value of zero.
Important, when you don't have a fully populated PCM data table.
771
IC-CAP .csv data format:
The 2.d and 3rd column is reserved for the DeviceID and the device Name.
Then indicated by the keyword 'IP', follows a unlimited number of columns,
describing the instance parameters (L, W, SA, etc)
Finally, indicated by the keyword 'PCM', follows a unlimited number of columns,
describing the measured PCM values.
Blank entries in the PCM values are supported.
When reading a row, the value stays identical until another value is read. Therefore, you can specify e.g.
ION or the temperature values in every column, or simply in the first.
Loading the .csv file into Target Modeling GUI
1. First, inspect the loaded PCM data
772
2. Group your PCM data (Derive Subset)
3. Optionally, convert your PCM data to pseudo-sweep data (Derive Device Data)
4. Optionally, add conventional I-V measurements
Till now the data are imported and organized.

Now you have to set up the simulations for the imported .csv data.
773
The PCM data have been imported before, with no detailed information about how they were measured.
E.g. for ION, we imported the meas. value, but for making simulations, we need to tell IC-CAP what
specific vd and vg values have to be applied in the simulations we are going to perform now.
Supported simulators and models

Support for ADS, Spectre and HSPICE
Direct support for BSIM3, BSIM4, PSP, BSIMSOI
User defined models can be added e.g. HV-MOS, macro-models etc.
* This GUI pops up when you set up a new .csv data set,as the very first window to
be filled out.
The entered info goes to
Simulation Templates
So far, in the Data tab below, we have discussed the
Available Outputs
Devices
Now: Simulation Templates
774
After the model and the simulation details have been defined, we need to specify how to
simulate e.g. ION, VTH etc.,i.e. the bias voltage values.
While the Simulator, and the model were already specified right at the beginning, even before importing
the .csv data file, it is now time to specify what *kind of simulation* need to be performed on the PCM
data. In the Data ModelFile, is basically a DUT TestCircuit for the ION, ILIN, IOFF, ISAT etc., and another
DUT TestCircuit for the VTxx and the GMxx data.
Right-click on 'Simulation Templates' to add the simulation circuits (DUT TestCircuits in the Data-
ModelFile).
Simulation bias conditions
The PCM data were meas. with specific stimulus conditions. The simulator needs to apply
them *too*.
So far:
Data are imported and organized

Corresponding simulation settings have been made
Now you will apply the PlotOptimizer to fine-tune, adjust, update etc. your Model
Parameters.
Applying PlotOptimizer
A tandem of manager and data base ModelFile.

A GUI provided by the manager ModelFile to perform all requested tasks.
775
776
Tutorial
How Target Modeling Works: Opening and Understanding the Example
In this tutorial we will cover:
How to start the Target Modeling Package

How to load the predefined example
Viewing the device table information
Viewing the available PCM data
Viewing the simulation templates
Viewing the subsets of devices for scaling diagrams
Defining I-V curves from PCM data
Handling the optional conventional I-V curves
Display diagrams
Using the PlotOptimizer to extract the model
Generating an HTML report
How to start the Target Modeling Tool
1. Under IC-CAP Examples, load the file: Target_Modeling.mdlfrom

Examples/model_files/target_modeling
2. Open the new dedicated OnTarget window by double-clicking on the
"Target_Modeling" icon.
Loading a Predefined Example
777
1. Click on the "Examples" icon to load a predefined example

2. Select the BSIM4_NMOS_example file and click "Open"
3. The Target Modeling GUI gets populated with data.
This example already includes

in Tab 'Data':
Device PCM data of all loaded devices

A subset of predefined scaling PCM data
I-V Setups obtained from PCM data
Conventional I-V Setups
in Tab 'Diagrams':
A set of predefined multiplots diagrams
in Tab 'Options':
ADS is selected as simulator
Navigating the User Interface: Tab 'Data'
Exploring the Example:Device Table
778
1. In the "PCM Data" tree, right click on "90nm" (this is the name of the set of loaded
devices) and select "Device Info"
2. This will show the "Device Table" for the selected PCM Data.
3. Device information was read from the PCM data Excel spreadsheet
Exploring the Example:Available PCM data
1. In the "PCM Data/90nm" tree, open the "Available Outputs" tree

2. This will show the available PCM data. This information was also read from the PCM
data Excel spreadsheet.
3. For each PCM value, the spreadsheet may include a spread (Min and Max values)
Exploring the Example:Simulation Templates for PCM
779
1. In the "PCM Data/90nm" tree node, open the "Simulation Templates" tree
2. This shows the two predefined templates for simulating the PCM data: single-point
Id_Vd and VTH.
3. Right click on one of the two templates and select "Configure Templates..." to open
the bias definitions for I or V simulations.
4. Each PCM data has its own bias definition to be used by the simulations later
Understanding the Example:Creating a subset of

devices
1. In the "All" devices tree node, right click and select "Derive Subset"
2. Name the subset and click "Next"
3. Select X-Axis and Temperature
4. Select Devices
5. Closing the dialog will create the subset of devices and automatically create scaling
diagrams
Understanding the Example:Viewing subset of

devices
1. In the "PCM Data/90nm" tree node, open the "L-Scale W=5u" and "Devices" sub-tree
2. This is a predefined, customer-selected scaling, as described in the slide before
3. Right click on "L-Scale W=5u" and select "Device Info...".
780
4. This will show the information for the devices which have been grouped into this
subset
5. You can select "Configure Subset..." to modify the device selection
Understanding the Example:Defining I-V Setups from

PCM data
1. In the "PCM Data/90nm" tree node, open the "Conductance" tree

2. "Conductance" is a predefined I-V diagram (id_vd)
3. Right click on "Conductance" and select "Configure Device Data...".
4. This opens a wizard that helps the user to select PCM data, configure an I-V curve
and select the devices to be used for simulation.
Understanding the Example:THE PLOTS
1. Select the Tab "Diagrams"

2. All the available plots are defined in the "Single Diagrams" tree (on the left).
3. Display single plots by double-clicking on their names.
4. Multiple plots diagrams are defined in the "Multiplots" tree (on the right).
5. Display MultiPlots by double-clicking on their names.
Understanding the Example:Using PlotOptimizer to

tune the parameters
1. Once a Plot is open, you can apply the IC-CAP PlotOptimizer

2. Select the parameters to be tuned/optimized
3. Run the Tuner or the Optimizer
4. Note that the diagrams show Min/Max spreads for each PCM data
Understanding the Example:Report Generation

781
1. Select the "Report" Tab

2. Single and Multiplots diagrams are available on the left side tree.
3. The report structure is shown on the right side.
4. User may add/delete/move diagrams in the report structure
5. Select "Reports/Configure" and hit "Generate Report" to automatically generate HTML
report.
Example of generated HTML documentation
Conclusions
On Target Modeling enables IC-CAP users to:
Adjust a simulation model which was extracted from one set of golden diesto data
representing the overall trend of the process.
Perform comparisons between different simulated models and silicon data.
Provide documentation of the improved model.
782
Direct Measurement Data Reduction for
Scalable Parameter Extraction
This is an updated slideset of the one presented at the Europ. IC-CAP User Meeting in Prague, 2003,
by Pietro Brenner, Infineon Munich, and Franz Sischka, Agilent-EEsof.
It now features the LSYNC sweep of IC-CAP, available since IC-CAP 2006B, Update3 (Sept.2007)
Also, a new IC-CAP demo has been developed. See

demo_features/1_BASIC_MDLG_EXAMPLES/24_Scaled_Parameters_Mdlg/scaled_parameterized_modeling.mdl
Introduction
What is the fastest way to get scalable model parameters for extensive transistor libraries
without performing a lot of single transistor measurements and a lot of parameter
extractions ?
Perform direct extraction of normalized parameters on normalized measurement data.
Provided you know:
transistor layout measures,

vertical transistor geometry and transistor configurations,
absolute, geometry independent parameters (e.g. NF, VAF, VAR )and geometrically
normalized parameters: e.g. is_a, is_p, cj_a, cj_p, re_a, rb_sh, rb_l
as well as the correct parameter scaling rules :
e.g. IS = is_a * AREA_E + is_p * PERIM_E,

CJ = cj_a * AREA_E + cj_p * PERIM_E,
tf0 = tf0_a * (1+ a1*PERIM_E/AREA_E)
You can calculate parameter sets for arbitrary device geometries applying parameter
calculation scripts (see e.g. Tradica /1/).
New Approach for Direct Extraction of Geometrically

Normalized Device Parameters
Basic Idea
The basic idea for Direct extraction of Geometrically Normalized Device Parameters is as
follows:
Measure DC- and C(V)-characteristics from a set of deviceswith different geometries,

e.g.
3-5 different emitter lengths LE and
4-5 different emitter widths WE.
Transform measurements of device X Cx(v) into normalized per-area CAREAx(v) and
per-perimeter CPERIMx(v) pseudo-measurements, e.g.
Perform multiple linear regression on equation (2) for each bias point vs. slope leads
to perimeter-normalized meas.data CPERIM(v) and the intercept points lead to area-
normalized meas.data CAREA(v).
Perform parameter extraction only on the normalized data plots CPERIM(v) and
CAREA(v).
Only the parameters of the per-area and per-perimeter normalized capacitance
pseudo-measurements have to be extracted.
Example: Diode DC Characteristics
1. Measure data of devices with different geometry as a function of area and perimeter.
783
2. In a second step, display the data as a function of area and perimeter, after a known
scaling function like e.g.
3. From the y-axis intersect, calculate iaarea(v) and from the slope, calculate
iaperimeter(v), for every individual 'v'.
4. Display the normalized per-area and per-perimeter measurements against the
original stimulus.
5. Extract the per-area (IS_area, N_area) and per-perimeter parameters (IS_perim,

N_perim)
784
NOTE: In the Plot 'Change of sweep order', the curves represent lines, provided that the prerequisite
equation is fulfilled. As can be seen, for low va, i.e. low ia values, the noise overlies the prerequisite, and
for high va values, i.e. the ohmic range of the diode current, the prerequisite is not met too.
Nevertheless, we can apply linfits to all traces of va, and later exclude those data in the parameter
extractions behind the Plots 'y_intersects' and 'slopes'. See the specified parameter-extraction-boxes in
these Plots.
NOTE:
If the traces represent lines with a positive slope, the meas. data include dependencies of both, area
and perimeter.
If the traces in Plot 'Change of sweep order' represent flat lines (slope=0), then there is no
dependency of perimeter for the meas. data.
If the traces represent lines with a negative slope, then the assumed (drawn) area and parameter
values are not valid for the produced device on the wafer.
The area and perimeter are typically both smaller on the real device than in the drawn masks.
If the per-area parameters (Plot y-intersects) are negative, then the parimeter effect dominates the
area effect, and thus, the prerequisite of splitting into area and perimeter is no longer valid.
For details on the area-perimeter scaling, see J.Berkner, Kompaktmodelle fuer Bipolartransistoren,
Expert-Verlag, Renningen, Germany, ISBN 3-8169-2085-3, Chapter 7.1.2
Example: Diode CV Characteristics
IC-CAP Implementation
An IC-CAP Model file example has been developed for the proposed scaled modeling
method,
for area and perimeter scaling,

