qt5wb14729 Nosplash

UNIVERSITY OF CALIFORNIA SAN DIEGO
Robust Physical Design and Design Technology Co-Optimization Methodologies at

Advanced VLSI Technology
A dissertation submitted in partial satisfaction of the

requirements for the degree
Doctor of Philosophy
in
Electrical Engineering (Computer Engineering)
by
Minsoo Kim
Committee in charge:
Professor Andrew B. Kahng, Chair

Professor Chung-Kuan Cheng
Professor Sujit Dey
Professor Farinaz Koushanfar
2023
Copyright
Minsoo Kim, 2023
All rights reserved.
The dissertation of Minsoo Kim is approved, and it is acceptable in quality and
form for publication on microfilm and electronically.
University of California San Diego

2023
iii
DEDICATION
To my family.
iv
TABLE OF CONTENTS
Dissertation Approval Page . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii
Dedication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv
Table of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v
List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii
List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiv
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xviii
Vita . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxi
Abstract of the Dissertation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxiv
Chapter 1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 New Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 Challenges for Modern IC Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.2 Challenges for Advanced Technology Nodes . . . . . . . . . . . . . . . . . . . . . . . 3
1.1.3 Challenges for Design-Technology Co-Optimization . . . . . . . . . . . . . . . . 4
1.2 This Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
Chapter 2 General Physical Design Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

2.1 Optimal Bounded-Skew Steiner Trees to Minimize Maximum k-Active Dynamic
Power . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.1.1 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.1.2 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.1.3 Experimental Setup and Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.1.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.2 CoRe-ECO: Concurrent Refinement of Detailed Place-and-Route for an Efficient
ECO Automation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.2.1 CoRe-ECO Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.2.2 Switchbox Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.2.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
2.3 Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
Chapter 3 Technology-Aware Physical Design Methodology . . . . . . . . . . . . . . . . . . . . . 59

3.1 Diffusion Break-Aware Leakage Power Optimization and Detailed Placement in
Sub-10nm VLSI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.1.1 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.1.2 2nd DB-Aware Leakage Optimization and Placement . . . . . . . . . . . . . . . . 65
v
3.1.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
3.2 Detailed Placement for IR Drop Mitigation by Power Staple Insertion in Sub-
10nm VLSI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
3.2.1 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
3.2.2 Our Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
3.2.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
Chapter 4 Design-Technology Co-Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

4.1 PROBE2.0: A Systematic Framework for Routability Assessment from Technol-
ogy to Design in Advanced Nodes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
4.1.1 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
4.1.2 The PROBE2.0 Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
4.1.3 Routability Assessment and Learning-Based Kth Prediction . . . . . . . . . . . 115
4.1.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
4.2 A Novel Framework for DTCO: Fast and Automatic Routability Assessment
with Machine Learning for Sub-3nm Technology Options . . . . . . . . . . . . . . . . . . . 136
4.2.1 DTCO Framework and Configuration for Sub-3nm Nodes . . . . . . . . . . . . 137
4.2.2 Routability Assessment and Block-Level Area Case Study . . . . . . . . . . . . 139
4.2.3 Machine Learning (ML)-Assisted Routability Assessment . . . . . . . . . . . . 143
4.2.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
4.3 PROBE3.0: A Systematic Framework for Design-Technology Pathfinding with
Improved Design Enablement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
4.3.1 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
4.3.2 Standard-Cell Library and PDK Generation . . . . . . . . . . . . . . . . . . . . . . . . 148
4.3.3 Power Delivery Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
4.3.4 Enhanced Artificial Designs for PPAC Exploration . . . . . . . . . . . . . . . . . . 159
4.3.5 Cell Width-Regularized Placements for More Realistic Routability As-
sessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
4.3.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
Chapter 5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
vi
LIST OF FIGURES
Figure 1.1. The scaling rates of fin pitch, contacted poly pitch (CPP) and standard-
cell height have slowed down over time, and device architectures have
also evolved from fin field-effect transistor (FinFET) to complementary
field-effect transistor (CFET) [104]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
Figure 1.2. Power and performance requirements for applications [89]. . . . . . . . . . . . . . 3
Figure 1.3. Local layout effect [128]. (a) Saturation voltage and effective current shifts
caused by the second diffusion break effect. (b) Interlayer dielectric (ILD)
stress in a TCAD simulation, which is a function of the spacing between
the first and the second diffusion breaks. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
Figure 1.4. Resistance of BEOL metal layers (Mx) increases exponentially as technol-
ogy advances [89]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
Figure 1.5. Overview of this thesis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
Figure 2.1. An example floorplan view of a memory-dominant chip [175]. The red
boxes indicate memory components in the chip. . . . . . . . . . . . . . . . . . . . . . . 11
Figure 2.2. “Cartoon” of floorplan and clock tree structures for a memory-dominant
SOC design, which motivates our k-active dynamic power minimization
formulation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
Figure 2.3. Terminals with uniform and non-uniform grids on an example floorplan.
(a) Terminals with uniform grids. (b) Terminals with non-uniform grids
based on channel intersection graph. (c) Terminals with non-uniform grids
after the unusable grids. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
Figure 2.4. Example trees for flexible clock sources. (a) A tree with a fixed clock
source. (b) A tree with a movable clock source, with optimal source
location. (c) A tree with the tree from (b) and the original source location
from (a). (d) A tree with a movable clock source that has a forbidden area. 25
Figure 2.5. An example set of non-dominated trees for a 12-terminal pointset. The
weighting factors and bounds are shown as (α, β , L, B). (a) (0, 0, 5,
0.6 · M). (b) (0, 100, 4, 0.6 · M). (c) (0, 100, 5, 0.6 · M). (d) (100, 0, 7,
0.4 · M). (e) (0, 100, 7, 0.4 · M). (f) (0, 100, 7, 0.2 · M). . . . . . . . . . . . . . . . . 28
Figure 2.6. Example trees with various weighting factors in our objective function. (a)
A tree solution to minimize the total wirelength. (b) A tree solution to
minimize PMax(1). (c) A tree solution to minimize PMax(2). . . . . . . . . . . 29
vii
Figure 2.7. Scatter plots for (a) skew-wirelength, (b) PMax(1)-wirelength and (c)
PMax(2)-wirelength for the 12-terminal pointset with various weighting
factors α and β . Additional points that do not correspond to the figure
legend are obtainable with other values of α, β , L and B. . . . . . . . . . . . . . . . 30
PMax(2)-wirelength for k-active sinks and k-consecutive-active sinks. . . . 32
PMax(2)-wirelength for fixed and flexible clock source locations, as well
as flexible clock source locations with a forbidden 2 ≤ x ≤ 4, 2 ≤ y ≤ 4 area. 32
Figure 2.10. Framework overview. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
Figure 2.11. Grid-based place-and-route architecture. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
Figure 2.12. Placement adjustments. (a) Initial placement. (b) Horizontal shifting. (c)
Horizontal flipping. (d) Cell swapping. (e) Vertical shifting. . . . . . . . . . . . . 40
Figure 2.13. Pinlength extension. (a) Initial placement. (b) Pinlength extension. . . . . . . 41
Figure 2.14. Switchbox generation. (a) Local window with one DRC, displayed by the
commercial tool [144]. (b) Visualization of the generated switchbox. . . . . . 42
Figure 2.15. Placement constraints. (a) Relative positions between two instances in the
same placement row. (b) Boundary condition of each placement row. . . . . 45
Figure 2.16. Dynamic cell allocation (DCA). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
Figure 2.17. Weighted extendable pin candidates. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
Figure 2.18. Grid-based conditional design rules: (a) MAR, (b) EOL and (c) VR. . . . . . 48
Figure 2.19. Overall ECO flow using CoRe-ECO framework. . . . . . . . . . . . . . . . . . . . . . . 51
Figure 2.20. Example trends of the number of DRCs by the ECO routing and the
proposed CoRe-ECO iterations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
Figure 2.21. DRC reductions by CoRe-ECO rounds for AES (4T/ET/MPO2). . . . . . . . . 56
Figure 2.22. Example of refinement operations in the proposed ECO flow (Index 38
case in Table 2.7). (a) Switchbox with three DRCs. (b) Routable solution
with placement adjustments and pinlength extensions. . . . . . . . . . . . . . . . . . 57
viii
Figure 3.1. Examples illustrating DB types and placement constraints. (a) Layout with
four SDB cells (Type-I). (b) Mixed-DB layout with three SDB cells plus
one DDB cell (Type-II). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
Figure 3.2. Target of our work. We seek to minimize actual, 2nd DB-aware leakage
power with no timing degradation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
Figure 3.3. Overview of our overall optimization flow. The red box indicates steps that
we implement. A commercial P&R tool is used for all other steps in the
flow. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
Figure 3.4. Normalized derating values for (a) leakage power and (b) cell delay accord-
ing to the 2nd DB distance. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
Figure 3.5. Examples of 2nd DB distance change due to cell relocation and sizing. The
red color of D2C2 indicates the 2nd DB distance of C2 and the green color
of D2C1 indicates the 2nd DB distance of C1. . . . . . . . . . . . . . . . . . . . . . . . . . 67
Figure 3.6. Runtime analysis for AES design. We use Type-II, 60% initial utilization
and 0.5ns clock period. 1st loop denotes the first iteration of our optimiza-
tion and 2nd loop denotes the second iteration. Sizing includes sensitivity
calculations and cell moves for gate sizing, Vt swapping and DB swapping. 74
Figure 3.7. Studies of leakage power with varying target clock period and initial
utilization. Leakage increment from the 2nd DB effect vs. (a) target clock
period and (b) initial utilization. Leakage recovery by our optimization vs.
(c) target clock period and (d) initial utilization. . . . . . . . . . . . . . . . . . . . . . . 76
Figure 3.8. Illustration of power staple insertions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
Figure 3.9. The case represented by D[i][1][4][2][2]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
Figure 3.10. Overall flow of our work. The red box indicates steps that we implement.
Commercial tools [144] [148] are used for all other steps. . . . . . . . . . . . . . . 87
Figure 3.11. (a) Sensitivity of #staples to displacement range x∆ and (b) sensitivity of
runtime to displacement range (x∆ ). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
Figure 3.12. Sensitivities of (a) #staples and (b) total cell displacement, to the displace-
ment factor (α). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
Figure 3.13. Design examples with different lengths of power staples. (a) Power staple
length = 2 row heights. (b) Power staple length = 10 row heights. . . . . . . . . 91
ix
Figure 3.14. (a) Number of staples (total, VDD and VSS) that can be placed, versus
balance factor (β ) (AES with utilization 0.60) and (b) number of staples
that can be placed, versus utilization by different strategies (AES with
utilization between 0.60 and 0.75). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
Figure 3.15. Heatmaps of vectorless dynamic IR drop for AES. . . . . . . . . . . . . . . . . . . . . 95
Figure 4.1. The DTCO process consists of three main stages: Technology, Design
Enablement, and Design. Figure is redrawn from [174]. . . . . . . . . . . . . . . . . 99
Figure 4.2. Overall flows for PROBE [64] and PROBE2.0. (a) PROBE only evaluates
BEOL stack options. (b) The PROBE2.0 flow includes standard-cell layout
generation [26] [82], design enablement and routability assessment. (c)
The PROBE2.0 flow uses a trained learning-based model to predict Kth . . . 105
Figure 4.3. Placements for PROBE [64]. (a) A mesh-like placement based on a 2-input
cell. (b) A cell width-regularized placement. The orange-striped cells are
considered to be neighbors of the blue-striped cell. The blue-striped cell is
swapped with a randomly-selected neighbor. . . . . . . . . . . . . . . . . . . . . . . . . . 106
Figure 4.4. A grid-based standard-cell architecture and technology parameters for

standard cells. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
Figure 4.5. DR-MAR, DR-EOL and DR-VR design rules. . . . . . . . . . . . . . . . . . . . . . . . . 109
Figure 4.6. Example of pin shape selection in a standard cell (2Fin, 5RT, 2MPO and
EUV-Loose). (a) Initial layout of OAI21 X1 from SMT-based layout
generation. (b) The OAI21 X1 layout after RPA-based pin shape selection. 114
Figure 4.7. Example of standard-cell layout with top-metal-only pin shape selection.
We show an initial layout of AOI21 X1 (2Fin, 5RT, 2MPO and EUV-
Loose). In this case, the ZN pin has only the M2 pin shape, and the M1
and V1 shapes become OBS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
Figure 4.8. Power and ground pin (PGpin) examples: (a) BPR, (b) M1 and (c) M1+M2.
The signal pin shapes (gray) are the same regardless of PGpin. . . . . . . . . . 115
Figure 4.9. Example of knight’s tour-based connections in a placement of 2-input cells.

