An Alternate Test Strategy

for Differential RF Mixers

J. Liaperdos* L. Dermentzoglou, A. Arapoyanni Y. Tsiatouhas
Dept. of Information Technology *Dept. of Informatics Dept. of Computer Science
and Telecommunications and Telecommunications University of Ioannina
Technological Educational Institute of Kalamata University of Athens Ioannina, Greece
Sparta, Greece Athens, Greece

Abstract—In this work we develop a new strategy for the built- and to accurately predict circuit performance characteristics
in testing of differential RF Mixers which follows the alternate (e.g. Gain, 1-dB compression point, etc.) from the correspond-
test paradigm, whose objective is to predict rather than to directly ing alternate test response [3]. Appropriate selection of the
measure a circuit’s performance characteristics from its response
to a suitable test stimulus. The proposed strategy adopts the use set of observables which compose the test response crucially
of the Local Oscillator’s (LO) signal as the test stimulus applied determines prediction accuracy [4]. Performance prediction
to the inputs of the Mixer. The self-mixing of the LO signal forces is based on nonlinear models, which could be developed
the Mixer to operate in homodyne (zero IF) mode, producing DC using statistical regression methods (i.e. multivariate adaptive
levels at its “IF” outputs. These DC levels (IF+, IF-) are used as regression splines (MARS) [5]) applied on an initial extended
the main voltage observables, together with the DC component
of the voltage drop across the Mixer’s current source (Vtail) set of measured observables. Since a single test configuration
which is introduced as a third observable leading to a significant and a single test stimulus for multiple specifications is used in
improvement in performance prediction accuracy, as proved by this case, generation and acquisition of the test response is less
the application of the proposed strategy to a typical differential complicated compared to the conventional testing approach.
RF Mixer designed in a 0.18µm CMOS technology. It has also
been found that there is a further increase in prediction accuracy The alternate test paradigm is used in this work to develop
if two different sets of measurements are performed for these a strategy for testing differential RF Mixers. The proposed
observables by varying the Mixer’s supply voltage (Vdd). test methodology is based on the selection of proper mixer
node voltage observables to form the alternate test response
I. I NTRODUCTION and their use to derive predictive models for desired circuit
performance characteristics. Furthermore, it is shown that a
In conventional RF Integrated Circuit (IC) testing, automatic
significant increase in alternate-test efficiency is observed if
test equipment (ATE) is used to measure the performance char-
the test response is extended by the incorporation of replicas
acteristics of the circuit under test (CUT), one by one. Upon
of the voltage observables which correspond to different values
conclusion of such tests, the CUT is classified as fault-free if
of the mixer’s supply voltage. The paper is organized as
each characteristic lies within an acceptable range, according
follows: In Section II the proposed test strategy is presented.
to specifications. Although these measurements are simple,
In Section III, the test procedure is evaluated and simulation
they require a variety of test resources which, together with the
results on a typical RF Mixer case study are provided. Finally,
long test application times, increase the total manufacturing
the conclusions are quoted in Section V.
cost. In many cases testing cost turns to be unacceptable,
tending to be comparable to the rest manufacturing cost [1].
II. A LTERNATE T EST S TRATEGY OVERVIEW
A main contributor to the raise of conventional RF testing
cost is the expensive automatic test equipment (ATE) used to
The proposed alternate test strategy adopts the use of the
measure the performance characteristics of the circuit under
Local Oscillator’s (LO) signal as the test stimulus at the inputs
test. Limitations on conventional testing are posed by the
of the Mixer [4], [6], [7]. The self-mixing of the LO signal
difficulty of the ATE to directly access all, or even part of the
forces the Mixer to operate in homodyne (zero IF) mode,
internal nodes of an IC, especially in System-on-Chip (SoC)
generating DC levels at its “IF” outputs. These DC levels
or System-on-Package (SoP) designs. Although some internal
(IF+, IF-) are used to compose the alternate test response,
signals can be made available to the external tester, frequency
together with the DC component of the voltage drop across the
limitations due to lower speed of the I/O interface may not
Mixer’s current source (Vtail) which is introduced as a third
permit their direct observation.
observable, leading to a significant improvement in prediction
To overcome the cost and inabilities of functional testing, accuracy, as shown in the following section. A further increase
the concept of alternate test was proposed [2]. The objective of in prediction accuracy is achieved if the alternate test response
the alternate test methodology is to find a suitable test stimulus consists of two different sets of measurements acquired for