DC and CV.
Other scaling functions can be implemented in a very similar way.
See
demo_features/1_BASIC_MDLG_EXAMPLES/24_Scaled_Parameters_Mdlg/scaled_parameterized_modeling.mdl
Results
Diode DC Modeling Result, all devices

Measured (red) and simulated (blue) diode DC characteristics for all different geometries
785
NPN Junction Capacitance Modeling Result, all devices

Measured (red) BE and BC junction capacitance's and simulated (blue) BE and BC
junction capacitances of all different geometries
Advantages of Direct Measurement Data Reduction
Advantages:
Best fit to measured data is implicitly given

Required measurement data processing is easy to implement
Single transistor parameter extraction on a few edge devices
Much lower effort and much faster parameter extraction
Requirements:
Scaling rules for device characteristics

Device characteristics must be measured for all devices at the same stimulus
conditions
Availability of enough device geometries and geometry variations
Summary
Transforming measured data from a lot of different device geometries directly into
per-area, per-perimeter and per-sheet normalized characteristics
Saves plenty of parameter extraction work
Gives automatically best fit to measurement data base
Provides inherent information on data distribution Statistics.
All of the required normalized parameters can be extracted with few efforts using
intelligent test structures.
Using the presented simple test structures in production measurement setups (PCM)
will lead to higher accuracy in determining important device parameters and will give
excellent parameter statistics.
References
1. M. Schroeter, "TRADICA An Integrated Modeling Tool Linking Process And Circuit

Design"Model Parameter Generation and Sizing of Integrated Bipolar Transistors.
Manual Version 5. Nov. 2002.
2. J.Berkner, Kompaktmodelle fuer Bipolartransistoren, Expert-Verlag, Renningen,
Germany, ISBN 3-8169-2085-3, Chapter 7.1.2
786
Statistics
Contents
Parametric Statistics (iccapmhb) On the Web

Non-Parametric Statistics (iccapmhb)
Articles
Knowledge Center
Technical Support
Training
Videos
787
Parametric Statistics
Contents
Measurement Based Statistics (iccapmhb)

Model Parameter Based Statistics (iccapmhb)
Conclusions (iccapmhb)
788
Measurement Related Statistics
A wafer, containing different dies with test structures on them, is tested. Each test die
contains different MOS transistor geometries (large, narrow(s), short(s), small(s)), but
also other test structures like sheet resistance or diodes. The wafer is loaded onto the
chuck of the prober, and the prober then steps from test die to test die. All pads of the die
are contacted together using a switching matrix. Selecting the specific connections, all
different test structures (modules) on the die can then be measured independently. The
following figure schematically represents the type of measurements used in this section:
Wafer Scheme of a die containing MOS test transistors, sheet resistors, and diodes modules
Measurement related statistics can be performed by two different methods: either building
the mean and sigma value of measurement data during the measurements itself (on the
fly), or to firstly perform the measurements, store the results in an IC-CAP .mdm file and
do the mean and sigma analysis based on the whole data of the .mdm file. Examples can
be found at the following location:
demo_features/9iccap4experts/7statistics/1meas_data_stats/stats_during_meas.mdl
demo_features/9iccap4experts/7statistics/1meas_data_stats/stats_on_mdm.mdl
Referring to the IC-CAP file stats_during_meas.mdl, the measurements are performed,

and, provided they were valid, the data is added to the existing mean and sigma data
array. The statistics plot then shows the actual measurement and the updated mean and
mean+-sigma plots. The following figure shows such a result, based on the measurement
of resistor modules on 60 dies.
Measurement-based 'on the fly' mean and sigma statistics
Since after every measurement, the user is asked if the measurement was valid (can of
course also be done by some macro), there are no outliers (bad measurement data)
included in this type of measurement statistics.
When all the test structures have been measured, a set of model parameters can be
extracted from the mean data trace, as well as the mean+-sigma measurement data
traces.
The other method of measurement based statistics, which is covered in the second IC-CAP
file mentioned above, is related to already performed and stored measurements. In this
case, IC-CAP has controlled the wafer prober, a switching matrix if necessary, and
performed the measurements of different transistors. During that procedure, all the data
have been stored in an IC-CAP .mdm file.
For an example about such an automated measurement procedure, see file

demo_features/9iccap4experts/6wafertest/prober_matrix_mdm.mdl
In order to visualize the acquired measurement data, the complete .mdm file is re-
imported into the setups of DUT mdm_data of our model file. This is depicted in fig.3
(top). For this example, module containinga a large MOS transistor was selected: Both, vG
and vB are swept, while vD is small (100mV). The 3rd order sweep is then the die
number.
789
Such type of MOS measurements is usually performed for the modeling of the threshold
voltage and the mobility parameters of the MOS model.
Setup to re-import all the measurement data for the large transistor of all dies
Visualization of the data including the outliers
In order to simplify the statistics, we will reduce the bias sweep to only 1st order. I.e. we
will re-import only measurement data belonging to a constant vB. In this way, we have
only one single bias sweep (vG), and the 2nd order sweep is simply the die number. Fig 3
shows the resulting plots.
Selecting a single Bulk voltage out of the measurement data
In the next step, the measurement outliers have to be marked, such that they are not
considered when calculating the mean and sigma curves. Back to the IC-CAP file, this is
done by the transform 'mark_outliers'. It asks the user to specify the die (transistor)
number of the outlier(s), and then enters a flag into the setup variable outliersi . In order
to identify the outlier die numbers, click with the middle mouse button on the specific
trace of fig.4.
IC-CAP returns a data index, consisting of firstly the total point index, followed by a
number, like e.g. 19/25. This number means: 19th first-order bias sweep (vG) of die
no.25. This no.25 is what has to be entered when running transform 'mark_outliers' in
order to set the flag.
After the transform 'mark_outliers' was run, we are now ready to calculate the mean and
sigma values of our measurement data array. This is done by the transform
'box_data_statistics', and finally 'Mean' and 'Sigma'. After these programs have been
executed, program 'find_golden_die' identifies the very die which is closest to the
calculated mean data. This is depicted with the plots of fig.5.
790
Calculated mean, mean+-sigma and identified 'golden die', related to measurements within the marked sub-range,
and with eliminated outlier die #53, based on the 'large' transistor measurement data
About The Difficulties To Identify The Golden Die

Using Measurement-Related Statistics
So far, the basically two different methods of measurement-related statistics have been
introduced. Let's now consider the task to identify the real 'golden die', containing those
transistor modules which should be used for model parameter extraction. Back to fig.5, it
can be shown that the identified large 'golden die' transistor number is heavily dependent
on the user-selected subrange of measured data. Referring to the data in the actual model
file setup 'large_idvg_vb0', we obtain the following identified die number depending on the
box limits defined by X_LOW and X_HIGH (defining the two Setup Variables and running
program RUN_ALL):
X_LOW X_HIGH die #

0.5 6 74
0.5 4 74
1 3 5
2 3 5
4 5 9
In other words: based on a single Bulk voltage bias of the transconductance curve and
based on different sub-ranges of measured data, we can either identify die #74, #5 or #9
to be the representative 'golden die' for later parameter extractions. If we would apply
this procedure also to the other Bulk bias voltages, we would get even more die numbers
which could be candidates for the mean 'large', i.e. the 'golden' large MOS transistor for
later model parameter extraction.
This confusing picture becomes even more weird, when taking also the other
measurement ranges of a specific transistor module into account. For this experiment, we
will consider the data of the short transistor, because it is with this device where the demo
file also has output characteristics data measured in addition to the so far discussed
transfer curves.
Before we start, we check all available plots of this short transistor for outliers. All
available plots, linear and logarithmic ones, of both measurement domains and for all bias
conditions (setups meas_data_stats_short/short_idvg_vbxx and short_idvd;vgxx) , are
inspected. This gives the following outlier die numbers:
25, 26, 27, 28, 29, 61, 71
Neglecting these outlier dies, we get the following table of results, when applying
measurement-related statistics to the data:
791
Setup Note X_LOW /* X_HIGH /* die # % error
short_idvg_vb0 vB=0 all data >0.5 /1 all data >0.5 /1 89 1.77
vB=-1.5 all data >0.5 /1 89 6.54
15 vB=-3 95 7.03
3 all data >0.5 /1
short_idvd_vg1 vG=1 all data all data 95 1.28

vG=2 all data all data 95 0.52
2 vG=3 all data all data 96 0.68
3 vG=4 all data all data 96 1.07
4
short_idvg_vb0 vB=0 1 3 119 0.15
vB=-1.5 1 3 118 0.34
15 vB=-3 1 3 119 0.32
3
short_idvd_vg1 vG=1 1 3 95 0.17
vG=2 1 3 95 0.05
2 vG=3 1 3 96 0.01
3 vG=4 1 3 89 0.02
4
1.5 2 4 119 0.05
15 3 2 4 119 0.12
3
vG=2 2 4 95 0.10
2 vG=3 2 4 96 0.01
3 vG=4 2 4 89 0.01
4
vB=-1.5 3 5 89 0.04
15 vB=-3 3 5 88 0.06
3
vG=2 3 5 119 0.16
2 vG=3 3 5 96 0.05
3 vG=4 3 5 89 0.02
4
Note:
/* to be consistent, instead of Setup Variables, DUT Variables are used
/1 to exclude measurement limitations (pA resolution), only vG >=0.5 is used
From this table, we have now the choice to select either die no. 16, 88, 89,95, 96 100,
118 or 119 which might be considered as the 'golden' short transistor module. If we would
then also take the large transistor modules, the narrows and the smalls into account, the
spread of choosable 'golden dies' would even be enlarged!
Remember: candidates for the 'golden' large transistor die were die numbers 5, 9, 74, and
these numbers do not show up with the candidate numbers of the short 'golden' die.
Therefore, we are now really hit by the question:

How can we solve this dilemma to identify the die which contains the typical, i.s. the
'golden' transistor modules LARGE_golden, SHORT_golden, NARROW_golden etc.?
The only way out of this 'golden die' selection dilemma is non-parametric statistics,
covered in the next topic.
792
Model Parameter Related Statistics
As we have seen in the preceding chapters, modeling of electronic components often
means to determine quite a number of model parameters. This is especially true for MOS
transistors and their spread of different geometries. As a consequence, the modeling
department is under a permanent time pressure related to measurements, to model
parameter extraction and, thus, to cost. Therefore, other approaches should be taken into
account to respond to these challenges.
This is possible, provided we have enough statistical data in order to relate the big
number of model parameters to a smaller number of dominant parameters. In case adding
even production process data to this type of analysis, these process data then can become
the dominant parameters. This means that, provided the process has not changed, and we
get new process data, we are ready to calculate the modeling parameters with a given
certainty of e.g. 90% out of these process data without a need for further modeling
measurements or parameter extractions.
This procedure is called 'model parametric statistics'. Additionally, this type of statistical
analysis can, if combined with process and device simulation results, also predict the
performance of transistors which are based on new designs, without manufacturing them
first.
Semiconductor modeling hierarchy from process simulation to model libraries
The hierarchy of simulation tools or statistical influences is illustrated in the above figure.
The lowest simulation (statistical analysis) hierarchy is the process simulation, which
models the different manufacturing steps in a semiconductor production. Results of these
process simulations are given in form of doping profiles and device geometries of a single
transistor. With this information, the device simulator calculates the electrical behavior of
this single transistor, when external voltages or currents are applied to it. The results are
typically given as a current vs. voltage. From these curves, the model parameters can be
extracted for simulation of the transistor in a circuit simulation program like SPICE. These
single transistors are the basis for further more complex circuits e.g. digital standard cells.
Let's come back to the model parameter related statistics. In order to find the dominant
parameters, be it process or model parameters, out of the big total number of
parameters, a big amount of extractions has to be performed first. As a general rule, for N
model parameters, there should be at least 10*N parameter sets been extracted, better
are N*N parameter sets. For a MOS transistor, this may be up to thousands of parameter
sets.
The most commonly used tools for model parameter statistics are factor or principal
components analysis. While basically describing the same effect, their difference lies in the
definition of the fluctuation contexts. Interesting enough, both methods have not been
developed for technical applications originally, but instead for social science. Their main
objective was to 'measure' non-measurable properties like intelligence etc. In such
statistical experiments, test persons are asked to fulfill some tests like solving
mathematical problems, puzzles or other tests. From the test results, it is attempted to
measure the non-measurable property 'intelligence'.
Principal Component Analysis

In order to discuss briefly the properties of the two main statistical methods, we will start
with some general infos about the principal components analysis (PCA). Beginning with N
correlated statistical variables (model parameters), this method calculates a user-defined
amount of m<n uncorrelated variables: the principal components. These new principal
components describe the statistical spread of the original n correlated variables. The
important fact is, however, that in practice, less than n principal components are required
to describe the original n correlated variables.
Note
These principal components should be considered as virtual variables, which do not necessarily represent
physical parameters, and which, therefore, cannot be measured physically.
Definition wise, the principal components can be considered as a linear function of the
original correlated variables X:
793
with:
PC non-measurable principal component, m<n

w weighting factors
X normalized measurable, possibly correlated statistical variable
(here: the n model parameters)
If the variables X have a wide range of values, they should be first normalized as follows:
with:
!mrps_equation2.1.gif!mean of the statistical variable

!mrps_equation2.2.gif!standard deviation of the statistical variable
x statistical variable
!mrps_equation2.3.gif!normalized statistical variable
After this normalization, all the variables have the mean value 0 and a standard deviation
of 1. This normalization is especially important for device model parameters, because their
numerical values can range from very low values (e.g. saturation current IS in the 10-20
range) to very big ones (doping concentration NCH in the 10 16 range).
In order to calculate the weighting factors w in (1), the correlation matrix of the
normalized statistical variables (model parameters) has to be calculated first. Out of this
correlation matrix, the eigenvector U and the eigenvalues can be derived. After that, the
principal components can be calculated by:
The following figure visualizes the effect of the principal components, related to modeling
parameters. Starting point were 50 MOS model parameters. Each of them, because
normalized, would explain 1/50 = 2% of the total variance of the MOS transistors. After
the principal component analysis has been performed, the first principal component can
explain 44.7 % of the variance of the original set of model parameters! The second one
explains another 23.6 %. Both together can explain already 68.3 %. And if we take the
first 5 principal components into account, over 92% of the MOS transistor variance can be
explained.
Comparing the variance of the original normalized 50 model parameters and the 8 obtained principal components>
This analysis result can also be interpreted that among the originally 50 model
parameters, there are 5 dominant model parameters. Therefore, the remaining last step is
to associate the dominant model parameters with the determined principal components.
Factor Analysis
The other method to reduce the number of model parameter is the factor analysis. While
the so far discussed principal components method assumes that the total variance of a
statistical variable contributes to the principal component, the factor analysis is based on
the assumption that the variance of a statistical variable is composed of two parts: one
part is correlated with the variances of the other variables, while the second part is related
only to the variable itself.
with:
794
F non-measurable common factor
e variable specific factor
X normalized observable statistical variable
v factor weighting
A little example should explain this idea:

In production, the weight and the diameter of a metal cylinder are measured. It is obvious
that these two parameters are correlated, because a bigger diameter means immediately
a bigger weight. However, each property is measured by another measurement
procedure, which, on the other hand, is not ideal and adds by itself some error to the
measurement result. This measurement specific error is not correlated with the other
measurement parameter.
Like with the principal component analysis, there are less factors F than observable
statistical variables X. And, again, some few dominant factors can cover the major part
(usually >90%) of the original model parameter fluctuation.
Practical Example
Referring to the modeling of MOS transistors from a single wafer or also from different
wafers from different production lots, we need to create a spreadsheet of extracted model
parameters first. It is very important to note that the spreadsheet should be based on
directly extracted model parameters and not more than required on parameter
optimization. As already stated in the previous chapter about measurement related
statistics, finding model parameters by optimization may add so-called optimizer noise to
the model parameter values. This is due to the fact that, depending on the starting
conditions before the optimization, the mathematical curve fitting between measured and
simulated data can be the same for different sets of model parameter values.
After the parameter spreadsheet is filled out, we have to check the validity of the data
first.
Note
best statistical results are obtained if the spread of all the model parameter values is Gaussian. Of course,
in reality, this is difficult to achieve, and therefore a pre-inspection of the data is very important.
First, we should try to get an overview about the trend of the extracted model parameters
versus the die number. Usually, this die number is related to certain manufacturing dates,
lot numbers etc. Although the total spread of the model parameter for all dies on all
wafers may be quite Gaussian, this graphical analysis can give us some insight about the
process fluctuations included in our statistical analysis. In the example on the left of fig.3,
it is obvious that our data is affected by changes in the production process. The parameter
TOX, given in that plot, indicates that the process (oxide thickness) has been adjusted at
least four times within the time slot related to our extracted model parameters. If,
however, we only had plotted the histogram of all the TOX data, this property would not
have become obvious.
Analyzing the distribution of the spreadsheet's model parameters versus production fluctuations
Note
an example for such a graphical analysis can be found under
demo_features/9iccap4experts/tstatistics/2param_stats/2visu_statistics/visu_stats.mdl and running macro
PARAM_DISTRIBUTION
In a second step, when applying histogram plots to the parameter data directly, we can
check the model parameter spread itself, independently of the manufacturing date. This is
done in the following figure:
Distribution of a model parameter (VTH) to check for outliers>
795
In the above figure, the outliers represent so-called spot-defects on the wafer: shorts due
to dust particles, too big leakage currents etc. The inner part of the distribution, marked
with 'accepted range', is what we want to include in our statistical analysis. Because this
range represents the effects, which we are interested in for our statistical analysis:
tolerances like mask positioning, implant effects, temperature effects during production,
sputtering energy etc. Therefore, the spreadsheet rows, containing outliers, have to be
deactivated before we apply statistical analysis.
This can be done by activating the actual parameter column and clicking Date/Date Filter.
An interesting aspect of model parameter analysis is to check for multi-modal parameter

distributions and how to avoid them. Fig. 5 gives an example, based on the shipped
example file examples/icstat/bsim3.sdf . Here, it becomes obvious that the data stems
from different test dies (5 dies on different wafer locations, the small stepping) and that
the wafers stem from different production lots (the big steps)
Bi-modal distribution of the model parameter TOX
Since TOX is always related to the MOS transistor's VTH, the bi-modal parameter
distribution suspected in fig.5 becomes also visible with the scatter plot of these two
parameters. This is depicted in fig.6.
Scatter plot of the bi-modal distribution of the MOS parameters TOX and VTH
Now, the suspicion of bi-modal parameter distribution has been proofed. Therefore, we
display the histogram plot of parameter TOX and select the upper sub-range for further
statistical analysis. The new histogram of this sub-range is given in the upper right side of
fig.7. It represents a much nicer Gaussian distribution than the original data set.
Eliminating bi-modal parameter distributions using the histogram plot
In a next step, we have to check the distributions of the other model parameters of the
selected sub-range. In this step, the main objective is to non-linearly transform the model
parameter values in order to obtain histogram plots with as similar as possible Gaussian
distributions before we can start with the very statistical analysis.
796
Note
for the non-parametric analysis of the previous chapter (measurement related statistics), the Gaussian
distribution of the model parameters was not required.
MOS parameter PDIBL is a candidate for such a check. As can be seen in fig.8, the
distribution of this parameter looks still quite confusing (although we have already
deactivated the bi-modal TOX-VTH distribution). After a logarithmic conversion of the
parameter values, however, we obtain a very nice Gaussian distribution.
Converting the model parameter non-linearly in order to achieve a distribution as close as possible to Gaussian
After all, the model parameters have been checked and possibly transformed for Gaussian
distribution, so that a final outlier elimination of the model parameters can be performed.
From the 'Statistical Summary' window of the statistics package, we obtain the mean and
standard deviation of the (transformed) model parameters, and, using a small macro in
IC-CAP, which can access the rows and columns of the statistics parameter spread sheet,
we can deactivate all those dies whose parameter values are outside the mean?sigma (or
similar) limits.
Then, we are finally ready to go for the model parameter related statistics, and, after that,
for the parametric statistical analysis. We click the menu pick 'Analysis', and, as can be
seen in fig. 9, we simply follow the procedures top-down:
Steps for the model-parameter related statistics
Description of the Steps