Numbers in red circles indicate the knight’s tour ordering. . . . . . . . . . . . . . 117
Figure 4.10. Kth results from knight’s tour-based cell-level routability assessments
across various track heights and cell masters. 15 cell masters from 6T, 7T
and 8T standard-cell libraries are used in this experiment. . . . . . . . . . . . . . . 124
x
Figure 4.11. Cell-level routability assessments showing #DRCs versus K with different
pin shape selection methods. RPA-Best (B) and RPA-Worst (W) pin shape
selection methods are shown. Solid lines are for cells with RPA-Best, and
dashed lines are for cells with RPA-Worst. . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
Figure 4.12. Design-level assessments with various technology parameters. . . . . . . . . . . 127
Figure 4.13. Kth results of design-level routability assessments with various design
parameters. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
Figure 4.14. Comparison between Kth and achievable utilization for the AES and LDPC
designs. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
Figure 4.15. Comparison between golden Kth and predicted Kth and extracted feature
importance; (a) training data, (b) testing data, (c) error distribution on
testing data with kernel density estimation (KDE) plot and (d) extracted
feature importance. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
Figure 4.16. (a) Overall flow of PROBE2.0 [24] and (b) machine learning-based Kth
prediction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
Figure 4.17. An example of neighbor-swap operation. The blue-colored cell is swapped

with one of the red-colored neighboring cells. . . . . . . . . . . . . . . . . . . . . . . . . 138
Figure 4.18. Standard-cell layout of the nine libraries in Table 4.16. OAI21 X1 of Lib1
to Lib9 are illustrated in (a) to (i), respectively. We show a LEF view (cell
boundary and pin) of layout of OAI21 X1 per library. . . . . . . . . . . . . . . . . . 139
Figure 4.19. Results for cell-level routability assessments. (a) #DRC vs. Kth plot for
NAND2 X1 and NAND3 X1. (b) #DRC vs. Kth plot for OAI21 X1 and
OAI22 X1. (c) Kth values for 15 cell types per four standard-cell libraries. 140
Figure 4.20. Design-level routability assessment for the nine standard-cell libraries. (a)
#DRC vs. Kth plot for AES design. (b) Kth values for AES, JPEG and
VGA designs. We use 0.6 Util for AES, JPEG and VGA and 0.15 Util for
LDPC design. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
Figure 4.21. Achievable density and block-area case study. (a) Plot for Kth vs. achiev-
able utilization for AES and LDPC designs. (b) Achievable block-area
study with the nine libraries and AES design. The arrows show block-level
area and cell height differences in the case study, as also shown in Table 4.17. 141
Figure 4.22. Comparison between golden Kth and predicted Kth . (a) Results for 128
training data sampled by LHS. (b) Results for 320 testing data. (c) Error
distribution on the testing dataset with kernel density estimation (KDE) plot. 143
xi
Figure 4.23. Scope of PROBE. PROBE1.0 [64] and PROBE2.0 [24] (Section 4.1 of this
thesis) address AC given BEOL and FEOL/BEOL, respectively. PROBE-
3nm [27] (Section 4.2 of this thesis) studies routability with sub-3nm
configurations. PROBE3.0 provides true full-stack PPAC pathfinding. . . . 146
Figure 4.24. Automatic standard-cell library and PDK generation (Design Enablement)
in the PROBE3.0 framework. In addition to technology and design pa-
rameters in the PROBE3.0 framework, other technology-related inputs are
required. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
Figure 4.25. Example standard cells (AND2 X1). The cells are generated by our
standard-cell layout generation with the following parameters (Fin, RT,
PGpin, CH): (a) Lib1 (2Fin, 4RT, BPR, 5T), (b) Lib2 (2Fin, 4RT, M0, 6T),
(c) Lib3 (3Fin, 5RT, BPR, 6T) and (d) Lib4 (3Fin, 5RT, M0, 7T). . . . . . . . . 153
Figure 4.26. Cross-section view of four PDN options in the PROBE3.0 framework: (a)
Frontside PDN (PFS ), (b) Frontside PDN with BPR (PFB ), (c) Backside
PDN (PBS ) and (d) Backside PDN with BPR (PBB ). . . . . . . . . . . . . . . . . . . . . 155
Figure 4.27. Power tap cells for (a) PFB and (b) PBS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
Figure 4.28. Four power tap cell insertion results: (a) Power tap cells for PFB (2CPP
width) with Column; (b) power tap cells for PFB with Staggered; (c) power
tap cells for PBS (6CPP width) with Column; and (d) power tap cells for
PBS with Staggered. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
Figure 4.29. IR drop analysis flow for (a) FSPDN and (b) BSPDN. For the IR drop flow
for BSPDN, we delete all the signal and clock routing after P&R and build
power stripes for BSPDN. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
Figure 4.30. ML-based parameter tuning for ANG. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
Figure 4.31. Comparison between real and artificial designs by t-SNE [91]. (a) t-SNE
visualization for real and artificial (ANG, design names followed by ∗)
designs without our parameter tuning flow. (b) Real and artificial (ANG)
designs with our ML-based parameter tuning flow. . . . . . . . . . . . . . . . . . . . 166
Figure 4.32. Two example clustered cells NAND X1 AND X1 and INV X1 OR X1,
in clustering-based cell width-regularization. (a) Schematic view, and (b)
physical layout view assuming Lib2. Here, the maximum clustered cell
width wmax is 12CPP. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
Figure 4.33. Cell width distributions pre-clustering (i.e., original netlist) and post-
clustering (i.e., by CWR-FC) for (a) AES, (b) JPEG, (c) LDPC and (d)
VGA. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
xii
Figure 4.34. PPAC tradeoffs for JPEG with four standard-cell libraries (Lib1-4) and four
PDN structures. We measure improvements relative to traditional PDN
(PFS ), to show the benefits of BSPDN and BPR: (a) Performance-Power,
(b) Performance-Area and (c) Energy-Delay Product-Area. . . . . . . . . . . . . . 177
Figure 4.35. IR drop-Area plots for JPEG with four standard-cell libraries (Lib1, Lib2,
Lib3 and Lib4). (a) JPEG with 4RT (Lib1/2). (b) JPEG with 5RT (Lib3/4). 178
Figure 4.36. PPAC tradeoffs for “artificial” JPEG with four standard-cell libraries (Lib1,
Lib2, Lib3 and Lib4). (a) Performance-Power. (b) Performance-Area. (c)
Energy-Delay Product-Area. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
Figure 4.37. IR drop-Area plots for “artificial” JPEG with four standard-cell libraries
(Lib1, Lib2, Lib3 and Lib4). (a) JPEG with 4RT (Lib1/2). (b) JPEG with
5RT (Lib3/4). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
Figure 4.38. Kth and achievable utilization for (a) AES and (b) JPEG, with various
libraries and power delivery methodologies. . . . . . . . . . . . . . . . . . . . . . . . . . . 182
xiii
LIST OF TABLES
Table 2.1. Notations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
Table 2.2. Average ILP runtime when α = 1, β = 0.5, L = M − B. M denotes the

maximum distance between source and sink. . . . . . . . . . . . . . . . . . . . . . . . . . . 33
Table 2.3. ILP runtime for fixed and flexible clock source locations in Figure 2.4 and
non-uniform grids. We set variables α = 1, β = 0.5, L = M − B. M denotes
the maximum distance between source and sink. . . . . . . . . . . . . . . . . . . . . . . . 34
Table 2.4. Notations for the proposed CoRe-ECO framework. . . . . . . . . . . . . . . . . . . . . 38
Table 2.5. Standard-cell architectures. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
Table 2.6. Experimental statistics. RT = the number of routing tracks, DRSet =