2nd Pan-Hellenic Conference on Electronics and Telecommunications - PACET΄12, March 16-18, 2012, Thessaloniki, Greece
 MARS
Models
V1N
C1=f1(V1,V2, ,Vq)
V2N
C2=f2(V1,V2, ,Vq)
MARS 500Ω 500Ω
1 Analysis
2
N
VqN Ck=fk(V1,V2, ,Vq)

Fault-free Mixer
Instances
300µm 300µm 300µm 300µm
340nm 340nm 340nm 340nm
Figure 1. Predicive models’ training procedure
200µm 200µm
340nm 340nm

the above observables by varying the Mixer’s supply voltage

20kΩ 3.14nH 3.14nH
(Vdd). Vtail

A preliminary training procedure, through which the pre-

dictive models are prepared, is necessary before the main 7µm . W = 300µm
testing procedure is conducted, both described in the following 500nm L 500nm

subsections.
Figure 2. The RF Mixer under consideration.
A. Training

During the training procedure, a set of models dedicated to TABLE I. RF MIXER PERFORMANCE CHARACTERISTICS
the prediction of k performance characteristics (C1 ,C2 , . . . ,Ck ) Parameter Value (Typical)
of interest (e.g. gain, 1-dB compression point etc.) is derived. fRF 1.9GHz
fIF 150MHz
Initially, q voltage observables V1 ,V2 , . . . ,Vq are optimally LO power 5.0dBm
selected so that they offer the most accurate prediction feasible Gain 4.9dB
NF SSB 11.7dB
for the k performance characteristics. Then, Monte Carlo Input IP3 10.5dBm
simulations are performed over N fault-free mixer instances, Input 1dB CP 0dBm
taking into account process variations and device mismatches. Supply Voltage 3.3V
Power Consumption 15mW
These simulations provide N training sets of the q voltage
observables. Afterwards, a set of k regression models, de-
scribed by mapping functions f1 ,f2 , . . . ,fk are generated by
MARS (Multivariate Adaptive Regression Splines) analysis transistors transform the input voltage into current, which is
[5], using the above N training sets of observables as input further commutated to the complementary IF outputs each LO
(see Fig.1). Each mapping function fi (1 ≤ i ≤ k), relates the period through output load resistances R1/R2. Degeneration
performance characteristic Ci to the set of voltage observables inductors L1/L2 are introduced to improve linearity. The main
(Ci = fi (V1 , V2 , . . . , Vq )). device parameters of the design are presented in Fig. 2 while
the Mixer’s typical performance characteristics are summa-
B. Testing rized in Table I. Among the characteristics shown in Table I,
Gain (G), 1-dB Compression Point (1 dB CP) and 3rd order
During the actual testing of a fabricated Mixer the voltage input referred Intercept Point (IP3) are considered as the target
observables are measured. Then the performance characteris- performance characteristics of interest.
tics are predicted using the regression models already avail-
able. If all predicted performance characteristics fall within A. Alternate Test Response
their specification-defined tolerance bands, then the mixer is
considered to be fault-free and it is shipped to the end-user. The DC levels (IF+, IF-) mentioned in the previous section
are used as the main voltage observables, together with the
III. A C ASE S TUDY – S IMULATION R ESULTS DC component of Vtail (see Fig. 2). Therefore, the set of
voltage observables used for the application of our method
The effectiveness of the proposed test strategy has been
are IF+, IF- and Vtail. It has been found that prediction
evaluated by simulations on a typical RF Mixer which is
accuracy increases significantly if two different measurements
presented in Fig. 2. The Mixer was designed in a standard
are performed for these observables: One by operating the
0.18µm CMOS technology (Vdd=3.3V) with an intermediate
mixer using its nominal voltage supply value (Vdd=3.3V), and
frequency (IF) of 150MHz. Its topology is based on the double
another by using a lower value (e.g. Vdd=3V). Both voltage
balanced Gilbert cell Mixer [8].
triads form the Alternate Test Response (ATR) from which
The circuit consists of an RF stage (transistors M1/M2) the regression models are derived to be used for performance
and a differential LO stage (transistors M3–M6). The RF prediction, according to Fig. 3.
 Vdd=3V Test Response Acquisition Mode