Correlation Analysis:
This step returns the model parameter correlation matrix. The correlation analysis has to
be performed first, giving the relationship among the model parameters. Fig.10 shows the
table and also a schematic visualization for easier interpretation. All the correlation factors
are printed and all the parameter relationships are plotted against each other. If there is a
circle-like 'cloud', then there is no relationship among the two parameters. If the 'cloud'
goes upwards, then the dependency is a linear one with positive slope, and if it tends
downwards, the dependency goes with a negative slope. If we have several 'clouds' within
a specific sub-plot of fig.10, this would hint to another bi-modal parameter distribution
which should be eliminated before further proceeding with the statistical analysis.
Fig.10: correlation matrix and a schematized scatter plot matrix of the parameters.
797
Factor Analysis:
This step allows you to:
Select the number of factors and the type of statistical analysis method (e.g. factor
analysis, principal component analysis etc.).
Returns the Factor Loadings spreadsheet containing the factors, their relationship to
the parameters and another spreadsheet containing the factor variances and, last not
least, the cumulative factor variances.
Then, applying factor analysis (step 2), we select the number of factors (here 8) and get
the factor loading matrix, see fig.11. First of all, we should inspect the distribution of the
factors. It is important to have, based on our 50 model parameters, not more than ~10
dominant factors which can explain >90% of the original model parameter variations, and
also that the first factor is big compared to the second and so on. Topic 1 is ok, because,
at the lower right in fig. 11, the cumulative percentage of the selected 8 factors is 92.7%.
A further inspection of the distribution of the factors, depicted in fig.12, also indicates that
the basis of our statistical analysis, the quality of the model parameters, was quite good.
This can be seen from the fact, that the first factor is represented by a big variance value,
followed by rapidly decreasing values of the next factors. This is an indication that there
are real dominant model parameters present in our parameter spreadsheet.
(see IC-CAP demo file

demo_features/9iccap4experts/tstatistics/2param_stats/2visu_statistics/visu_stats.mdl
and running macro CUMUL_PARAM_VARIANCE)
Another important check at this point is the data in the upper table of fig.11. Here, we can
identify the association of the factors with the original model parameters. We should
check, if the factors, selected by the mathematical algorithm, are the wanted dominant
model or process parameters or not. The user can change these associations in order to
help with achieving this goal.
Hint: it has been shown that, when applying optimization for model parameter extraction
rather than the recommended direct model parameter extraction, the automatically
selected factor associations would link the factors to 2nd order model parameters and not
to the desired 1st order parameters!
Factor loading matrix after having applied principal components analysis with 10 factors
Cumulative variance of the calculated factors: the first factor alone can explain
already ~35% of the model parameter variance, while the first three factors together
can explain more than 65%, and the 8 factors altogether explain more than 92%
of the variance.
Generate Equations
submenu /Factors: returns the Equation spreadsheet, containing the context between the
798
parameters and the normalized factors
submenu /Dominant Parameters: calculates the set of the new, dependent model
parameters based on the independent model parameters.
With this step, we have nearly reached our goal: to calculate the new model parameters
out of the factors, so that they have values and a spread as similar as possible to the
original, really extracted ones. First, after selecting the submenu /Factors, we obtain a list
of the new model parameters as a function of the factors. Fig.13 shows that list.
Finding the context between the determined factors and the model parameters.
Calculating the dependent model parameter based on the independent parameters
identified by the analysis depicted in the upper part of fig.11.

Then, after selecting the submenu Dominant Parameters, we get the final goal of our
attempts: the list of the dependent model parameters and their linear context with the
independent ones. The independent ones were identified with the factor loadings matrix in
fig.11.
In other words, this last spreadsheet of our statistical analysis closes the loop between the
model parameter <> factor context of fig. 13, and the factor <> dominant model
parameter context of fig.11.
This ends the description of how to identify the dependent model parameters and of how
to relate them to the independent ones.
But, we can also continue with our statistical experiment and apply methods, which can
produce artificial complete model parameter sets with basically the same properties like
the underlying old ones. Such a method could be used for design centering analysis.
Based on the previous statistical results, we can now also perform a parametric analysis
(step 4 in fig.9). Parametric means, that we can use the results of our statistical analysis,
and apply parametric techniques to predict sets of model parameters, which behave like
the underlying original model parameter sets.
Parametric Analysis
submenu /Factor Equations: Asks for selecting the parametric analysis method, e.g. Monte
Carlo Analysis, and for the
number of outcomes (model parameter sets).
Returns a new spreadsheet containing the calculated parameters, based on the statistical
factors from above.
799
New set of model parameters generated by the factors of the principal component analysis
First, we specify which statistical method we want to apply to generate a spreadsheet of

new model parameters (which are based on the determined dominant factors). Choosing
Monte Carlo analysis for example, (which alters the dominant factors F1..Fn between -1
and +1 in fig.13), we get the new model parameter sets as shown in fig.15. These sets of
parameters have basically the same statistical properties like the original sets of model
parameter, i.e. the same mean and standard deviation, however, is has been calculated
out of the dominant factors/parameters.
Parameter Based Statistics: Wrap-Up

This means that applying parameter related statistical analysis methods (factor analysis,
principal component analysis) to device modeling, we are able to reduce the number of
model parameters to the dominant ones. This works well, provided the spread of the
underlying fabrication process has been described properly by eliminating outliers,
eliminating multi-modal data fluctuations, and transforming model parameters to
Gaussian distribution. Furthermore, we get a deeper insight into the process fluctuations
and its stability.
Also, by applying parametric statistical methods like Monte Carlo Analysis to these
techniques, we can also predict the performance of new, not yet produced components,
based on the knowledge of the dominant factors of our production process and the
expected new values of the dominant process parameters.
Literature
/1/: T. Gneiting, Handout of the Workshop on Statistical Modeling, Nov.4 1997,
HP Office Taufkirchen/Munich, Germany (in German)
/2/: IC-CAP 5.0, Manual Statistical Analysis, June 1997, HP Part no. 85190-90050
/3/: T. Gneiting, I. P. Jalowiecki, "The Prediction of Circuit Performance Variations for

Deep Submicron CMOS Processes", IEEE International Symposium on Defect and Fault
Tolerance in VLSI Systems, Boston/MA, November 1996
/4/: T. Gneiting, I. P. Jalowiecki, "The Influence of Process Parameter Variations on the

Signal Distribution Behaviour of WSI Devices", IEEE Transactions on Components,
Packaging and Manufacturing Technology, Part B, Vol. 18, No. 3, August 1995
/5/: Chow, "Statistical Circuit Simulation of a Wideband Amplifier: A Case Study in Design
for Manufacturability", Hewlett-Packard Journal, October 1990
/6/: J. Power, D. Barry, A. Mathewsen, W. Lane; "Worst Case Simulation Using Principal
Component Analysis Techniques: An Investigation", Microelectronic Engineering 15-1991,
pp 213-216, Elsevier
800
Conclusions
After the two basic statistical principles, measurement related and model parameter
related statistics have been introduced, the question still remains, how to combine them in
order to save time without sacrificing accuracy.
A possibility might be to combine the non-parametric statistics, however applied to

process data (Process Control Measurement data) instead of measurement data, with the
model parameter related, factor analysis based statistics. In this case, the PCM data from
production, where the required big amount of data is acquired anyhow, are written to an
ASCII file and then loaded into the statistics spreadsheet. Then, applying non-parametric
statistics to these PCM data, we can identify the 'golden die' and the boundary dies (from
the dies containing the PCM modules). Provided these dies contain, besides the PCM test
modules, also the required transistor modules (see figure below), it is exactly those
transistors which are then measured and modeled conventionally. The obtained model
parameter sets are then exported one after the other to a new statistics spreadsheet, and
a factor analysis based statistics can be applied. This method would automatically exclude
outliers, and generate reliable model parameter sets, so that model parameter filtering
should be nearly not necessary.
Measurements of the test modules for PCM and Modeling
Applying the non-parametric statistics to PCM data can also be used to identify those few
dies which should be used for model parameter extraction, if the absolute minimal number
of measurements is requested in order to get just the 'golden model parameter set' and a
few boundary model parameter sets.
Publications
David McFeely, Dung Pham: Generating Statistical Models in IC-CAP,
Hewlett-Packard Circuit Modeling Seminar, 1992.
Farshid Iravani, Masa Habu, Ebrahim Khalily: Statistical Modeling Tools, Methods and
Applications for Integrated Circuit Manufacturability, Hewlett-Packard Circuit Modeling
Seminar, 1994.
Dan Stoneking: Statistical Circuit Design and IC-CAP's Non-Parametric Boundary Analysis,
Proceedings of the 4th European IC-CAP User Meeting, Berlin/Germany, Oct.1.-2., 1997
J.A. Power et al: Relating Statistical MOSFET Model Parameter Variabilities to IC

Manufacturing Process Fluctuations Enabling Realistic Worst Case Design, IEEE
Transactions on Semiconductor Manufacturing, August 1994.
J. Edward Jackson: A Users Guide to Principal Components, Wiley, 1991.
J.C. Zhang and M.A. Styblinski: Yield and Variability Optimization of Integrated Circuits,
Kluwer Academic Publishers.
Sam Kachigan: Multivariate Statistical Analysis: A Conceptual Introduction, Radius Press
A very detailed paper on principal components analysis and correlation rotation is:
K.McKarthy, A.Mathewson, Extraction of Statistically Valid Parameters for IC Design,
IEEE International Conference on Microelectronic Test Structures ICMTS 1999 Tutorial
Short Course, March 15-18,1999, Gothenburg, Sweden.
801
Non-Parametric Statistics
Contents
Introduction to Non-Parametric Statistics