Design rule set, Impr./Incr. = Improvement/Increment ratio over ECO
routing, #V/#H/#F/#P = the number of refined instances by vertical shift-
ing/horizontal shifting/horizontal flipping/pinlength extension. . . . . . . . . . . 50
Table 2.7. Detailed experimental results of the proposed ECO flow (AES, 4T, ET,
MPO2, 74 DRCs). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
Table 3.1. Preliminary study of the 2nd DB effect. Design nomenclature specifies
design name (AES or VGA), filler insertion strategy (Small (sm) or Standard
(std)), and design type (Type-I or Type-II). WS denotes the worst slack of a
design. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
Table 3.2. Notations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
Table 3.3. Experimental results for all the design blocks. . . . . . . . . . . . . . . . . . . . . . . . . 71
Table 3.4. Experimental results for design space exploration. We use AES Type-II
design with 60% initial utilization and 0.5ns target clock period. . . . . . . . . . 75
Table 3.5. Experimental results for sensitivity to the tolerance. The 2nd DB-aware leak-
age power of the design is 0.190mW. Reloc., 1st loop and 2nd loop denote
runtime of the relocation step, and of the first and the second iterations of
the optimization, respectively. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
Table 3.6. Notations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
Table 3.7. Design information. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
xiv
Table 3.8. Total #staples placed when we apply two rounds of optimization with
total displacement range (x∆ ) = 6. Each row represents design AES with
utilization between 0.60 and 0.75. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
Table 3.9. IR drop vs. staple length. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
Table 3.10. Experimental results using double-row optimization. . . . . . . . . . . . . . . . . . . . 93
Table 4.1. Standard-cell architectures in our experiments. . . . . . . . . . . . . . . . . . . . . . . . . 109
Table 4.2. Four BEOL stack options. R denotes routing resource. . . . . . . . . . . . . . . . . . 111
Table 4.3. Details of PDN options. All numbers are pitches in units of µm for each
layer. “P” indicates that we use a given layer only for PDN at maximum
area density, and do not allow the layer to be used for signal routing. The
width of M5 stripes is 0.96µm and the width of M6/M8 stripes is 1.296µm. 111
Table 4.4. Technology and design parameters in our experiments. . . . . . . . . . . . . . . . . . 112
Table 4.5. Placed design information for AES, mesh-like placements and knight’s tour
based topologies. Mesh-like placements and knight’s tours have 115 × 115
placed instances. Area denotes total instance area in the designs. Area is in
units of µm2 and wirelength (WL) is in units of µm. . . . . . . . . . . . . . . . . . . . 118
Table 4.6. Rent parameters for mesh-like and knight’s tour-based placed netlists. We
use rectangle sampling-based methods I and II for the evaluation method,
Types I and II for the pin counting method, and geometric mean for the
averaging method in [165]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
Table 4.7. Input features of the Kth prediction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
Table 4.8. Default parameters in our experiments. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
Table 4.9. List of 39 standard cells per generated library. . . . . . . . . . . . . . . . . . . . . . . . . . 123
Table 4.10. Kth results with RPA-Best and RPA-Worst pin shape selection methods. This
table shows the Kth values corresponding to the data shown in Figure 4.11. 126
Table 4.11. Kth results with various Fin, MP, RT, PGpin and MPO. . . . . . . . . . . . . . . . . . 128
Table 4.12. Results on training data of Top-5 ML models by using AutoML. The three
GBM models (GBM 1, GBM 2 and GBM 3) have different configurations
for hyperparameters. The training runtime is limited to one hour with 80
CPU threads. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
xv
Table 4.13. Results of DL modeling with AutoML. DL 1h (resp. DL 24h) denotes a DL
model with a 1-hour (resp. 24-hour) runtime limit for training. . . . . . . . . . . . 132
Table 4.14. Model accuracy results with different amounts of training data (TD), i.e., 20,
40, 60 and 80% of the total data (2460 Kth values), with a one-hour limit and
80 threads. Testing data are fixed for an apples-to-apples comparison. ME
and AE respectively denote maximum and average error of Kth prediction. 132
Table 4.15. Technology and design parameters in this work. We use all the parameter
types defined in [24]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
Table 4.16. Technology parameters for the nine standard-cell libraries which we study
in this work. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
Table 4.17. Results for achievable block-level area case study, expressed as overhead of
Lib B relative to Lib A, i.e., (AreaB − AreaA )/AreaA × 100% . . . . . . . . . . . . . 142
Table 4.18. Layer definition in the PROBE3.0 technology. . . . . . . . . . . . . . . . . . . . . . . . . 150
Table 4.19. Key features of the PROBE3.0 technology. . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
Table 4.20. List of 41 standard cells per generated library. . . . . . . . . . . . . . . . . . . . . . . . . . 152
Table 4.21. PDN configurations for FSPDN and BSPDN. A pair of power (VDD) and
ground (VSS) stripes are placed every pitch, while maintaining the spacing
between VDD and VSS. Density denotes the percentage of routing tracks
occupied by PDN. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
Table 4.22. Definition of topological parameters in ANG [78] [140]. . . . . . . . . . . . . . . . . 160
Table 4.23. Topological parameters for real netlists from OpenCores [161] and artificial
netlists generated by [78]. Design names followed by ∗ indicate ANG-
generated artificial netlists. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
Table 4.24. Parameter sets for training and testing. We train our ML model with ANG
input parameters and post-P&R output parameters. The total number of
datapoints is 4 × 6 × 6 × 5 × 5 × 6 = 21600. Testing is performed in the
ranges around given target parameters, according to the step sizes. . . . . . . . . 163
Table 4.25. Topological parameters for target, input and output netlists. The design
names followed by ∗∗ indicate ANG-generated artificial netlists with ML-
based ANG parameter tuning. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
Table 4.26. Comparison of the width-regularized clustered netlist produced by CWR-FC

([C]) with the original flat netlist ([A]) and a width-regularized clustered
netlist induced from a placement of the flat netlist ([B]). . . . . . . . . . . . . . . . . 172
xvi
Table 4.27. Technology and design parameters in our experiments. . . . . . . . . . . . . . . . . . 174
Table 4.28. Tools and versions in our experiments. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
Table 4.29. Kth comparison for the JPEG design with cell width-regularized placements
([A]) and clustering-based cell width-regularized placements ([C]). Util
denotes real utilization with 0.6 initial utilization. . . . . . . . . . . . . . . . . . . . . . 183
xvii
ACKNOWLEDGMENTS
I would like to express my gratitude to my advisor, Professor Andrew B. Kahng, for
serving as the chair of my committee and providing invaluable guidance throughout my Ph.D.
journey.
I also want to thank my fellow labmates at the UCSD VLSI CAD Laboratory, including
Nikhil Dakwala, Sayak Kundu, Jiantao Liu, Bodhisatta Pramanik, Ravi Varadarajan, Zhiang
Wang, Mingyu Woo, Chien-Yi Yang and Dooseok Yoon, as well as former lab members, including
Payal Agarwal, Dr. Wei-Ting (Jonas) Chan, Dr. Ahmed Taha Elthakeb, Dr. Kwangsoo Han, Dr.
Chia-Tung (Mark) Ho, Dr. Seungwon Kim, Dr. Hyein Lee, Hsin-Yu (Desmond) Liu, Dr. Uday
Mallappa, Tushar Shah, Siddharth Singh, Dr. Vaishnav Srinivas, Sriram Venkatesh, Dr. Lutong
Wang and Dr. Bangqi Xu, for their assistance and thought-provoking discussions. Special thanks
go to Dr. Tuck-Boon Chan, Dr. Ilgweon Kang, Dr. Daeyeal Lee, Dr. Jiajia Li and Dr. Dongwon
Park for their guidance during my Ph.D. studies.
I am also grateful to my thesis committee members, Professor Chung-Kuan Cheng,
Professor Sujit Dey and Professor Farinaz Koushanfar, for their time, encouragement and
insightful comments.
I would like to acknowledge the invaluable feedback and support of my industrial
collaborators, including Suhyeong Choi, Dr. Hamed Fatemi, Dr. Jose Pineda de Gyvez, Sunik
Heo, Dr. Jinwook Jung, Chunghee Kim, Hayoung Kim, Dr. Jaehwan Kim, Dr. Gi-Joon Nam and
Dr. S. C. Song.
I extend my gratitude to my old fellows from South Korea, Jinhwan Choi, Sangha Chung,
Sangyeop Kim, Youngmin Kim, Seungheon Lee and Byeong-Geun Sun, as well as my beloved
companions, Hyunwoo Baek, Woojae Park, Eunjae Shin and Sooyoung Son.
Finally, I would like to express my heartfelt appreciation to my parents, Hongjin Kim and
Jungsuk Kang, my mother-in-law, Nankyung Kim, and the rest of my family for their endless
love and encouragement. I owe a special debt of gratitude to my wife, Youri Jeong, whose
unwavering love and support enabled me to complete this journey, and my lovely daughters, Joy
xviii
and Skylar, who gave me the strength I needed every day with their love. This thesis is dedicated
to my family.
The material in this thesis is based on the following publications.
Chapter 2 contains a reprint of Hamed Fatemi, Andrew B. Kahng, Minsoo Kim and
Jose Pineda de Gyvez, “Optimal Bounded-Skew Steiner Trees to Minimize Maximum k-Active
Dynamic Power”, Proc. International Workshop on System-Level Interconnect Problems and
Pathfinding, 2020; and Chung-Kuan Cheng, Andrew B. Kahng, Ilgweon Kang, Minsoo Kim,
Daeyeal Lee, Bill Lin, Dongwon Park and Mingyu Woo, “CoRe-ECO: Concurrent Refinement
of Detailed Place-and-Route for an Efficient ECO Automation”, Proc. IEEE International
Conference on Computer Design, 2021. The dissertation author is a main contributor to, and a
primary author of, each of these papers.
Chapter 3 contains a reprint of Sunik Heo, Andrew B. Kahng, Minsoo Kim and Lutong
Wang, “Diffusion Break-Aware Leakage Power Optimization and Detailed Placement in Sub-
10nm VLSI”, Proc. Asia and South Pacific Design Automation Conference, 2019; and Sunik Heo,
Andrew B. Kahng, Minsoo Kim and Lutong Wang and Chutong Yang, “Detailed Placement for
IR Drop Mitigation by Power Staple Insertion in Sub-10nm VLSI”, Proc. Design, Automation
and Test in Europe, 2019. The dissertation author is a main contributor to, and a primary author
of, each of these papers.
Chapter 4 contains a reprint of Chung-Kuan Cheng, Andrew B. Kahng, Hayoung Kim,
Minsoo Kim, Daeyeal Lee, Dongwon Park and Mingyu Woo, “PROBE2.0: A Systematic
Framework for Routability Assessment from Technology to Design in Advanced Nodes”, IEEE
Transactions on Computer-Aided Design of Integrated Circuits and System 41(5), 2022; Chidi
Chidambaram, Andrew B. Kahng, Minsoo Kim, Giri Nallapati, S. C. Song and Mingyu Woo,
“A Novel Framework for DTCO: Fast and Automatic Routability Assessment with Machine
Learning for Sub-3nm Technology Options”, Proc. IEEE Symposium on VLSI Technology,
2021; and Suhyeong Choi, Jinwook Jung, Andrew B. Kahng, Minsoo Kim, Chul-Hong Park,
Bodhisatta Pramanik and Dooseok Yoon, “PROBE3.0: A Systematic Framework for Design-
xix
Technology Pathfinding with Improved Design Enablement”, in submission to IEEE Transactions
on Computer-Aided Design of Integrated Circuits and Systems, 2023. The dissertation author is
a main contributor to, and a primary author of, each of these papers.
All co-authors, listed in alphabetical order, including Professor Chung-Kuan Cheng,
Dr. Chidi Chidambaram, Suhyeong Choi, Dr. Hamed Fatemi, Sunik Heo, Dr. Jinwook Jung,
Professor Andrew B. Kahng, Dr. Ilgweon Kang, Hayoung Kim, Dr. Daeyeal Lee, Professor Bill
Lin, Dr. Giri Nallapati, Dr. Chul-Hong Park, Dr. Dongwon Park, Dr. Jose Pineda de Gyvez,
Bodhisatta Pramanik, Dr. S. C. Song, Dr. Lutong Wang, Mingyu Woo, Chutong Yang and
Dooseok Yoon, have kindly approved the inclusion of the aforementioned publications in my
thesis.
xx
VITA
2011 B. S., Electrical Engineering,