Normal Mode

Normal Mode
phase 1 phase 2
IF+ (Vdd=3V) (Vdd=3.3V)
Vdd=3.3V IF- sub-phase 1 sub-phase 2 sub-phase 3 sub-phasesub-phase
1 3
Vtail (IF+) (IF-) (Vtail) (IF+) (Vtail)
Predictive Performance
IF+ Model Characteristic
IF- (e.g.1dB-CP)
1
Vtail
2
Alternate Test T T T 3T
Response (ATR)
1 Select_V:0 2 Select_V:1

Figure 3. Performance prediction based on the Alternate Test Response.

Figure 5. Timing diagram for the ATR acquisition procedure.
LO
= e-fuse

LO+ LO-
domain can be exploited to deliver the test control signals
(Select V<0:1>) shown in Fig. 4.

MIXER
Aiming to avoid the influence of the LO signal on the RF
RF+ IF+
signal path in the normal mode of operation, through the first
LNA RF− IF−
switch, low cost electrical fuses (e-fuses) (or laser-cut fuses),
Vtail commonly used in memory circuit repair operations [9] and
Switch-1
Select_V<0:1> data converters trimming [10], can be optionally exploited to
Switch-2
VS
2 eliminate the LO-RF test path after the completion of the test
Select_V:0 Select_V:1 LPF
procedure.
VDC A simplified timing diagram for the mixer in ATR acquisi-
tion (test) mode, which is initiated when at least one of the
Figure 4. Auxiliary circuitry for ATR acquisition.
two test control signals (Select V<0:1>) is high, is shown in
Fig. 5. This diagram corresponds to the acquisition of both
voltage triads (IF+, IF-, Vtail) which compose the ATR and
B. Acquisition of Voltage Observables
implements the ATR acquisition procedure. The “Select V”
A possible implementation that allows the acquisition of signals determine the element of the ATR that is obtained,
the ATR could follow the scheme presented in Fig. 4, which while the total time of the operation depends mainly on the
illustrates the case of a differential RF Mixer in a receiver. acquisition time (T) required for each voltage level and is
An analog switch (Switch-1) selects the input signal to the determined by the settling time of the circuitry involved in
Mixer either in normal operation or during the acquisition of analog to digital conversion.
the ATR voltage observables. In the latter case, the Mixer’s It is well known that analog and especially RF circuits
input is disconnected from the LNA and connected to the LO. are extremely susceptible to capacitive output loads that
In standalone Mixer circuits the voltage observables can be may degrade their performance. In addition, the used analog
directly processed by an external tester, with the addition of switches affect normal operation. It should be consequently
an extra pad for Vtail. On the contrary, in embedded Mixers expected that the insertion of the auxiliary circuitry into
with no direct access to their outputs, a second analog switch a finalized RF Mixer design may change its performance
(Switch-2) can be utilized as shown in Fig. 4. This is used characteristics setting them outside the specifications. To avoid
to select a voltage observable among IF+, IF- and Vtail, one this phenomenon, the Mixer and the auxiliary circuitry are
at a time. The output signal VS of the second switch can designed in common, so that the former meets the design
be connected to an analog test bus (e.g. IEEE 1149.4 std.) specifications in the presence of the latter.
that may exist to provide test facilities for the whole analog
domain. Then, the VS signal can be processed by an external C. Alternate Test Efficiency – Evaluation
analog tester.
The efficiency of the proposed alternate test strategy has
Alternatively, in case of a SoC application, a common RC been evaluated by the calculation of the prediction’s coef-
Low-Pass Filter (LPF) can be added at the output of the second ficients of determination (R2 ) [5] for different sets of ATR
switch to reject any high frequency components and to provide voltage observables, by using a sample of 100 mixer instances
a DC voltage signal VDC. The VDC signal is converted to for which performance characteristics were obtained by 3σ
the digital domain by the Analog-to-Digital Converter (ADC) Monte Carlo simulations and then compared to their predicted
embedded in the system and then processed by a digital tester counterparts. A training set of 500 instances was used to build
or by the Digital Signal Processor (DSP) that is also embedded the predictive models, following the procedure described in
in the system, according to the procedure proposed in Section Section II-A. During these 3σ Monte Carlo simulations the
II. In addition, existing scan testing facilities in the digital LHS (Latin Hypercube Sampling) option was activated, since
 5.5 1.5 12