Identify the 'golden device' and the 'limits devices'
Non-Parametric Boundary Analysis
Improving The Manufacturability Of Electronic Designs
802
Introduction to Non-Parametric
Statistics
Non-parametric statistics is a method which identifies in the multi-dimensional model
parameter room that die which is closest to the center of that room: the 'golden die'.
Therefore, this method takes into account all measurement setups and overcomes the
'golden die' identification problems described in the above chapter. Of course, with this
method, it is required that the model parameters for all the measured setups have been
determined accurately.
Let us refer to IC-CAP file

demo_features/9iccap4experts/7statistics/meas_vs_param_stats.mdl
Again, we try to simplify the things as much as possible. Therefore, for this experiment,
we choose a diode for the statistical analysis. Only two types of measurements, DC and
CV, are used, over a total of 98 dies. Outliers have been eliminated from the measured
data, before they were stored to an IC-CAP .mdm file. Fig. 6 depicts the results:
DC characteristics of 98 diode modules
CV characteristics of 98 diode modules
803
Note
Back to the IC-CAP file, setup 'dc_all' and 'cv_all' contain the re-imported .mdm data, and in both setups,
transforms 'meas_data_statistics', 'Mean' and 'Sigma', are used to perform the measurement-related
statistics. Transform 'find_golden_die' again identifies the die which is closest to the calculated mean data.
Like with the example about MOS transistors, we have the same dilemma: for DC, it is the
diode module on die 59 which is closest to the mean of all DC data (0.45% error), while
for CV, it is rather the diode module on die 5 with an error of 0.03%.
As stated above, the goal of this section is to export the model parameters to the
statistics spreadsheet. Before we do that, the measurement data distribution can be
analyzed. This can be done with transform 'check_distrib'. This transform cuts through the
die distribution for a given diode bias voltage and displays the result. Fig. 7 gives an
example for V~0.8V.
Checking the measurement data distribution for a given DC bias voltage
After this final check has been performed, we are now ready to extract the model
parameter for both, DC and CV, and to export them into the statistics spreadsheet. This is
performed by macro 'mdm2stats'.
It should be mentioned that it is especially important to use as much as possible 'direct

extraction' instead of parameter optimization. 'Direct extraction', where the model
parameters are calculated very accurately and directly from the measured data, produces
always a clear and direct relationship between the measured data and the model
parameters. Optimization, however, can add so-called 'optimization noise' to the model
parameters. This effect comes from the fact that different starting conditions for the
parameter values will lead to possibly the same fitting quality, however with slightly
different final parameters. And this slight variations in the final model parameters can
affect statistical analysis very badly. Modeling people call this effect 'optimizer noise'.
Therefore, again, it is important to use rather the 'direct parameter extraction' method.
(This is done with the extractions of the actual model file).
After the macro has been executed, the model parameter spreadsheet is filled up, see
fig.8.
Note
a complete IC-CAP .sdf file of the extracted DC and CV diode parameters can be loaded into the statistics
package from
demo_features/9iccap4experts/7statistics/meas_vs_param_stats_full.sdf
804
Filled statistics parameter spreadsheet
For the scatter plot analysis depicted in fig.9, the parameter columns Diode.IS and
Diode.N are already marked.
Scatter plot of the parameters Diode.IS and Diode.N
As shown in figures 9 and 10, as well as in table 1, different data analysis methods can be
applied to the extracted model parameter sets of the 98 dies, in order to check for
Gaussian distribution or correlation effects.
histogram plots of the distribution of model parameters Diode.IS and Diode.N
RS Diode.IS Diode.N Diode.CJO Diode.M Diode.VJ

Mean 5.17326 1.01001e-17 0.98782 6.31167e-13 0.414133 0.912216
Variance 0.04052 1.95562e-40 2.14889e-05 1.10783e-29 1.21248e-06 .60971e-05
Std.Deviation 0.20129 1.39843e-20 0.00464 3.32840e-15 0.001101 0.004012
Skewness 4.47606 -0.28516 0.69743 -0.016377 0.238159 0.238159
Kurtosis 24.9556 -0.10185 0.47870 -0.801842 2.291947 13.49792
Minimum 5.08958 1.0065e-17 0.98782 6.24985e-13 0.410591 0.8975
Maximum 6.58694 1.0132e-17 1.00132 6.37994e-13 0.418707 0.93625
Median 5.09256 1.0101e-17 0.98782 6.30994e-13 0.413798 0.9125
Med.abs.dev. 0.00213 8.7e-21 0.00248 2.003e-15 0.000983 0.00125
Rob.Est.Scale 0.00316 1.2898e-20 0.00367 2.9696e-15 0.001457 0.001853
With the use of 'non-parametric statistical analysis', we are now ready to identify the real
existing die which contains the 'golden module' of our diodes. As shown in fig.11, we
specify 10 boundary points (diode modules) which can be considered to enclose 80% of
the parameter samples.
805
Non-parametric statistical analysis to identify the golden model parameter set nd, thus,
the golden diode module, as well as 10 boundary parameter sets, which will include 80%
of the 98 parameter sets.
From this analysis, see fig. 12, diode module no.53 is identified to represent best the
'golden device' for both, measured DC and CV data. Modules #6, 13, 14, 36 etc. represent
the boundary chips on the wafer.
Result of the non-parametric statistical analysis
After having executed the non-parametric statistical analysis, we can re-execute a scatter
plot for the most important DC and CV model parameter in order to check the validity of
the identified 'golden' and 'boundary' modules. The two plots of fig. 13 show the result of
such a check. It can be seen clearly, that the identified die #53 represents that diode
module whose DC and CV data represent best the 'golden' module in question. Also, the
boundary modules look very reasonable, again covering both domains, the DC and CV
data.
DC parameters IS vs. N
CV parameters VJ vs. CJO
806
CV parameter M vs. CJO
Using macro 'stats2mdm' of our IC-CAP model file, we can export the model parameters
of our identified diode modules back into IC-CAP, run a simulation and display this
simulation result with the measurement distribution. As can be seen in fig. 14, the golden
no. 53 fits nearly ideally the mean CV curve, but is a little more aside the mean DC trace
(see the
IC-CAP file plots for details). However, it is the module which represents best both, the DC
and CV domain.
Identified 'golden diode module' no.53 and its location in the measurement data.
Note
if we had calculated simply the mean data of the DC curves and the mean data of the CV curves, and if we
had then extracted, based on these two mean data traces, the 'golden model parameters', this set of
parameters would correspond to no real existing diode. (Because the mean trace of the DC and the mean
trace of the CV would not correspond to a specific existing module).
And if we had done so for our MOS transistor example, with the many different transistor geometries and
measurement ranges, the result would have been even more non-physical and, therefore, even more
doubtful.
Back to our starting example with the MOS transistors, a similar analysis can be
performed. The result of a non-parametric statistical analysis (with the mentioned outliers
taken into account) is die no.19 with its different transistor geometry modules, which is
closest to the mean data of all dies, all modules and measurement ranges.
Besides this non-parametric type of statistical analysis, the IC-CAP statistics package
covers also the 'classical' factor analysis and principal component analysis method, as well
as the ability to generate Monte Carlo based sets of model parameters, with a distribution
corresponding to the original model parameter distribution. More details about these
methods can be found in the following chapter about parameter-related statistics.
Applying Non-Parametric Statistics Directly To

Measured Data
Besides referring to extracted model parameters, non-parametric statistics can also be
applied to measured data directly. In this case, a representative selection of measurement
data points is gathered during the measurement of the whole wafer in an ASCII file and
this file is imported into the statistics tool. Representative means that some characteristic
measurement points, e.g. 10-20 points of an output characteristics, are exported to that
ASCII file and not all the data. Especially noisy data have to be ignored for this type of
analysis.
After this ASCII file is imported into IC-CAP's statistics, it may look like the following
table:
chip# gummel1 ... gummelN output1 ... outputM cv1 ...

1 1.3u 24u 144u 1.3m 1.5m 1.6m 1.2p 1.5p
2 1.4u 25u 146u 1.4m 1.6m 1.7m 1.1p 1.4p
3 1.2u 24u 145u 1.3m 1.5m 1.7m 1.2p 1.6p
4 1.3u 23u 144u 1.5m 1.7m 1.8m 1.3p 1.5p
Applying a non-parametric analysis to this data will again identify the chip which is closest
to the mean of all data of all measurement setups. Compared to the measurement-based
807
type of statistics of the first chapter, this method includes all kinds of measurement
setups and is, therefore, identifying the real 'golden die' for all types of measurements
and not simply that one related to a specific measurement setup.
Literature about non-parametric statistics

IC-CAP 5.0, Statistical Analysis Manual, June 1997, HP Part no. 85190-90050
808
Identify the 'golden device' and the
'limits devices'
In this section you will learn:
Conventional Statistics Methods (e.g. Factor Analysis)

Introducing The 'Non-Parametric Statistics' Concept
How does 'Non-Parametric Statistics' work?
An Example
Conclusions
Conventional Statistics Methods

Example: Factor Analysis
measure many devices

extract model parameters
add process parameters (PCM) to statistics spreadsheet
transform each parameter distribution to Gaussian
apply statistics
find dominant parameters (= parametric statistics)
apply e.g. Monte Carlo Analysis to generate new parameters which represent the
original distribution
What are Corners, Boundaries, and Monte Carlo

statistical models, obtained with conventional
statistics?
This is a simple geometric representation of what corners, boundaries, and Monte Carlo
are. In 2 dimensions, we select 4 evenly spaced models which lie on an N sigma boundary
of the joint PDF (probability density function). Corners are outside the joint PDF (in
purple), which produces too pessimistic results. Parametric boundary models (in red) are
right on the boundary of joint PDF so it provides a more realistic worst-case model. Worst-
case modeling includes both corner and boundary methods.
Monte Carlo (in black) is a sampling from the complete distribution that leads to a yield
estimate when each sample is simulated in the final circuit and the circuits performance is
checked against the pertinent performance specifications,
Introducing The Non-Parametric Statistics Concept

This type of statistics analyzes the hyper-cube of parameters:
Identifies the chip closest to the parameter center

Chips closest to an e.g. 90% enclosure hyper space
Therefore, it can easily be applied to model parameters and also to measurement data
What means typical component and limits

components identification ?
limits device: the user specifies the enclosure percentage for the data of the hyper space.
This kind of statistics is called 'Non-Parametric'
809
Non-Parametric Statistics Works on Gaussian
Distribution
Non-Parametric Statistics Works on Non-Gaussian

Distributions
Conventional Statistics vs. Non-Parametric

Statisticswrap-up
Mode 2 may lead to better congruence between boundary model and the boundary
devices in terms of final electrical behavior
Mode 2 is better because we do not expend considerable effort extracting many models
that we ultimately may throw away. If we are going to do direct or generated Monte Carlo
yield estimation then the effort is not in fact wasted.
Mode 2 is also better because one get two outputs nominal and worst-case vectors for
both the measurements and the models.
How does 'Non-Parametric' Statistics work?