Yonsei University, Seoul, South Korea
2013 M. S., Electrical Engineering,
Korea Advanced Institute of Science and Technology, Daejeon, South Korea
2013–2017 Engineer, Samsung Electronics, Gyeonggi-do, South Korea
2021 C. Phil., Electrical Engineering (Computer Engineering),
2023 Ph. D., Electrical Engineering (Computer Engineering),
All papers coauthored with my advisor Professor Andrew B. Kahng have authors listed in
alphabetical order.
PUBLICATIONS
Suhyeong Choi, Jinwook Jung, Andrew B. Kahng, Minsoo Kim, Chul-Hong Park, Bodhisatta
Pramanik and Dooseok Yoon, “PROBE3.0: A Systematic Framework for Design-Technology
Pathfinding with Improved Design Enablement”, in submission to IEEE Transactions on
Computer-Aided Design of Integrated Circuits and Systems.
Chia-Tung Ho, Alvin Ho, Matthew Fojtik, Minsoo Kim, Shang Wei, Yaguang Li, Brucek
Khailany and Haoxing Ren, “NVCell 2: Routability-Driven Standard Cell Layout in Advanced
Nodes with Lattice Graph Routability Model”, Proc. International Symposium on Physical
Design, 2023, pp. 44-52.
Chung-Kuan Cheng, Andrew B. Kahng, Ilgweon Kang, Minsoo Kim, Daeyeal Lee, Bill Lin,
Dongwon Park and Mingyu Woo, “PROBE2.0: A Systematic Framework for Routability Assess-
ment from Technology to Design in Advanced Nodes”, IEEE Transactions on Computer-Aided
Design of Integrated Circuits and System 41(5) (2022), pp. 1495-1508.
Andrew B. Kahng, Minsoo Kim, Seungwon Kim and Mingyu Woo, “RosettaStone: Connecting
the Past, Present and Future of Physical Design Research”, IEEE Design & Test 39(5) (2022), pp.
70-78.
Chung-Kuan Cheng, Andrew B. Kahng, Ilgweon Kang, Minsoo Kim, Daeyeal Lee, Bill Lin,
Dongwon Park and Mingyu Woo, “CoRe-ECO: Concurrent Refinement of Detailed Place-and-
Route for an Efficient ECO Automation”, Proc. IEEE International Conference on Computer
Design, 2021, pp. 366-373.
xxi
Chidi Chidambaram, Andrew B. Kahng, Minsoo Kim, Giri Nallapati, S. C. Song and Mingyu
Woo, “A Novel Framework for DTCO: Fast and Automatic Routability Assessment with Machine
Learning for Sub-3nm Technology Options”, Proc. IEEE Symposium on VLSI Technology, 2021,
pp. 1-2.
Hamed Fatemi, Andrew B. Kahng, Minsoo Kim and Jose Pineda de Gyvez, “Optimal Bounded-
Skew Steiner Trees to Minimize Maximum k-Active Dynamic Power”, Proc. ACM/IEEE
International Workshop on System-Level Interconnect Problems and Pathfinding, 2020, pp. 1-8.
Austin Rovinski, Tutu Ajayi, Minsoo Kim, Guanru Wang and Mehdi Saligane, “Bridging Aca-
demic Open-Source EDA to Real World Usability”, Proc. IEEE/ACM International Conference
on Computer-Aided Design, 2020, pp. 1-7.
Vidya A. Chhabria, Andrew B. Kahng, Minsoo Kim, Uday Mallappa, Sachin S. Sapatnekar
and Bangqi Xu, “Template-based PDN Synthesis in Floorplan and Placement Using Classifier
and CNN Techniques”, Proc. Asia and South Pacific Design Automation Conference, 2020, pp.
44-49.
Tutu Ajayi, Vidya A. Chhabria, Mateus Fogaça, Soheil Hashemi, Abdelrahman Hosny, Andrew B.
Kahng, Minsoo Kim, Jeongsup Lee, Uday Mallappa, Marina Neseem, Geraldo Pradipta, Sherief
Reda, Mehdi Saligane, Sachin S. Sapatnekar, Carl Sechen, Mohamed Shalan, William Swartz,
Lutong Wang, Zhehong Wang, Mingyu Woo and Bangqi Xu, “Toward an Open-Source Digital
Flow: First Learnings from the OpenROAD Project”, Proc. ACM/IEEE Design Automation
Conference, 2019, pp. 76:1-76:4.
Tutu Ajayi, David Blaauw, Tuck-Boon Chan, Chung-Kuan Cheng, Vidya A. Chhabria, David
K. Choo, Matteo Coltella, Sorin Dobre, Ronald Dreslinski, Mateus Fogaça, Soheil Hashemi,
Abdelrahman Hosny, Andrew B. Kahng, Minsoo Kim, Jiajia Li, Zhaoxin Liang, Uday Mallappa,
Paul Penzes, Geraldo Pradipta, Sherief Reda, Austin Rovinski, Kambiz Samadi, Sachin S. Sapat-
nekar, Lawrence Saul, Carl Sechen, Vaishnav Srinivas, William Swartz, Dennis Sylvester, David
Urquhart, Lutong Wang, Mingyu Woo and Bangqi Xu, “OpenROAD: Toward a Self- Driving,
Open-Source Digital Layout Implementation Tool Chain”, Proc. Government Microcircuit
Applications and Critical Technology Conference, 2019, pp. 1105-1110.
Sunik Heo, Andrew B. Kahng, Minsoo Kim, Lutong Wang and Chutong Yang, “Detailed
Placement for IR Drop Mitigation by Power Staple Insertion in Sub-10nm VLSI”, Proc. Design,
Automation & Test in Europe Conference & Exhibition, 2019, pp. 824-829.
Sunik Heo, Andrew B. Kahng, Minsoo Kim and Lutong Wang, “Diffusion Break-Aware Leakage
Power Optimization and Detailed Placement in Sub-10nm VLSI”, Proc. Asia and South Pacific
Design Automation Conference, 2019, pp. 550-556.
xxii
Minsoo Kim, Chong-Min Kyung and Kang Yi, “An Energy Management Scheme for Solar-
Powered Wireless Visual Sensor Networks Toward Uninterrupted Operations”, Proc. IEEE
International SoC Design Conference, 2013, pp. 23-26.
xxiii
ABSTRACT OF THE DISSERTATION
Robust Physical Design and Design Technology Co-Optimization Methodologies at

Advanced VLSI Technology
by
Minsoo Kim
Doctor of Philosophy in Electrical Engineering (Computer Engineering)
University of California San Diego, 2023
Professor Andrew B. Kahng, Chair
The semiconductor industry has achieved remarkable progress by adhering to Moore’s

Law in the past few decades. As a result, technology has continuously scaled down and advanced
to the 2nm and 3nm nodes by 2023. The consistent scaling of advanced technologies has made it
possible to utilize them in various applications of modern IC designs, such as mobile, data center,
automotive, graphics, the Internet of Things (IoT) and artificial intelligence (AI), which may
demand high performance and/or ultra-low power consumption. However, the recent slowdown
in the traditional Moore’s Law scaling rate has presented significant challenges. Therefore,
considerable efforts have been devoted to physical design and design-technology co-optimization
xxiv
to optimize the advantages of advanced technology nodes for different applications.
This thesis presents robust physical design and design-technology co-optimization
methodologies that aim to maximize the benefits of advanced technologies and optimize power,
performance, area and cost in modern IC design. The proposed methodologies are categorized
into three main directions: (i) general physical design methodologies, (ii) technology-aware
physical design methodologies and (iii) design-technology co-optimization methodologies.
To address challenges in modern IC design, this thesis presents two works: (i) bounded-
skew Steiner tree optimization for clock tree synthesis to minimize active power and (ii) concur-
rent refinement of detailed place-and-route (P&R) for efficient engineering change order (ECO)
automation.
To address challenges specific to advanced technology nodes, this thesis presents two
works: (i) leakage power optimization with the awareness of local layout effects and (ii) detailed
placement for IR drop mitigation by power staple insertions.
Finally, to address challenges in design-technology co-optimization at advanced tech-
nology nodes, this thesis presents three works: (i) PROBE2.0: A systematic framework for
routability assessments, (ii) a routability study using the PROBE2.0 framework with 3nm tech-
nology configurations and (iii) PROBE3.0: A systematic framework for power, performance,
area and cost explorations, with improved design enablement.
xxv
Chapter 1
Introduction
In the last several decades, the semiconductor industry has successfully followed Moore’s
Law as a guide for the long-term planning of advanced technology development. As a result,
semiconductor technologies have consistently scaled down to the 2nm and 3nm nodes, with the
number of transistors in a single chip now exceeding tens of billions due to aggressive transistor
density scaling. These advanced technologies bring various capabilities to applications, including
mobile, data center, automotive, graphics, the Internet of Things (IoT) and artificial intelligence,
with significantly improved performance and/or reduced power consumption.
1.1 New Challenges
Despite the semiconductor industry’s past success, the trend of Moore’s Law scaling has
recently slowed down. Figure 1.1 illustrates this slowdown, where the fin pitch, contacted poly
pitch (CPP) and standard-cell height scaling rates have flattened over time. Consequently,
the slowdown has presented numerous challenges across many applications, and both the
semiconductor industry and academia have made significant efforts to mitigate the slowdown
and address the challenges posed by various applications and advanced technology nodes.
1.1.1 Challenges for Modern IC Design
In modern Integrated Circuit (IC) designs, various applications target different power
and performance levels. Figure 1.2 illustrates the major applications in modern IC design and
1
Figure 1.1. The scaling rates of fin pitch, contacted poly pitch (CPP) and standard-cell height
have slowed down over time, and device architectures have also evolved from fin field-effect
transistor (FinFET) to complementary field-effect transistor (CFET) [104].
their general power and performance requirements. High-performance computing demands high
performance with high power consumption, while IoT aims to achieve long operational lifetimes
with ultra-low power consumption. Across these applications, the ultimate goal of modern IC
design is to optimize power, performance, area and cost (PPAC) within a reasonable turnaround
time. However, as the target PPAC is dependent on the application and design, optimizing for
various applications can be challenging in modern IC design.
Aggressive PPAC optimizations introduce further challenges in modern IC design. One
such challenge is the demand for aggressive power reduction. Although low power methodologies
have enabled significant power reductions while maintaining or even improving performance,
the recent demand for ultra-low power consumption in automotive, mobile and IoT applications
comes with strict power constraints. Another challenge is the routability problem and complex
design rule check violations (DRC) caused by aggressive area reduction in physical design.
2
Figure 1.2. Power and performance requirements for applications [89].
The complexity of design rules and constraints in advanced nodes exacerbates this problem,
necessitating additional human engineering resources to address the problem.
1.1.2 Challenges for Advanced Technology Nodes
As technology advances, aggressive scaling of critical dimensions and pitches poses

new challenges in placement, clock tree synthesis, routing and timing analysis. One of these
challenges arises due to the increasing impact of layout-dependent device parameter shifts.
This local layout effect (LLE) becomes more significant at advanced nodes due to aggressive
scaling. LLE can cause power and performance variations in devices, resulting from standard-cell
architectures and/or placement in physical design. For example, in standard cells, diffusion
breaks isolate two neighboring devices. In advanced nodes, these breaks can have a stress
effect on timing and leakage power. Figure 1.3(a) illustrates the saturation voltage and effective
current shifts caused by LLE from the second diffusion break. Figure 1.3(b) shows the interlayer
dielectric (ILD) stress in a TCAD simulation, which depends on the spacing between the first
and second diffusion breaks.
3
Figure 1.3. Local layout effect [128]. (a) Saturation voltage and effective current shifts caused
by the second diffusion break effect. (b) Interlayer dielectric (ILD) stress in a TCAD simulation,
which is a function of the spacing between the first and the second diffusion breaks.
Another challenge is the high resistance of interconnects due to aggressive BEOL metal
pitch scaling. Figure 1.4 illustrates the unit resistance of interconnects for recent technologies,
where the 5nm technology has a 9X larger unit resistance for the lowest BEOL metal layers (Mx)
compared to the 16nm technology. This leads to severe timing degradation from interconnects
and voltage (IR) drop problems.
1.1.3 Challenges for Design-Technology Co-Optimization
Design-Technology Co-Optimization (DTCO) is an optimization process that aims to