11.5
the prediction accuracy drastically increases, as shown in the
1
third row of Table II and graphically represented in Fig. 7.

1dB CP, predicted (dBm)

5

IP3, predicted (dBm)

G, predicted (dB)

11
0.5

4.5 10.5 The efficiency of the alternate test procedure can be further
0
10 evaluated by calculating the error regarding the classification
4
-0.5
9.5 of a mixer’s instance as compliant to its specifications or not.
3.5
3.5 4 4.5 5 5.5
-1
-1 -0.5 0 0.5 1 1.5
9
9 10 11 12
In this study we consider that specifications for the mixer
G, actual (dB) 1dB CP, actual (dBm) IP3, actual (dBm)
under consideration (see also Table I for typical values) are
Figure 6. Prediction accuracy plots for G, 1dB CP and IP3; prior work [6].
as follows: 4dB≤G≤5dB, 1 dB CP≥1dBm and IP3≥8dBm.
(Dashed line: Actual=Predicted) Using the same sample of 100 mixer instances as above, we
have calculated the number of compliant circuits based on both
5.4 1.5 12

5.2
simulation-derived performance and alternate test predictions.
11.5
5
1
The results, summarized in Table III, show that the prediction
1dB CP, predicted (dBm)

IP3, predicted (dBm)

G, predicted (dB)