First of all: Identify outliers and eliminate them
810
Then:Select some spot data from DC, CV and Spar and export them
to the statistics spreadsheet
Then, the Data Sets are Normalized

e.g., for a diode, id(at vd): is in the mA range C(at vd): is in the fF range Spar at freq and
vd: are -1 ... 1 in MAG, 0 ... 360' in PHASE
-> Normalize each data set to A Unit Hypercube
Estimating Density Using Nearest Neighbors Method

Number of NN = 7
The user needs to understand the trade off between setting Density Estimator Percentage,
which directly controls the number of nearest neighbors, either too high or too low. If set
811
too high, density estimate is overly smooth and the algorithm can fail to resolve peaks in
the PDF that are close together. Setting it too low leads to a different problem. The
problem is that the distance averaging is a Monte Carlo process and consequently
introduces Monte Carlo noise into the density estimate. This noise can lead to poorly
chosen nominal and boundary models. The problem is more acute for the nominal model,
as the nominal model is in a pronounced region of low slope in the density estimate.
Spatial Diversity
Spatial diversity is done by oversampling and then iteratively forming all possible pairs
and discarding a point from the pair that is closest together. The procedure terminates
when only the desired number of boundary points remain.
Do not know best limits for Spatial Diversity Oversampling factor. At the present 1.5 to
3.0 seems like a good range. Clearly, it cannot be 1 or lower since 1 implies no
oversampling. Runs with it set above 3 show no clear advantage. High settings do have a
significant performance penalty for large numbers of requested boundary points.
Nominal and Boundary Point Selection
The Result
812
Note
The 'golden' device is NOT the mean(DC), mean(CV), mean(Spar) and the 'boundary' devices are NOT the
sigma(DC), sigma(CV), sigma(Spar). The 'golden' device is rather the device which is the most
centered to the 'cloud' of dataof all types of measurements and the 'boundary' devices are thosewhich
include a given data variation for all types of measurements.
An Example: Diode with measurements from DC to S-

parameters
measure and save data
identify outliers, export to stats
model individual die
measurement results (here full curves, but spot data wouldbe sufficient), outlier
identification, export of spot data to statistics spreadsheet
Note
Only performing spot datameasurementswould be sufficient as well.
813
Quick-extraction parameters, PCM data and/or spot data fromoperating points are
imported into the statistics spreadsheet
The non-parametric conditions are set and the statistics is applied
The result in table form:
The result graphically:
814
Now, that the 'golden' and boundary devices are identified on the wafer, we measure
them completely, and model them (and only them)
Modeling Results
The model parameters obtained will represent the centereddevice and an e.g. 80%
(boundary) yield parameter setrelated to the original spot measurement data distribution,
depending on the selected enclosure rate.
A simple method to identifythe 'golden device' and

the 'limits devices'has been proposed.
It has been implemented in IC-CAP.
This method reduces the amountof measurements and time considerably.
815
In this section, you will learn about:
Motivation
Circuit and System Design Methods
The Non-Parametric Boundary Analysis Algorithm
Validating the Algorithm
Applying the Tool
Conclusion
Motivation
Fabrication processes inherently stochastic
Boost yields and reduce design cycle times
Control process tolerances
Accurate nominal models
Assess manufacturability before production
Note
Assess manufacturability via yield estimation and worst-case model analysis.
Circuit and System Design Methods

Non-statistical design:
Traditional methods: Cut and try. Use hopefully nominal models. No assessment of
manufacturability prior to production.
Sensitivity based analysis: Add sensitivity analysis to traditional flow. Better but no
direct assessment of manufacturability.
Statistical design
Yield optimization: Must be able to estimate yield and need a global optimization
method. Best of the statistical design techniques.
Response surface methods: Probe circuit behavior with design of experiments so as
to build a low order polynomial model of the circuit's performance curves. No direct
information about yield.
Yield estimation: Direct or generated Monte Carlo. Ensemble directly or (PDF fit and
sample)
Worst-case analyses
Corner
Parametric boundary
Non-parametric boundary
What are Corners, Boundaries, and Monte Carlo?

This is a simple geometric representation of what corners, boundaries, and Monte Carlo
are. In 2 dimensions, we select 4 evenly spaced models which lie on an N sigma boundary
of the joint PDF (probability density function). Corners are outside the joint PDF (in
purple), which produces too pessimistic results. Parametric boundary models (in red) are
right on the boundary of joint PDF so it provides a more realistic worst-case model. Worst-
case modeling includes both corner and boundary methods.
Monte Carlo (in black) is a sampling from the complete distribution that leads to a yield
estimate when each sample is simulated in the final circuit and the circuits performance is
checked against the pertinent performance specifications,

A statistically robust method for obtaining nominal and worst-case vectors from a
collection of vectors drawn from an arbitrary multidimensional stochastic process
Works on Gaussian PDFs
And on Non-Gaussian PDFs !
816
Primary Inputs to Non-Parametric Boundary Analysis
Real valued dataset of S sample vectors with each vector having P parameters
Number or Boundary Points
desired number of worst-case models
Percent Enclosed
Percentage of data enclosed by boundary points
We conjecture that Percent Enclosed may be a lower limit on the yield under
some circumstances. First, each boundary model when simulated must indicate
that the circuit in which it is embedded passes all of its performance
specifications. Second, the mapping functions from the model parameter space
to the specific circuit performance space are monotonic. We have neither proven
this mathematically or demonstrated it empirically at this time.
Secondary Inputs
Density Estimator Percentage
Number of nearest neighbors for density estimation
Diversity Oversampling
Oversampling factor for ensuring spatial diversity
Distance Metric
Formula for distance calculation
Outputs
Nominal vector
Boundary vectors
Both are selected, not constructed
Normalize the Dataset
Outlier screening is necessary because outliers cause the true data to lie along some short
817
interval of one or more of the unit parameter axes. Statistical variation in that parameter
is lost in distance calculation.
Logarithmic transforming is necessary for similar reason because the higher magnitude
data skews the results. Variance is more evenly reported if the data is logarithmically
transformed to make it linearly varying.
Estimating Density Using Nearest Neighbors Method
User needs to understand the trade off between setting Density Estimator Percentage,
which directly controls the number of nearest neighbors, either too high or too low. If set
too high, density estimate is overly smooth and the algorithm can fail to resolve peaks in
the PDF that are close together. Setting it too low leads to a different problem. The
problem is that the distance averaging is a Monte Carlo process and consequently
introduces Monte Carlo noise into the density estimate. This noise can lead to poorly
chosen nominal and boundary models. The problem is more acute for the nominal model,
as the nominal model is in a pronounced region of low slope in the density estimate.
Nominal and Boundary Point Selection
Spatial Diversity---An Illustration
Spatial diversity is done by oversampling and then iteratively forming all possible pairs
and discarding a point from the pair that is closest together. The procedure terminates
when only the desired number of boundary points remain.
Do not know best limits for Spatial Diversity Oversampling factor. At the present 1.5 to
818
3.0 seems like a good range. Clearly, it cannot be 1 or lower since 1 implies no
oversampling. Runs with it set above 3 show no clear advantage. High settings do have a
significant performance penalty for large numbers of requested boundary points.
Validating the Algorithm

Constant density curves for 2-D distributions
2-D Joint Gaussian Distribution with Nominal and

Boundary Points
Handles this bimodal distribution nicely. Ellipses are distorted due to unequal scales on
axes.
Discuss what happen when the distributions are moved close together. NEAREST
NEIGHBORS TOO HIGH.
Point out how the nominal can wander around for NEAREST NEIGHBORS TOO LOW.
Hypershell in 320-D Spherical Gaussian Distribution
Conditional Density Histogram on 4-D Bimodal

Gaussian Distribution
Conditional density histogram are done by selecting point from the distribution such that
all but one of the parameters are constrained to be in some reasonably narrow band of
values for each of the parameters. The remaining parameter is allow to range over its
819
natural region of support. This is equivalent to saying that we wish to repeatedly slice up
the space, except for the last parameter. We then form a histogram on the sample points
that are in the constrained region and a histogram over the boundary points that are also
in that same region. This procedure is somewhat related to the computing of margin
density histograms.
In out case, the expected outcome is to see two approximately Gaussian sample
histograms with the boundary points be in the tails of the two distribution. This is the
result that we observe.
Applying the Tool
Mode 2 may lead to better congruence between boundary model and the boundary
devices in terms of final electrical behavior
Mode 2 is better because we do not expend considerable effort extracting many models
that we ultimately may throw away. If we are going to do direct or generated Monte Carlo
yield estimation then the effort is not in fact wasted.
Mode 2 is also better because one get two outputs nominal and worst-case vectors for
both the measurements and the models.
2-D Analysis of BSIM3 Data
Analysis of Diode Data
Conclusion
Statistical circuit design still beckons
820
More upfront infrastructure and effort
Fewer problems and higher yields during production
Non-parametric boundary analysis is a powerful new tool for statistical circuit design
Easy to use
Accommodates wide range of stochastic processes
Provides nominal and worst-case models via a statistical robust technique
821
Improving The Manufacturability Of
Electronic Designs
Dan Stoneking, Hewlett-Packard,
IEEE Spectrum, June 1999, Volume 36, Number 6
This material appears here with permission of the IEEE. Such permission of the IEEE does
not in any way imply IEEE endorsement of any of Hewlett-Packard's products or services.
Internal or personal use of this material is permitted. However, permission to
reprint/republish this material for advertising or promotional purposes or for creating new
collective works for resale or redistribution must be obtained from the IEEE by sending a
blank email message to [email protected]. By choosing to view this document,
you agree to all provisions of the copyright laws protecting it.
After many weeks or months of laboring over your design with the best available models
and a favorite simulator, you and your management commit it to production. You breathe
a sigh of relief as the first manufactured units work their way to final test. Then, just as
you are transitioning to the next project, the bad news comes: yields are unacceptably
low.
A likely cause is that you overlooked the effects of the stochastic nature of the
manufacturing process. Or perhaps you assumed that random fluctuations in the myriad
of process steps would more or less cancel each other out--and this time they seem to be
canceling rather less than more. Whatever the reason, too many of the units failed to
measure up and became scrap.
Now management is on edge. Customers are unsettled and looking around for options.
The part can be produced, but the low yield means that manufacturing costs will be too
high. Yields can be improved, but that will require more engineering time. In retrospect, it
seems likely that the application of statistical circuit design techniques could have largely
prevented these problems.
Nonparametric boundary analysis