maximize the benefits of advanced technology nodes by exploring most beneficial design and
technology configurations. Fast and accurate PPAC explorations are key players in this process
as they help to explore new transistor architectures, materials, standard-cell architectures and
design knobs. In addition, scaling boosters are introduced at advanced technology nodes to
maintain the traditional Moore’s Law density scaling rate. These scaling boosters provide more
design and technology configurations and broader solution spaces for the PPAC explorations.
However, the exploration of these new technology options is becoming increasingly difficult and
complicated, and requires thorough evaluations.
Today’s DTCO methodologies face several limitations in advanced nodes. One of the
4
Figure 1.4. Resistance of BEOL metal layers (Mx) increases exponentially as technology
advances [89].
major challenges is the limited exploration of the entire solution space due to a lack of human
engineering resources. The process is not fully automated and requires significant manual
efforts from human engineers, such as creating standard-cell layouts and process design kits
(PDK). Another challenge is that the current ring oscillator-based PPAC explorations do not
always match the results of real chips as they do not address all aspects of design, such as
routability. Moreover, PPAC is design-dependent, and the results from ring oscillators cannot
be applied to all types of designs. Finally, due to the competitive nature of the semiconductor
industry, critical information related to advanced technologies remains under the proprietary
domain of semiconductor companies, making it inaccessible to academia. As a result, academic
researchers are unlikely to make significant innovations in DTCO research as they cannot conduct
experiments with real values corresponding to modern IC technologies.
1.2 This Thesis
This thesis presents robust physical design and DTCO methodologies to tackle challenges
posed by advanced technology nodes. The methodologies proposed in this thesis can be classified
5
Figure 1.5. Overview of this thesis.
into three main categories, as illustrated in Figure 1.5.
• General physical design methodologies.
• Technology-aware physical design methodologies.
• Design-technology co-optimization (DTCO) methodologies.
First, to address the challenges for modern IC design, the general physical design
methodologies chapter of this thesis presents approaches to optimize bounded-skew Steiner trees
for clock tree synthesis to minimize active power, and to concurrently refine detailed placement
and routing for efficient engineering change orders (ECO) automation.
Second, to address the challenges for advanced technology nodes, the technology-aware
physical design methodologies chapter of this thesis presents approaches to optimize leakage
power with the awareness of local layout effects, and to mitigate IR drop by detailed placement
that enables more power staples to be inserted.
Last, to address challenges for DTCO, the design-technology co-optimization chapter of
this thesis presents a systematic framework for routability assessments, a routability study by
6
using the PROBE2.0 framework with 3nm technology configurations and a systematic framework
for PPAC explorations with improved design enablements.
The reminder of this thesis is organized as follows.
• Chapter 2 presents two works in the context of general physical design methodologies.
The first work addresses clock power optimizations in modern IC designs. With more
memory hierarchies and various modules embedded in modern VLSI, on-chip memory
accesses are increasingly required due to the high cost of off-chip memory accesses
and the lower power density of memory fabrics. For memory-dominated chips, product
specifications and electronic device designers focus on the maximum power consumption
across all power usage scenarios, where a portion of memories are active and others
are turned off by clock/power gates. We introduce and study k-active dynamic power
minimization in bounded-skew trees, where we seek to minimize the maximum dynamic
power consumption when at most k clock sinks are active in the memory-dominant designs.
The experimental results give new insight into the tradeoff of maximum k-active dynamic
power and wirelength, and of skew and wirelength. The second work addresses the
challenges posed by aggressive area reduction in IC designs, which results in post-route
DRCs caused by complex design rules at advanced technology nodes. We propose CoRe-
ECO, a concurrent refinement framework for efficient automation of the ECO process.
The framework efficiently resolves pin accessibility-induced DRCs by simultaneously
performing detailed placement, detailed routing and cell replacement.
• Chapter 3 presents two works in the context of technology-aware physical design method-
ologies. The first work addresses local layout effects (LLE) in physical design. In a
standard cell-based design, a diffusion break (DB) isolates neighboring devices and affects
their delay and leakage power. In sub-10nm designs, the type of DB (single or double diffu-
sion break) used in the library cell layout can significantly impact device performance and
leakage. The 2nd DB effect is a type of LLE in which the distance to the second-closest DB
7
affects device performance. We develop 2nd DB-aware leakage optimization and detailed
placement heuristics. The second work addresses the high resistance of interconnects.
High resistance of BEOL layers causes high IR drop at advanced nodes. To solve this
problem, pre- or post-placed power staples are inserted in pin-access layers to connect
adjacent power rails and reduce power delivery network resistance, at the cost of reduced
routing flexibility, or reduced power staple insertion opportunity. We propose dynamic
programming-based single- and double-row detailed placement optimizations to maximize
the power staple insertion in a post-placement flow. We further propose metaheuristics to
improve the quality of results.
• Chapter 4 presents three works in the context of design-technology co-optimization method-

ologies. We first present a systematic framework for routability assessments. The key
challenge for DTCO is that large engineering efforts and long timelines are needed to
develop design enablements and perform implementation studies in order to assess tech-
nology options. We describe a new framework that can systematically evaluate a measure
of intrinsic routability, Kth , across both technology and design choices. We focus on
routability since it is a critical factor in the scaling of area and cost. The experimental
studies demonstrate the assessment of routability impacts for advanced-node technology
and design options. We demonstrate learning-based Kth prediction to reduce runtime, disk
space and commercial tool licenses needed to implement our framework. Also, we present
a routability study with this framework with 3nm technology configurations to show
block-level area scaling is reversed from cell-level scaling in cell height <150nm regime,
using a machine learning (ML)-assisted DTCO methodology to optimize block-level area
accounting for routability. Last, we present a systematic framework for design-technology
pathfinding with improved design enablements. The goal of our work is to build a “con-
figurable” and open-sourced PDK generation with promising scaling boosters, such as
backside PDN and buried power rails, and study PPAC given technology and design param-
8
eters. In addition, we propose our ML-based parameter tuning flow and clustering-based
cell width-regularized placements to generate more realistic artificial designs. We conduct
experiments to present PPAC evaluation capabilities of the framework with our predictive
technology and we achieve comparable results with studies from industry.
• Chapter 5 summarizes our conclusions in physical design and design-technology co-

optimization.
9
Chapter 2
General Physical Design Methodology
This chapter presents two general physical design methodologies. The first work ad-
dresses clock power optimizations in modern IC designs. With more memory hierarchies and
various modules embedded in modern VLSI, we introduce k-active dynamic power minimization
problem in bounded-skew trees, where we seek to minimize the maximum dynamic power
consumption when at most k clock sinks are active in the memory-dominant designs. The
experimental results give new insight into the tradeoff of maximum k-active dynamic power and
wirelength, and of skew and wirelength. The second work addresses the challenges posed by
aggressive area reduction in IC designs, which results in post-route DRCs caused by complex
design rules at advanced technology nodes. We propose CoRe-ECO, a concurrent refinement
framework for efficient automation of the ECO process. The framework efficiently resolves pin
accessibility-induced DRCs by simultaneously performing detailed placement, detailed routing
and cell replacement.
2.1 Optimal Bounded-Skew Steiner Trees to Minimize

Maximum k-Active Dynamic Power
Static Random-Access Memory (SRAM) is a key component of modern systems-on-chip

(SOC), as a result of several inexorable trends. First, off-chip memory accesses entail long
latencies and large energy consumption. Thus, modern SOCs have larger memory hierarchies
10
Figure 2.1. An example floorplan view of a memory-dominant chip [175]. The red boxes
indicate memory components in the chip.
and integrate more modules, trading away high-cost off-chip memory accesses for lower-cost
on-chip memory accesses. For example, microcontroller units contain multiple “masters”, such
as central processing units (CPU), memory protection units (MPU), floating-point processing
units (FPU), digital signal processors (DSP), etc. More than one of these masters may access
memory components simultaneously in many applications. Second, the lower power density of
memory fabrics can help reduce the need for “dark silicon”. Last, even as advanced technology
nodes continue to be aggressively scaled, SRAM scaling has been slower than other (logic)
scaling: device scaling (e.g., to single-fin transistors in 6T bitcells) causes susceptibility to
variations and read/write instability [166]. Due to these trends, as well as emergence of new chip
architectures for AI and machine learning, the area occupied by SRAM blocks now commonly
approaches or exceeds 60% of the total die area in modern SOCs [43].
Figure 2.1 shows a die shot of a recent SOC product in the mobile application processor
space. The reference [175] notes that the width and height of the SOC are around 9mm, and that
there is significantly more memory content relative to the previous-generation product.
11
Figure 2.2. “Cartoon” of floorplan and clock tree structures for a memory-dominant SOC design,
which motivates our k-active dynamic power minimization formulation.
Low-power design, always a significant challenge, becomes more important in modern