11
4.8
0.5 error in the characterization of mixer instances as compliant
4.6

4.4 0
10.5
with the specifications or not is reduced if the proposed
4.2
10
strategy is applied, remaining significantly lower compared to
-0.5
4 9.5
other alternate test strategies.
3.8
-1 9
4 4.5 5 -1 -0.5 0 0.5 1 1.5 9 10 11 12
G, actual (dB) 1dB CP, actual (dBm) IP3, actual (dBm)
IV. C ONCLUSION
Figure 7. Prediction accuracy plots for G, 1dB CP and IP3; this work.
(Dashed line: Actual=Predicted) In this paper, an alternate test strategy suitable for differen-
tial RF Mixers is presented. The LO signal is used as a single
TABLE II. PREDICTION ACCURACY (R2) FOR DIFFERENT test stimulus, drastically reducing testing time and cost. Highly
SETS OF ATR VOLTAGE OBSERVABLES (V)
accurate regression models are developed for the prediction
R2 (%)
ID Voltage Observables - V of the mixer’s performance by the appropriate selection of
G 1dB CP IP3
1 IF+ ─ IF- 37.36 1.83 3.07 the alternate test response components, both by means of the
2 IF+, IF-, Vtail 96.81 55.76 50.66 mixer’s node from which they are obtained and the mixer’s
(IF+, IF-, Vtail) @Vdd=3.3V, supply voltage applied during their acquisition. Simulations
3 97.44 88.08 91.20
(IF+, IF-, Vtail) @Vdd=3V
have shown a significantly reduced test error compared to
existing alternate test strategies.
TABLE III. ALTERNATE TEST EFFICIENCY FOR DIFFERENT
SETS OF VOLTAGE OBSERVABLES (CASE STUDY)
Compliant* Instances (%)
ID Voltage Observables - V
Actual Predicted Error R EFERENCES
1 IF+ ─ IF- 87 100 14.94
2 IF+, IF-, Vtail 87 98 12.64 [1] SIA - The International Technology Roadmap for Semiconductors.
(IF+, IF-, Vtail) @Vdd=3.3V, [Online]. Available: http://public.itrs.net
3 87 90 03.45 [2] P. Variyam, S. Cherubal, and A. Chatterjee, “Prediction of analog per-
(IF+, IF-, Vtail) @Vdd=3V
* specifications: 4dB≤G≤5dB, 1dB CP≥-1dBm, IP3≥8dBm formance parameters using fast transient testing,” IEEE Trans. Comput.-
Aided Design Integr. Circuits Syst., vol. 21, no. 3, pp. 349–361, Mar.
2002.
[3] S. S. Akbay, A. Halder, A. Chatterjee, and D. Keezer, “Low-cost test
it improves the models’ accuracy by increasing the number of of embedded RF/analog/mixed-signal circuits in SOPs,” IEEE Trans.
Comput.-Aided Design Integr. Circuits Syst., vol. 27, no. 2, pp. 352–
generated samples at the tails of a distribution. 363, May 2004.
[4] E. Garcia-Moreno, K. Suenaga, R. Picos, S. Bota, M. Roca, and E. Isern,
It has been found that using the differential output voltage “Predictive test strategy for CMOS RF mixers,” Integration, the VLSI
of the mixer (IF+ – IF-) as a test observable, following the Journal, vol. 42, pp. 95–102, Jan. 2009.
approach proposed in [6], provides poor prediction efficiency [5] J. H. Friedman, “Multivariate adaptive regression splines,” Ann. Stat.,
vol. 19, pp. 1–141, 1991.
for the specific mixer design, as shown in the first row [6] A. Goyal, M. Swaminathan, and A. Chatterjee, “Self-calibrating embed-
of Table II and graphically represented in Fig. 6. The ded RF down-conversion mixers,” in Proc. IEEE Asian Test Symposium
introduction of Vtail and the use of the mixer’s single-ended (ATS), Taichung, Taiwan, 2009, pp. 249–254.
[7] L. Dermentzoglou, A. Arapoyanni, and Y. Tsiatouhas, “A built-in self-
outputs as voltage observables seem to lead to a significant test technique for RF mixers,” in IEEE Int. Sym. Design and Diagnostics
increase in prediction accuracy, as shown in the second row of of Electronic Circuits and Systems (DDECS), 2010, pp. 88–92.
Table II. It has also been proven that there is a further increase [8] S. A. Maas, Microwave Mixers. Artech House Publishers, 1993.
[9] J. Safran, A. Leslie, G. Fredeman, C. Kothandaraman, A. Cestero,
in prediction accuracy if two different sets of measurements C. Xiang, R. Rajeevakumar, K. Deok-Kee, L. Y. Zun, N. R. D. Moy,
are performed for these observables by varying the Mixers T. Kirihata, and S. Iyer, “A compact e-FUSE programmable array
supply voltage (Vdd). More specifically, if the acquisition of memory for SOI CMOS,” in Proc. IEEE Symposium on VLSI Circuits,
2007, pp. 72–73.
IF+, IF- and Vtail is performed not only for the nominal [10] L. Counts, “Analog and mixed-signal innovation: The process-circuit-
supply voltage (Vdd=3.3V in the used technology) but also system-application interaction,” in Proc. IEEE Int. Solid-State Circuits
for a second supply voltage (Vdd=3.0V in our experiments), Conf., 2007, pp. 26–32.

View publication stats