Statistical circuit design techniques analyze the yields of circuit designs whose underlying
components exhibit random fluctuations. These techniques can help produce more robust
designs by calling attention to areas where statistical variations are likely to combine in
such a way as to cause circuit failure. Nonparametric boundary analysis (NBA), see "How
nonparametric bounary analysis works" a technique introduced in Hewlett-Packard
EEsof's IC-CAP 5.0, permits yield analysis when the random fluctuations result from an
arbitrary stochastic process, in addition to well-studied processes such as the Gaussian.
Its ability to analyze data from all stochastic processes is why NBA takes the label
nonparametric, which indicates that it does not attempt to fit a parametric, closed-form
probability density function to the input data.
NBA captures the statistics of the process from data vectors-collections of parameter
values for all of a design's statistically variant components. These parameters can be just
about anything so long as a given set of values uniquely represents each component
sample. Herein lies NBA's flexibility. It allows the input data vectors to be formed from
measurements or from model parameters that represent each of the design's components.
Each component's model can have any number of parameters, and the number of
components in the design can be quite large.
The technique is a straightforward means of estimating a lower bound on yield, and

requires much less computational effort than the well-known technique of Monte Carlo
yield estimation. The Monte Carlo approach, on the other hand, provides an actual yield
estimate, not just a lower bound. It randomly samples the distribution of combinations,
performs a complete circuit simulation, and counts how many give an acceptable result.
For typical distributions, most of the computation effort is expended on simulations that
have near-nominal responses.
NBA requires less computational effort because it first examines the overall distribution of
model parameter values or measurements of components and then selects only those
combinations in the tails of the distribution-the combinations most likely to cause yield
problems-for a complete circuit simulation. The combinations selected enclose a user-
specified fraction of the distribution. NBA's lower bound on yield is obtained by simulating
all the selected combinations and determining whether they all result in circuits with
acceptable responses. If they do, then the yield will be at least as large as the user-
specified enclosure fraction. If not, then the enclosure fraction is reduced and the process
is repeated until a fraction is found for which all of the circuits work properly.
In addition to analyzing yield, the technique identifies a nominal model for each of the
statistically variant components in a design. Nominal models are useful for initiating the
design process because they emulate the various components' ideal behaviors.
Applying NBA
To see how NBA works, consider the voltage divider of Fig. 1. The voltage supply is ideal
with an output of 3 V. The resistors, however, are statistically variant. The NBA inputs are
pairs of R1 and R2 sample values. The responses of interest are V and I. By construction,
the R1-R2 distribution is Gaussian. The enclosure percentage for the R1-R2 distribution is
85 percent. Since the dataset is two-dimensional, the enclosure boundary is clearly
evident for the R1-R2 and I-V spaces. Note that the points in the I-V space that
correspond to the boundary points in the R1-R2 space also create a boundary. This
822
boundary encloses the same percentage-85 percent-of the I-V distribution. Note also the
position of the nominal points, shown in blue, in the input and response distributions.
They are both close to the known means.
The voltage-divider circuit example (top left)
The above figure has statistically variant resistors that illustrate equal-enclosure
percentages in the input (R1-R2) and response(I-V) spaces below . In the I-V plot, the
corner areas within the white rectangle but outside the area defined by the green
boundary points illustrate why nonparametric boundary analysis provides only a lower
bound on yield.
We can also see why NBA provides only a lower bound on the yield and why this bound
can be pessimistic. Compare the 85 percent enclosure in the I-V space with the
hypothetical acceptability region outlined in black. NBA provides only a lower bound
because the 85 percent enclosure is round and the acceptability region has corners that
are not enclosed. Grossly pessimistic yield estimates can be obtained by applying NBA
only once and noting that all the outcomes are acceptable. This situation would pertain if,
for example, the hypothetical acceptability region were expanded. This difficulty can be
corrected simply by iterating the NBA enclosure percentage upward until a failure is found.
NBA has a number of other constraints and limitations on its practical application. An
important constraint is that the functions that map the input space to the response space
must be approximately monotonic if the property of equal enclosures is to hold. Without
equal enclosures in the two spaces, the algorithm does not provide a lower bound on
yield. Approximately monotonic means that the mapping functions can have some ripple.
However, the ripple must be small relative to the overall change in the response functions
across the acceptable response region.
Although approximate monotonicity is difficult to guarantee in all cases, the author

conjectures that the vast majority of practical circuit designs are approximately
monotonic. The monotonicity of a design can be explored by artificially adjusting the
component parameters to trace paths along rays emanating from the nominal design
while verifying that the circuit responses are monotonically increasing or decreasing.
Another method consists of iteratively increasing the NBA enclosure percentage while
verifying for all pair-wise combinations of the response variables that the area (or volume)
of enclosure in the response space is constantly increasing.
Ultimately, even if the circuit is not approximately monotonic with respect to its
parameters, NBA can be used to explore for worst-case designs in the performance space.
Dimensionality considerations
Another concern when applying NBA is to limit the dimensionality (the number of real
numbers per input vector) of each design's parameter space. This is important for several
reasons. As the dimensionality goes up, the user must populate that high-dimensional
space with data. The amount of data needed grows rapidly. This is mostly a concern for
components that require many parameters for a unique representation. Getting enough
samples of the overall design is not normally a problem because of the large numbers of
combinations that can be formed from modest numbers of samples for each component.
Another point to watch regarding dimensionality is that the method's density estimate
degrades as dimensionality increases. The author has shown proper functionality with
dimensionalities as high as 320. Proper performance into low- to mid-thousands is viable.
Beyond the logistics of getting the data and algorithm performance, the user should be
aware of an interesting phenomena of very high-dimensional spaces: as the number of
dimensions goes to infinity, the volume contained in a very thin shell on the surface of a
hypersphere containing one unit of volume also goes to unity. That is, all the volume of a
high-dimensional hypersphere is on its surface. This phenomena ultimately limits NBA's
ability to return a meaningful set of boundary points for a given enclosure percentage.
823
Several tools are available to help the user of NBA limit dimensionality. Direct knowledge
and engineering judgment assist in selecting appropriate model parameters or
measurements to represent each component in the overall design. Obviously, parameters
to which the component's responses are not sensitive should be avoided. If using
measurements, do not choose data that shows little variation across the samples. Avoid
using parameters or measurements that are highly correlated with one another. Beyond
engineering judgment, employ correlation and sensitivity analyses to determine the
important parameters or measurements.
NBA has different data requirements from other statistical circuit design methodologies.
This is both good and bad. To apply other methods, it is normal to use a compact set of
statistics derived from the raw data-mean vectors and covariance matrices for Gaussian
distributions or more exotic distributions along with their pertinent statistics. NBA,
however, needs the raw data. This is not as onerous a demand as it may at first appear,
however. To get the compact statistics in the first place, the raw data must be gathered-
no extra work here. Although the volume of NBA's data is much larger, it is not
impractically so, and today's Internet makes user access simple and quick.
A big advantage of NBA is that it never requires that a probability density function be
fitted to the data. It thereby avoids a host of problems, not the least being the often
unjustified assumption that the data fits a Gaussian model. Even if the probability density
model is appropriate, fitting any model to any data inherently leads to an imperfect
emulation.
The last major concern for NBA is how large a circuit can be analyzed. To answer that
question, it is helpful first to consider the various statistical circuit design flows that use
NBA. As is turns out, there are several answers depending on the flow that one employs.
Statistical circuit design flows using NBA