SOCs. Clock distribution is an important aspect of any low-power methodology. And, reducing
the power consumed by clock distribution to SRAMs is a growing concern in memory-dominant
designs. Figure 2.2 shows a “cartoon” view of a memory-dominant SOC floorplan and a tree
structure for clock distribution; the figure captures the physical design challenges of clock
distribution to memory blocks. The figure also shows a motivation for our proposed k-active
dynamic power minimization problem formulation.
With memory-dominant designs, memory clusters (sometimes referred to as “banks”)
are defined by designers; SRAM instances in the same cluster are placed close together at the
floorplan stage (four memories in a single cluster are highlighted in Figure 2.2). Importantly, in
many low-power SOC products, the product specification guaranteed to OEM customers and
other electronic product designers will focus on the maximum power consumption across all
power usage scenarios wherein a portion of memories is “active” (marked in green in Figure 2.2)
and all other memories are turned off by clock/power gates. Figure 30 of [157] shows current
12
consumption of active memory partitions (clusters) for an MCU product. The figure shows that
the maximum difference of current consumption across the partitions is around 30%. Thus,
minimization of maximum power consumption for active memory clusters is a real and significant
issue at the SOC product level. Figure 2.2 also illustrates the insertion of repeaters in long global
segments of the clock distribution. In recent technology nodes, the repeater distance is typically
at most 150 to 200µm, while die width and height can be 5 to 7mm or even more. The sizes
of SRAM blocks and the SOC die, relative to repeater distances in advanced nodes, effectively
“linearize” the calculation of clock power and buffered RC delay for top-level clock distribution.1
k-Active Dynamic Power Minimization Problem. In usage scenarios where only k memory
clusters are active and the rest are turned off by power/clock gates, the clock power consumption
depends on the locations of active clusters and the wirelengths of active clock subtrees. Among
all such scenarios, the worst-case power becomes the maximum k-active dynamic power. We thus
propose and study the k-active dynamic power minimization problem, where we seek to minimize
the maximum dynamic power consumption when at most k clock sinks are active. Based on the
linearity of power consumption and buffered RC delay in top-level clock distribution, we can
model k-active dynamic power and skew based on the union of source-sink paths to a given set
of k active clock sinks.
k-Consecutive-Active Dynamic Power Minimization Problem. To address potentially more
realistic scenarios, we remove unrealistic power usage scenarios which overconstrain the opti-
mization and result in suboptimal clock tree constructions. Figure 2.2 shows a scenario where
there are two neighboring clusters, and a third cluster is placed far away from the others. Depend-
ing on the SOC architecture and fine-grain access given to applications, the optimization may not
need to worry about scenarios where two simultaneously active clusters or sinks are placed far
apart. We thus introduce the k-consecutive-active dynamic power minimization problem which
simplifies constraints in the integer linear program (ILP) formulation that we develop below,
1 Power and delay are linear in distance, when long global interconnects are buffered with a regular buffer
interval [5].
13
since only sinks that are physically close can be active simultaneously. In our study below, we
index clock sinks (SRAM clusters) based on geometric information, and establish k-active power
constraints only for groups of k active sinks that have consecutive indices.
Flow-Based Integer Linear Program (ILP) for Bounded-Skew Tree. In the recent work
of [51], the authors propose a flow-based ILP formulation to find optimal (spanning and Steiner)
trees, given locations of clock source and sinks, along with skew constraints. We build on
the work of [51] to address the k-active and k-consecutive-active dynamic power minimization
problems. We also apply non-uniform grids to the previous ILP to obtain better results which
more closely reflect routing channels in the SOC floorplan, and locations of memories. Finally,
we study a scenario where we seek the optimal location of a flexible clock source rather than a
fixed clock source; we show how to generally formulate the ILP for optimal tree construction
when there are flexible and/or forbidden locations for the clock source.
We make the following contributions.
• Building on the work of [51], we use a flow-based integer linear programming (ILP)
approach. We add new constraints to minimize a weighted sum of total clock distribution
wirelengths and the maximum wirelengths of subtrees that reach k-active sinks. Our
formulation addresses wirelengths of subtrees based on “k-active dynamic power”.
• Our ILP formulation also addresses k-consecutive-active dynamic power of clock trees to
remove unrealistic power usage scenarios.
• We apply non-uniform grids to the ILP problem to more accurately model terminal
locations without incurring additional complexity increase.
• Going beyond pre-fixed clock source locations, we study effects of flexible as well as
forbidden clock source locations. The former helps to reduce overall wirelength and
wirelengths of subtrees with k-active sinks. The latter reflects scenarios where clock source
locations are unavailable due to SOC floorplan constraints.
14
The remainder of this section is organized as follows. Section 2.1.1 summarizes related
previous works. Section 2.1.2 describes our main problem formulation and its several variants.
Section 2.1.3 describes the setup and results for our experiments. We conclude in Section 2.1.4.
2.1.1 Related Work
In this section, we review previous related works. Since we address power minimization
for clock tree constructions, we first review the literature for zero-skew tree (ZST) and bounded-
skew tree (BST) construction problems for clock trees. We also note several example approaches
for low-power buffered clock tree construction in physical design.
Zero-Skew Tree (ZST) and Bounded-Skew Tree (BST) Construction. [10] [15] [38] [69]
propose deferred-merge embedding (DME)-based zero-skew tree (ZST) constructions which
subsequently became the foundation of many works for the bounded-skew tree construction
problem. The works [30] [58] [70] address the bounded-skew Steiner tree construction problem
with heuristic approaches. [19] uses hierarchical clustering for the ZST problem with dynamic
programming. [132] proposes approximation algorithms for ZST and BST. Another part of
the literature addresses optimal or approximate rectilinear Steiner minimal tree (RSMT) con-
structions. [3] proposes ILP (set covering) formulations to find a Steiner tree given a set of
terminals and designated Steiner points. [98] finds optimal Steiner trees with specified topology
satisfying lower and upper pathlength bounds, using linear programming. [121] proposes exact
algorithms to find Steiner trees based on mathematical programming and dynamic programming.
[28] proposes FLUTE, a fast lookup table-based RSMT algorithm. FLUTE is optimal when
the number of terminals is less than ten, and for the range of practical instance sizes offers an
extremely competitive runtime-quality tradeoff point for the RSMT problem. [51] studies the
skew-cost tradeoff and suboptimality in previous interconnect tree constructions, and proposes
an optimal flow-based ILP approach.
Low-Power Buffered Clock Tree. Many works have been proposed to minimize clock power
in physical design. In some situations, considering only wirelengths in clock trees as a proxy
15
Table 2.1. Notations.
Notation Meaning
pi ith terminal (pi ∈ P, p1 is a root (source)).
vi th
i vertex (vi ∈ V, P ⊆ V ).
euw A directed edge from vertex vu to vertex vw (euw ∈ E).
λuw 0-1 indicator of whether euw is in a tree T .
cuw Cost of edge euw .
i
fuw 0-1 indicator of whether the flow to pi goes through euw .
dui Pathlength at vu along unique v1 -pi path.
muw Manhattan distance from vu to vw .
B Skew bound.
L Lower bound on pathlength from source v1 .
Si Pathlength of p1 -pi path.
Si, j Pathlength of p1 -pi -p j tree.
Ci, j Length of common edges of p1 -pi path and p1 -p j path.
i, j
fuw 0-1 indicator of whether the flow to pi and/or p j goes through euw .
PMax(1) Pathlength of the longest source-to-sink (pi ) path (p1 -pi ).
PMax(2) Length of the longest source-to-two-sinks (pi , p j ) subtree (p1 -pi ∪ p1 -p j ).
M A large constant.
for cost (power) is not enough for real-world designs due to buffers on interconnects. [120]
addresses buffered clock power minimization by proposing an algorithm to design a tree topology
and simultaneously insert buffers. [36] proposes a methodology for power-aware clock tree
planning, named LPClock, which constructs clock trees heuristically and inserts clock gating
cells to minimize clock power. [16] proposes a heuristic approach to construct a low-power
zero-skew clock network by buffer and clock gate insertions, given terminal locations and activity
information. By contrast, in this work, we focus on clock tree constructions in memory-dominant
SOC designs. By considering SRAM blocks, SOC die sizes and buffer distances in advanced
nodes, we observe that both the clock power and the signal propagation delay in clock subtrees
can be modeled as linear phenomena.
We are not aware of previous works that study minimization of subtrees across subsets of
k sinks. Below, we point out that traditional bounded-skew or skew-wirelength tradeoffs do not
directly address this new objective. On the other hand, directly minimizing the cost of k-sink
16
clock subtrees in memory-dominated SOC designs dimensions can reduce clock distribution
power in actual operation of IC products. Our work can provide guidance for clock distribution
in such designs.
2.1.2 Problem Formulation
In this section, we describe our problem formulation. First, we review the formulation of
[51]; we then modify it to address the k-active dynamic power minimization problem. Second, we
formulate the k-consecutive-active dynamic power minimization problem. Third, we incorporate
non-uniform grids into the ILP formulation to more flexibly reflect clock source and sink
locations. Last, we propose both flexible and forbidden clock source locations to further guide the
choice of optimal clock source locations for reduced tree cost and/or subject to layout constraints.
Table 2.1 shows the notations that we use in what follows.
Review of the ILP Formulation of [51]
We now review the ILP formulation for optimal bounded-skew spanning and Steiner tree
construction in the work of [51]. Let xi and yi be coordinates of terminal pi in the Manhattan plane.
P denotes a set of terminals pi over which a tree is to be constructed, where i = {1, 2, ..., |P|}.
By convention, p1 is a root (source) of the tree and the other terminals are leaves (sinks). A
graph G = (V, E) is constructed which includes terminals in P along with the Hanan grid [53]
and additional half-integer grids (centered lines) between the predefined Hanan grids.2 Every
point at the intersection of two gridlines corresponds to a vertex vi ∈ V , and directed edges euw
from vertex vu to vertex vw are defined where vu and vw are neighboring vertices in the grid.
Thus, each vertex vi can have up to four incoming edges and four outgoing edges. The objective
is to minimize the total cost of edges (i.e., total wirelength) in the constructed tree T .
2 In the work of [51], optimal spanning and Steiner trees are constructed on Hanan grids or half-integer grids.
17
Minimize: ∑ λuw · cuw
u,w
Subject to:
i
λuw ≥ fuw , ∀pi ∈ P, euw ∈ E (2.1)





 1 if vw = v1 , ∀vu ∈ V, pi ∈ P, i ̸= 1


i i
∑ fuw − ∑ fwu = −1 else if vw = vi (2.2)
u u 





 0 otherwise
i
∑ cuw · fuw ≥ L, ∀pi ∈ P, i ̸= 1 (2.3)
u,w
i
∑ cuw · fuw ≤ L + B, ∀pi ∈ P, i ̸= 1 (2.4)
u,w
Given a set of terminals P including source p1 , and pathlength lower and upper bounds L and
L + B, where B is a skew bound, the ILP finds a spanning or Steiner tree over P that satisfies the
pathlength bounds while minimizing ∑u,w λuw · cuw .
i denotes a binary indicator of whether the flow to p goes through e . Constraint (2.1)
fuw i uw
enforces λuw = 1 when any flow goes through euw . Constraint (2.2) enforces flow conservation
at all vertices except for the source (where the outgoing flow exceeds the incoming flow by the
number of sinks) and any sink (where the sum of incoming flow exceeds the outgoing flow by
one).
Skew bounds are enforced by Constraints (2.3) and (2.4), which apply the lower bound
(L) and the upper bound (L + B) to each source-to-sink pathlength.
18

dui = 0




 if vu = v1 , ∀ pi ∈ P, i ̸= 1


dui ≥ L else if vu = vi , ∀ pi ∈ P, i ̸= 1 (2.5)




dui ≤ L + B otherwise ∀ pi ∈ P, i ̸= 1



dui + M · (1 − fwu
i
) ≥ dwi + cwu , ∀vu , vw ∈ V, u ̸= w, pi ∈ P (2.6)
dui − M · (1 − fwu
i
) ≤ dwi + cwu , ∀vu , vw ∈ V, u ̸= w, pi ∈ P (2.7)
Constraints (2.5) to (2.7) prevent cycles in the output, i.e., they enforce a tree solution topology.
Without these constraints, cycles can be introduced by detouring to satisfy skew constraints.
To avoid cycles, we define a variable dui which denotes pathlength from the source to vertex
vu for a path to sink pi . For Constraint (2.5), the pathlength is zero when vertex vu is the
source. The pathlength must be maintained between lower and upper bounds for any vertices.
Constraints (2.6) and (2.7) are added for pathlength dui . When fwu
i = 1, a pathlength increases by
the edge cost cwu , which means that dui will be infinite in the presence of a cycle, and will not
satisfy Constraint (2.5).
if m1u + mui > L + B, ∀vu ∈ V, u ̸= 1, ..., |P|, pi ∈ P