Different statistical circuit design flows are possible for more complicated circuits Fig. 2 .
The first flow in the figure shows the application of Monte Carlo yield estimation (MCYE).
The last three flows are variations of NBA-the measured data (MD), model parameter
(MP), and reference model parameter (RMP) methods. All focus on designs in which
transistors are the prominent cause of circuit variability. Designs with other statistically
variant components would use identical flows. As a visual cue to the reader, similar
functionality steps carry the same color coding in all the flows.
In common with all the others, the MCYE flow assumes that the design under analysis has
a total of five transistors of different sizes light blue block . Two hundred samples of each
of the transistors are measured to produce 1000 measurement datasets light purple block
. One thousand models are then extracted orange block . Next, form as many
combinations of five transistors as it is practical to simulate (200-1000) by randomly
selecting a model from each pool of 200 transistor models green block . Then run full-
circuit simulations with the combinations and get a yield estimate gray block . If the yield
estimate is low, repeat as shown until the yield becomes satisfactory black block .
824
Statistical circuit design flows are diagrammed for comparison .
From top to bottom, they are:traditional Monte Carlo yield estimation; nonparametric
boundary analysis operating on measured data from statistically variant components;
nonparametric boundary analysis applied to model parameters of the components; and
use of a reference model and scaling rules to decrease effort needed for a nonparametric
boundary analysis.
The NBA-MD flow introduces an extra step into the MCYE flow-perform NBA after choosing
a minimum desired yield tan block . In addition, it transposes the steps for forming the
random combinations and extracting models. This transposition matters both from the
standpoint of how many combinations are considered (coverage of a high-dimensional
space) and because of the amount of effort expended on extracting the models and
running the full-circuit simulations.
The NBA-MP flow is essentially that of the MCYE flow. The NBA-MP flow has a "perform
NBA" block interjected between the "form random combinations" and "full-circuit
simulation" blocks of the MYCE analysis. This flow diminishes the effort in the iteration
loop but still requires extraction of a model for all the measurement datasets.
The NBA-RMP flow begins like the NBA-MP flow but relies on a priori knowledge of how to
translate the model parameter values of the measured reference transistor into the
parameter values for all the other transistors. If this translation can be done effectively,
less effort by far is needed in gathering the measured datasets.
The main difference between MCYE and the NBA methods is the number of full-circuit
simulations required to interrogate the nominal circuit design. For the present example,
which assumes typical counts, MCYE requires a factor of 10-50 more full-circuit
simulations per iteration. Given that simulation time can be large and that the full-circuit
simulations are in an iteration loop, significant savings in engineering time will result from
using NBA.
Another key point in comparing MCYE to NBA-MD and NBA-MP is the extent to which the
methods explore the input data space. Note the number of combinations formed for the
MCYE, NBA-MD and NBA-MP methods. The NBA methods accommodate far greater
numbers of combinations because they select only a tiny subset of the input points, which
then are run through full-circuit simulation. MCYE's weakness is that it runs all the
combinations through full-circuit simulation.
Comparing the NBA methods indicates that NBA-RMP requires the least effort and NBA-MP
the most. NBA-RMP enjoys its position because fairly few transistors are measured and
extracted. The weak link in NBA-RMP is the transistor scaling operation light green block .
For example, it implicitly assumes that all of a design's transistors are deterministically
825
related-that there are no intra-die variations.
NBA-MP takes the most effort, and may not offer any advantage in accuracy over NBA-
MD, which requires the mid-range effort. Low data dimensionality and data fidelity drive
accuracy in any NBA method. If the underlying models are sufficiently compact, then NBA-
MP offers lower dimensionality. All the same, this lower dimensionality from models as
compared to measurements is not a given. NBA-MD always offers higher fidelity than
NBA-MP because extracting a model from measurements always introduces noise into the
NBA input data.
How big is big?
Returning to the question of the size of circuit amenable to an NBA analysis, assume that
the maximum dimensionality of the input data to the NBA algorithm is 2000. Assume also
that the average number of measurement data or model parameters per statistically
variant device is 20. Then using the NBA-MD or NBA-MP methods would permit analysis of
circuits with 100 statistically variant devices-small to maybe modest-sized analog circuits.
Keep in mind that if identical devices are used more than once, then the maximum circuit
size is leveraged by the number of times each unique device is used.
Far larger circuits can be accommodated if the NBA-RMP method is employed. With this
method, each of 100 statistically variant reference devices can represent tens to even
thousands of individual devices in the circuit designs. The number of devices in an
amenable circuit then grows into the 1 000 to 100 000 range.
NBA in action
To see how NBA works, consider how it was actually used in the design of a microwave
amplifier. NBA-RMP was applied to a high-gain, wide-bandwidth HMMC5040 amplifier
manufactured by Hewlett-Packard's Microwave Technology Division. The amplifier's overall
design is pictorially rendered in the schematic of Fig. 3. The procedure begins with 108
samples of input data concatenated with response data. The inputs are data derived from
reference models of transistors in the process control monitor die. The response data are
amplifier performances of interest. Table 1 summarizes the input and response data.
The five transistors are the principal sources of circuit variability in this representation of
the Hewlett-Packard HMMC5040 RF amplifier above . The amplifier stages are connected
by matching networks of proprietary design.
The result of most interest is how well nonparametric boundary analysis helps determine a
lower bound for the yield. Since acceptable circuit responses are usually specified in terms
of independent upper and lower limits for each response, acceptability regions are
hyperboxes. (A hyperbox is the generalization of a rectangle in two dimensions and a box
in three dimensions to higher-dimensional spaces.) Three enclosure percentages are
compared with Monte Carlo yield estimates. The Monte Carlo estimates are then 34, 89,
and 97 percent for the hyperboxes enclosing the 25-, 50-, and 100-percent boundary
point sets. The expected result is that the analysis enclosure percentages represent lower
bounds on the actual yields.
Calculating a lower bound on yield is also possible with parametric worst-case methods so
long as the underlying data distributions are Gaussian. Inspection of Fig. 4 illustrates that
the compression versus large-signal (LS) gain response data are clearly non-Gaussian.
Also, we see overlays of the nominal point and the boundary points for 25, 50, and 100
percent enclosures. Visually, the nominal and boundary points appear consistent with the
response scatter although the analysis was done on the input variables.
826
The response data in this plot of compression versus large-signal gain right is clearly not
Gaussian, so parametric methods cannot be used for yield estimation. The nonparametric
method, on the other hand, does yield useful results.
The data presented is a projection of the responses, which are eight-dimensional, onto the
compression versus LS-gain plane. Interesting problems arise when a human being tries
to examine multidimensional boundaries on a two-dimensional sheet of paper-to wit, they
can't be seen. This does not mean that the boundary is absent, but rather that the
boundary points do not form an enclosed region in this two-dimensional projection.
Acknowledgments
The analysis of the HMMC5040 would not have been possible without the efforts of John
Wood and David Root. David Root and Ryan Zurakowski contributed a a number of
important ideas over the course of nonparametric boundary analysis's development. Many
thanks go to Brian Hughes for his insightful suggestions regarding the flow and
presentation of this article.
--------------------------------------------------------------------------------
To probe further
In Handbook of Statistical Methods for Engineers and Scientists, editor Harrison M.
Wadsworth supplies an accessible overview of statistical methods for the practicing
engineer (McGraw-Hill, New York, 1990). His handbook has considerable breadth, but is
geared for those who just want to know how to get the job done.
Michael D. Meehan and John Purviance present a lively review of statistical circuit design
in their book Yield and Reliability in Microwave Circuit and System Design (Artech House,
Boston, 1993). The authors discuss not only the many algorithms in the field, but also
when and how to apply these methods.
David W. Scott's book, Multivariate Density Estimation Theory, Practice, and Visualization
(John Wiley & Sons, New York, 1992), provides the reader with an introduction to density
estimation in multivariate (multidimensional) spaces.
Further information about the theory and application of nonparametric boundary analysis
is available in the form of U.S. Patent No. 5,835,891, "Device Modeling Using Non-
Parametric Statistical Determination of Boundary Data Vectors," which was issued to the
author, Dan Stoneking, on 10 November 1998. Dan also wrote "Statistical Circuit Design
and IC-CAP's Non-Parametric Boundary Analysis," which is in HP EEsof's Characterization
Solutions (Order No. 5965-8931EUS).
--------------------------------------------------------------------------------
About the author

Dan Stoneking joined Hewlett-Packard's EEsof Division in Santa Rosa, Calif., as a software
design engineer in 1994. His current responsibilities and interests include behavioral
modeling of nonlinear analog circuits and the advancement of statistical circuit design
techniques. Before joining HP, he was a senior engineer with M/A-COM in Lowell, Mass.
His work there focused on applying statistical techniques to improve both manufacturing
and design processes for electronic circuits.
--------------------------------------------------------------------------------
Spectrum editor: Michael J. Riezenman

IEEE Spectrum June 1999 Volume 36 Number 6
(c) Copyright 1999, The Institute of Electrical and Electronics Engineers, Inc.
How nonparametric boundary analysis works
Nonparametric boundary analysis (NBA) finds points lying near the surfaces of equi-
density that occur in collections of multidimensional data. The collections are either
measurement data or else model parameters that represent specific parts, transistors, or
components. The algorithm uses equi-density surfaces because the region interior to the
surface contains the most frequently occurring parts. The probability of manufacturing a
part that falls in the region is directly related to the manufacturing process yield. Further,
it is straightforward to find points near such a surface.
To perform these operations, the NBA algorithm must estimate density for all points in the
827
data collection and have a technique for finding the equi-density surface appropriate for a
given yield. The density estimator uses a nearest neighbor method figure at right . The
nearest neighbor method sorts the data, repetitively locates nearest neighbors for each
point, calculates distances, and finally computes density estimates.
A specific feature of NBA's method is the density metric formulation. This formulation is
advantageous because it does not aggravate the inherent Monte Carlo noise present in the
computed average distance to the nearest neighbors. A dimensionally correct density
estimate suffers from ever poorer signal-to-noise ratio as the dimensionality of the
problem increases. To some extent, the noise can be mitigated by increasing the number
of nearest neighbors. But although this smoothes the density estimate, excessive
smoothing can obscure important features in the density function. That, in turn, can lead
to erroneous results. A good rule of thumb in this regard is to keep the number of nearest
neighbors below the number of points in the smallest data cluster.
Once the density estimates are available for all the points, NBA sorts them into order per
their density estimate above . NBA then finds the points near a surface of equi-density.
The surface of interest is the one that encloses the desired percentage of the distribution.
NBA proceeds by placing a cutline in the sorted list that encloses the desired percentage of
points as shown. The boundary points are then those to the left of the cutline.
Since clustering of the returned boundary points is not desirable, NBA oversamples to the
left of the cutline to produce the candidate boundary points. Obviously, high oversampling
degrades the approximation that the candidate boundary points are close to the desired
equi-density surface. The algorithm has a heuristic constraint to prevent excessive
oversampling.
The candidate boundary points then pass into the spatial diversity algorithm, which
outputs the final boundary points to the user. The spatial diversity algorithm works by
iteratively forming all candidate boundary points pairs, identifying the pair closest
together, and then discarding one point from the closest pair.
The net result of this iterative process is that the final boundary points are spatially
distributed along the desired surface of equi-density
828

Iccap

Uploaded by

Copyright:

Available Formats

Iccap

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Iccap

Uploaded by

Copyright:

Available Formats

IC-CAP Modeling Handbook

IC-CAP Modeling Handbook

Author Name: Franz Sischka

1. Creating a CMOS Inverter Circuit

Creating a CMOS Inverter Circuit

Figure 1 : A Simple CMOS Inverter Circuit

Main Steps of the Semiconductor Process

Step Semiconductor Description

Semiconductor Process Flow

The following sequential steps describe the semiconductor process flow:

Process and Device Modeling

These high-speed measurements are performed on as many as available devices on the

Resistivity Continuity/Bridging Leakage/Breakdown Capacitance/ CV IV Curves

There is more information about the tests in the Appendix.

In addition to the standard traditional IV and CV measurements and curves, reliability

The following are three main areas of concern in Reliability Testing:

Hot Carrier Injection

Why a DC Analyzer has 4 SMUs?

Importance of Triaxial Cables

Sampling measurements involve repeating a constant (rather than incrementing) forced

In a sweep measurement, voltages or currents are increasingly swept between each

Up to 4 resources (ie SMUs, VSUs, or PGUs) can be defined as DC or AC stress sources as

CV measurement can reveal :

If a high frequency (>100kHz) measurement is made, the minority carriers (electrons)

If a low frequency capacitance measurement is used, an entirely different result is

Why a Switching Matrix?

Resistivity measurements are used to evaluate processes related to creation of conductive

The Film thickness, cross-section, purity

Continuity and bridging tests can reveal:

Particles / defects in film

Leakage / BreakdownIsolation Edge Effects

Dielectrics and diffused junctions can be very sensitive to contamination or defects

Reverse-bias testing can reveal:

Ion implant damage

Basics of DC and AC Characterization of

While the characterization of electronic components in the DC domain is relatively simple

Modeling of a nonlinear component always begins with the characterization of its DC

PlotOption Access Name and Value Examples

The Guarding (Triax Cable) Principle for Fast and

Triax Feedthru and Triax-Coax Converters From Trompeter

Oscillations during DC measurements

1. SMU Induced oscillations. These oscillations occur when an inductive load is

Methods to eliminate oscillations

There are some effective methods to eliminate oscillations:

Self-heating during DC measurements

What to do when self-heating becomes an issue for your device modeling:

CV Measurements - Calibration considerations

CV Meter Applying the Auto-balancing Method

For CV meters, the measurement principle is typically a so-called auto-balancing method.

Avoiding Measurement of Parasitic Capacitance

A so-called four-wire method is used for CV measurements. Similarly to the DC Kelvin

AttoGuard Enhanced CV Measurements

FemtoGuard surrounds the chuck at shield ground

S-Parameters and Characteristic Impedance Z0

Transmission Coefficient = Transmitted / Incident

Calculating S-Parameters From Voltages

Understanding the Smith Chart

Why a Smith Chart for the Sxx Parameters?

Interpreting S11 and S22 (Sxx)