i i
fuw = 0, fwu = 0, ∀euw ∈ E, ewu ∈ E (2.8)
19
if m1u + muu′ + mu′ i > L + B && m1u′ + mu′ u + mui > L + B
i
fuw + fui ′ w′ = 1, i
fwu + fui ′ w′ = 1, ∀euw ∈ E, ewu ∈ E (2.9)
i
fuw + fwi ′ u′ = 1, i
fwu + fwi ′ u′ = 1, ∀eu′ w′ ∈ E, ew′ u′ ∈ E (2.10)
Constraints (2.8) to (2.10) are added to improve ILP runtime by special handling of vertices
and edges that are irrelevant to the desired tree solution. Constraint (2.8) makes flow variables
i and f i zero when the sum of the Manhattan distances from source p to vertex v and
fuw wu 1 u
from vertex vu to sink pi is larger than the upper bound L + B. This is because the pathlength
of path p1 -pi going through vertex vu is always larger than the upper bound. Constraints (2.9)
and (2.10) address a forbidden scenario involving two vertices. When the sum of the pathlengths
from p1 going through non-terminal vertices vu and v′u to sink pi is larger than the upper bound
L + B, then any flows for incoming and outgoing edges going through vu and v′u cannot exist
simultaneously.
k-Active Dynamic Power Minimization
We extend the approach of [51] to address the k-active dynamic power minimization
problem, as follows. We first modify the objective function from the ILP formulation, then we
add six constraints to capture k-active dynamic power minimization problem for k = 1 and k = 2.
Our objective function is to minimize the weighted sum of total cost of edges in an output tree T
and the maximum k-active dynamic power. For k points other than p1 , we say that the k-active
dynamic power is the total cost of edges in the union of the corresponding k source-to-sink paths
in an output tree T . In addition, for a given output tree T , we define the maximum k-active
dynamic power, PMax(k), to be the maximum k-active dynamic power over all k-subsets of
{p2 , . . . , p|P| }. Thus, our objective function and constraints can be formulated as follows.
20
Minimize: ∑ λuw · cuw + α · PMax(1) + β · PMax(2)
u,w
Given a set of terminals P including source p1 , pathlength bounds L and L + B, and parameters
α and β , we construct a Steiner tree over P that satisfies the pathlength bounds while minimizing
∑u,w λuw · cuw + α · PMax(1) + β · PMax(2). The first term of the objective function is the total
cost of the edges. λuw = 1 when edge eu,w is included in the output tree T . The second and
third terms are the weighted k-active dynamic power where k = 1, 2 respectively. Based on the
principle of inclusion-exclusion, we can extend to any k-active costs from k = 1 and k = 2, at the
cost of additional wirelength summation terms and constraints that govern k-tuples of sinks.
Si ≤ PMax(1), ∀pi ∈ P (2.11)
Si, j ≤ PMax(2), ∀pi , p j ∈ P (2.12)

i
∑ cuw · fuw ≤ Si , ∀pi ∈ P (2.13)
u,w
Si + S j −Ci, j ≤ Si, j , ∀pi , p j ∈ P (2.14)

i, j i j
2 · fuw ≤ fuw + fuw (2.15)
i, j
∑ cuw · fuw = Ci, j , ∀pi , p j ∈ P (2.16)
u,w
Constraints (2.11) to (2.16) are added for the PMax(k) computation. In Constraint (2.11), Si
denotes a pathlength from the source p1 to sink pi and PMax(1) denotes the maximum pathlength
among all Si . In Constraint (2.12), Si, j denotes a length of a source-sinks (pi and p j ) subtree
and PMax(2) denotes the longest pathlengths among source-to-two-sinks (pi and p j ) subtrees.
Constraint (2.13) computes pathlength Si from p1 to pi . In Constraint (2.14), Ci, j is the common
pathlength between the source-to-sink paths for pi and p j . The constraint is therefore with
21
respect to the total edge length of a source-to-two-sinks (pi and p j ) subtree. In Constraint (2.15),
i, j
a new flow variable fuw is defined, which indicates whether edge euw is common to both of the
source-to-sink paths to pi and p j . Constraint (2.16) computes common pathlength Ci, j for sinks
pi and p j .
k-Consecutive-Active Dynamic Power Minimization
In this section, we describe a k-consecutive-active dynamic power minimization problem.

With this variant formulation, we capture locality of simultaneous memory accesses, and can
thus relax constraints for unrealistic power usage scenarios which overconstrain the optimization.
In our study, we index clock sinks based on geometric information and consider locality
according to adjacent indices. We treat the clock source as the origin, and associate each clock
sink with its theta angle with respect to this origin. Sinks are indexed (in counterclockwise order)
by theta angle.
We do not add new constraints to the ILP. Instead, we replace four constraints with the
following:
Si,i+1 ≤ PMax(2), ∀pi , pi+1 ∈ P (2.17)
Si + Si+1 −Ci,i+1 ≤ Si,i+1 , ∀pi , pi+1 ∈ P (2.18)

i,i+1 i i+1
2 · fuw ≤ fuw + fuw (2.19)
i,i+1
∑ cuw · fuw = Ci,i+1 , ∀pi , pi+1 ∈ P (2.20)
u,w
For the k-consecutive-active dynamic power minimization problem, Constraints (2.17), (2.18),
(2.20) and (2.19) replace Constraints (2.12), (2.14), (2.15) and (2.16). In this way, constraints
from the original k-active dynamic power minimization problem are relaxed for the k-consecutive-
active dynamic power minimization problem. Instead of addressing all the possible combinations
of k sinks, the relaxed constraints only address local groups of k sinks.
22
23
Figure 2.3. Terminals with uniform and non-uniform grids on an example floorplan. (a) Terminals with uniform grids. (b) Terminals
with non-uniform grids based on channel intersection graph. (c) Terminals with non-uniform grids after the unusable grids.
Non-Uniform Grid
In this section, we demonstrate how non-uniform underlying grids enable the ILP to
more flexibly capture locations of terminals and available routing resources in the Steiner tree
construction. A non-uniform grid is created for a given floorplan and macro (SRAM) placement.3
We consider channels between macros as well as clock pin locations (or midpoints of macros
in clusters) of macros in our non-uniform grids; in the limiting case, the grid can be said to
approach a channel intersection graph [33]. Removal of unusable edges in the grid can make the
problem sparser and improve ILP runtimes.
Figure 2.3 shows uniform and non-uniform grids on an example floorplan from a RISC-V
based design.4 Figure 2.3(a) shows uniform grids with locations of terminals. Note that we
would like to accurately capture the locations of clock pins (terminals) on SRAMs. However,
in a uniform grid these locations must be snapped to nearest grid intersections. This causes a
discrepancy between desired terminal locations and the terminal locations actually used on the
underlying uniform grids, leading to suboptimal tree solutions. On the other hand, Figure 2.3(b)
shows a non-uniform grid based on the macro channels and clock source locations, with removal
of unusable grids where terminals are not located. Figure 2.3(c) shows non-uniform grids after
removing unusable grids. In this way, we can use desired terminal locations of clock pins for our
tree construction and have more flexibility of terminal locations without adding more complexity
to the ILP.
Flexible or Forbidden Clock Source Location
Given fixed locations of a clock source and sinks, we find an optimal solution to minimize
costs while skew constraints are satisfied. However, in real designs, we might be able to move
a clock source location unless fixed locations are hard constraints. For example, if a clock
3As in [51], a graph is constructed based on the Hanan grid and additional centered lines between the predefined
grids.
4We synthesize netlists of SweRV wrapper design [170] [176] using [171], and perform place-and-route us-
ing [144], in a foundry 14nm FinFET enablement. We perform macro placement by using [144] and manually
adjust the macro placement. The design has 28 macros and 74K standard-cell instances.
24
25
Figure 2.4. Example trees for flexible clock sources. (a) A tree with a fixed clock source. (b) A tree with a movable clock source, with
optimal source location. (c) A tree with the tree from (b) and the original source location from (a). (d) A tree with a movable clock
source that has a forbidden area.
generator is placed and fixed at a specific area or pin assignments for top-level implementations
are given, we cannot move the clock source location. Otherwise, we can achieve better tree
solutions by moving clock source locations. To this end, we describe enablement of flexible
clock source locations.
To allow all vertices (except sinks) to be sources, we define a clock source at a new
virtual vertex v0 . In the graph G = (V, E) constructed, we add virtual edges from the virtual
vertex v0 to every vertex vi . Thus, each vertex vi has up to five incoming edges and four outgoing
edges. We then set the cost of each virtual edge to zero. The vertex that ends up connected to the
virtual vertex through a virtual edge becomes the actual clock source. We define two constraints
for flexible clock source locations, as follows.
∑ λ0u = 1, ∀e0u ∈ E (2.21)

0,u
λ0u = 0, ∀vu = vi , pi ∈ P, i ̸= 1 (2.22)
Constraint (2.21) ensures that only one virtual edge is used in tree solution T . Constraint (2.22)
is to avoid flows from a virtual vertex (clock source) to any sinks with zero cost.
Figure 2.4 shows example trees (|P| = 12) with a fixed clock source and a movable clock
source. Figure 2.4(a) shows an optimal tree with the fixed clock source at (3, 5.6). The total
wirelength of the tree solution is 25.6 when lower bound L = 3.44 and upper bound L + B = 8.6.
In this example, we set α = 1, β = 0.5 in the weighted sum of total wirelength and k-active
dynamic power. Given a virtual clock source, the tree solution in (b) makes a connection from
the virtual source to the vertex at (2.4, 3), which becomes the actual clock source. Figure 2.4(c)
shows a combined tree with the tree from Figure 2.4(b) and the original clock source location
from Figure 2.4(a). As seen in the data table of Figure 2.4, the tree of Figure 2.4(c) has larger
cost than the tree of Figure 2.4(a), as we expect (tree (a) is optimal for this particular source
location).
26
On the other hand, even if clock sources can be flexibly moved, they might not be
placeable at certain locations due to geometric constraints. To address this, the ILP can be
modified to reflect a “forbidden area” wherein clock sources cannot be placed. The following
constraint is added to define the forbidden area for clock sources.
λ0u = 0, ∀vu ∈ Vr , (2.23)
Constraint (2.23) prevents connections from the virtual edge to any vertices within a given
forbidden area, where Vr denotes the set of vertices in the forbidden area. Figure 2.4(d) shows
an example tree with flexible clock source locations and a forbidden area for the source. In this
example, we set the forbidden area to be 2 ≤ x ≤ 4, 2 ≤ y ≤ 4 – e.g., reflecting a situation where
clock pins or clock generator cannot be placed in the middle of the die area.
2.1.3 Experimental Setup and Results

Experimental Setup
We implement our work in C++ with CPLEX 12.8.0 [155] and Gurobi 9.0.1 [153] as our
ILP solvers. Our experiments are performed with four threads on a 2.6GHz Intel Xeon server. In
our objective function, we set the weighting factors α = 0, 1, 100 and β = 0, 0.5, 100. For lower
bounds, we set lower bound L = {4, 5, 6, 7, 8, 9}. For our runtime study, we set lower bound
L = M − B to avoid infeasiblility, where M denotes the maximum Manhattan distance between
source and sink. Also, we define the skew bound B using M multiplied by 0.2, 0.4, 0.6, 0.8 (i.e.,
B = 0.2 · M, 0.4 · M, 0.6 · M, 0.8 · M). We also set unbounded cases with zero lower bound and
infinite upper bound. In our experiments, we set two hours as the runtime limit for ILP.5
5The runtime limit is not applied to our runtime study in Table 2.2.
27
28
Figure 2.5. An example set of non-dominated trees for a 12-terminal pointset. The weighting factors and bounds are shown as (α, β ,
L, B). (a) (0, 0, 5, 0.6 · M). (b) (0, 100, 4, 0.6 · M). (c) (0, 100, 5, 0.6 · M). (d) (100, 0, 7, 0.4 · M). (e) (0, 100, 7, 0.4 · M). (f) (0, 100, 7,
0.2 · M).
29
Figure 2.6. Example trees with various weighting factors in our objective function. (a) A tree solution to minimize the total wirelength.
(b) A tree solution to minimize PMax(1). (c) A tree solution to minimize PMax(2).
Figure 2.7. Scatter plots for (a) skew-wirelength, (b) PMax(1)-wirelength and (c) PMax(2)-
wirelength for the 12-terminal pointset with various weighting factors α and β . Additional points
that do not correspond to the figure legend are obtainable with other values of α, β , L and B.
Experimental Results
Study of Weighting Factors (α and β ) in an Objective Function. We show example trees

and scatter plots for skew-wirelength, PMax(1)-wirelength and PMax(2)-wirelength tradeoffs
with various weighting factors α, β in our objective function. When α = β = 0, the results are
the same as those of [51]. In this way, we can compare our results to [51]. We observe that in
practice, there are not many non-dominated trees for a given instance, across all values of α, β ,
lower and upper bounds. A tree T1 is said to be dominated by tree T2 if the wirelength, skew,
PMax(1) and PMax(2) metrics of T2 are all equal to or better than those of T1 . Figure 2.5 shows
an example set of non-dominated solution trees for one instance in our experiment.
Figure 2.6 shows example trees with various weighting factors in our objective function.
Given a 12-terminal pointset, lower bound L = 6 and skew bound B = 0.4 · M, we find tree
solutions where α = β = 0, α = 100, β = 0 and α = 0, β = 100, respectively. We set lower
bound L = 4, 5, 6, 7, 8, 9 in this experiment. Figure 2.6(a) shows the tree solution which minimizes
the total wirelength of the tree. Figure 2.6(b) shows the tree solution which minimizes PMax(1).
The solution has the minimum total wirelength among all solutions with minimum PMax(1).
Compared to Figure 2.6(a), PMax(1) decreases from 9.4 to 8.6 while the total wirelength
increases. Figure 2.6(c) shows the tree solution which minimizes PMax(2). The solution has
the minimum total wirelength among all solutions with minimum PMax(2). Compared to
30
Figure 2.6(a), PMax(2) decreases from 14.6 to 13.6 while the total wirelength increases. From
this example, we can see that solution trees can be optimized for PMax(1) and PMax(2) by
varying the weighting factors in the objective function.
Figure 2.7 shows scatter plots for skew-wirelength, PMax(1)-wirelength and PMax(2)-
wirelength. Figure 2.7(a) shows tradeoff curves with various weighting factors. Compared to
the cases that minimize only total wirelength (α = 0 and β = 0, i.e., as in the work of [51]), the
tradeoff curves for nonzero weighting factors form a “bundle” of tradeoff curves. In terms of
the skew-wirelength tradeoff, we see that considering k-active sinks does not affect skew much.
Figure 2.7(b) gives a scatter plot for PMax(1)-wirelength, showing that PMax(1) decreases
when α is nonzero. Similarly, Figure 2.7(c) gives a scatter plot for PMax(2)-wirelength, showing
that PMax(2) decreases when β is nonzero. Figure 2.7(b) includes the solutions with nonzero
α and Figure 2.7(c) includes the solutions with nonzero β to show improved PMax(1) and
PMax(2), respectively.
Study of k-Consecutive-Active Sinks. As described above, we index sinks according to theta
angle about the clock source, in counterclockwise order. Figure 2.8 shows scatter plots for
skew-wirelength, PMax(1)-wirelength and PMax(2)-wirelength tradeoffs with k-active sinks
and k-consecutive-active sinks. We set weighting factors α = 1, β = 0.5 and lower bound
L = 4, 5, 6, 7, 8, 9. Figures 2.8(a) and (b) respectively show skew-wirelength and PMax(1)-
wirelength for k-consecutive-active sink solutions, similar to what was presented for k-active sink
solutions. Figure 2.8(c) shows that PMax(2) values of k-consecutive-active sinks are less than
those of k-active sinks. This is because PMax(2) of k-consecutive-active sinks only considers
local subtrees with consecutive indices of clock sinks. (Note that the k-consecutive-active sinks
optimization will lead to different solutions than the k-active sinks optimization only when the β
weighting term for PMax(2) is nonzero.)
Study of Flexible or Forbidden Clock Source Locations. Figure 2.9 shows skew-wirelength,
PMax(1)-wirelength and PMax(2)-wirelength scatter plots for flexible and fixed clock source
locations for the 12-terminal pointset in Figure 2.4. In this experiment, we set weighting factors
31
wirelength for k-active sinks and k-consecutive-active sinks.
wirelength for fixed and flexible clock source locations, as well as flexible clock source locations
with a forbidden 2 ≤ x ≤ 4, 2 ≤ y ≤ 4 area.
α = 1, β = 0.5 and lower bound L = 4, 5, 6, 7, 8, 9. Figures 2.9(a) to (c) shows that tree solutions
with flexible clock source locations have better metrics (total wirelength, PMax(1) and PMax(2))
than those with fixed clock source locations. Adding a forbidden area (reducing flexibility)
worsens results in (c), but skew-wirelength, PMax(1)-wirelength and PMax(2)-wirelength
remain superior to when the clock source location is fixed.
Study of Runtime. In order to compare ILP runtime, we define our pointsets with P =
{8, 10, 12, 14, 16}, where |P| is the number of terminals of trees. For this experiment, for
each |P| we randomly generate 10 sets of terminals within a seven by seven grid. We also
generate random grid edge costs in horizontal and vertical directions (i.e., between consecutive
32
Table 2.2. Average ILP runtime when α = 1, β = 0.5, L = M − B. M denotes the maximum
distance between source and sink.
ILP solver Skew bound |P| = 8 |P| = 10 |P| = 12 |P| = 14 |P| = 16

Unbounded 9.30 37.21 151.23 288.77 488.97
0.8 · M 10.07 12.17 32.69 56.57 100.31
Solver A 0.6 · M 22.63 31.34 66.38 110.35 175.87
0.4 · M 71.29 96.41 153.19 1082.44 2454.79
0.2 · M 20.19 638.81 2912.25 14979.97 35566.68
Unbounded 5.9 90.57 329.92 642.07 1454.97
0.8 · M 2.43 10.78 68.09 156.59 567.08
Solver B 0.6 · M 7.87 91.53 619.80 757.56 1583.59
0.4 · M 96.93 1606.85 2091.18 11528.67 27122.51
0.2 · M 37.02 17800.59 24265.66 74113.28 107977.12
x- or y-coordinate values), from a uniform distribution over the interval [0.5, 2.0]. We use two
different ILP solvers [153] [155] in this experiment.6
Table 2.2 shows the average ILP runtime. As shown in Table 2.2, ILP runtime increases
when the number of terminals increases in most cases. Even though there are outliers due to
infeasiblity of some of the problems, runtime increases when skew constraints get tighter (i.e.,
smaller skew bound).7 The runtime for Solver A is smaller than for Solver B in most of the
cases.
We also compare runtime for fixed and flexible clock source locations with non-uniform
grids. Table 2.3 shows ILP runtime for fixed and flexible clock source locations with two
different non-uniform grids, defined as follows. Sparse denotes non-uniform grids after removal
of unusable edges as in Figure 2.3(c). Dense denotes non-uniform grids where unusable edges
are retained, as in Figure 2.3(b). Due to the inclusion of a virtual source and edges, runs with
flexible clock source location have longer runtimes than corresponding runs with fixed locations.
We see that ILP solution runtime is reduced by removing unusable edges.
6 Inorder to avoid benchmarking the tools, below we report tool names only as Solver A and Solver B.
7We cannot explain why the runtime for unbounded cases is smaller than for any bounded cases in [51]. In our
studies, if we remove constraints for lower and upper bounds, the runtime slightly increases compared to the cases
of B = 0.8 · M in Table 2.2.
33
Table 2.3. ILP runtime for fixed and flexible clock source locations in Figure 2.4 and non-
uniform grids. We set variables α = 1, β = 0.5, L = M − B. M denotes the maximum distance
between source and sink.
Sparse Dense
Skew bound
Fixed Flexible Fixed Flexible
Unbounded 95.82 278.87 1964.88 2073.29
0.8 · M 7.13 132.41 117.19 1581.06
0.6 · M 34.89 432.06 500.11 3280.04
0.4 · M 143.31 36781.35 2607.33 163050.29
2.1.4 Conclusion
In this work, we have proposed a new k-active dynamic power minimization problem
that arises in clock distribution for memory-dominant SOC designs. We observe that in modern
technologies the scale of SRAM blocks and SOC die, relative to global repeater distances,
effectively “linearizes” the problem and makes it amenable to combinatorial optimization.
By extending the recent work of [51], we formulate an ILP for the k-active dynamic power
minimization problem as well as a k-consecutive-active dynamic power minimization variant,
The ILP minimizes the weighted sum of the total costs (wirelengths) and the maximum power
across all sets of k active sinks. We also propose non-uniform grids to capture flexible locations
of terminals in an output tree, and the handling of both flexible and forbidden clock source
locations to improve solution quality and solver runtime. Our experimental results give new
insights into the tradeoff for maximum k-(consecutive)-active dynamic power and wirelengths of
subtrees as well as skew and total wirelengths.
Our future directions include (1) implementations of our tree solutions as a guide for clock
tree synthesis in commercial place-and-route methodologies; (2) heuristics to combine ILP and,
e.g., problem decomposition strategies to address larger-scale instances of the k-(consecutive)-
active dynamic power minimization problem; and (3) runtime improvements of ILP solution,
particularly with flexible clock tree locations.
34
2.2 CoRe-ECO: Concurrent Refinement of Detailed Place-
and-Route for an Efficient ECO Automation
With the relentless scaling toward advanced technology nodes, increasingly sophisticated
IC fabrication constraints (e.g., fewer routing tracks, higher pin density and complicated condi-
tional design rules) bring rapidly increasing design complexity [110]. This leads to non-trivial
challenges for physical design, particularly in the routing stage. The number of remaining design
rule check violations (DRC) after place-and-route (P&R) has become one of the most crucial met-
rics for an automatic IC layout solution, since back-end designers must manually stitch/modify
all DRCs via implementation of engineering change orders (ECO) at the post-layout stage. In
particular, achieving the pin accessibility needed to resolve DRCs is a critical, time-consuming
engineering task just before tapeout, and is therefore a critical bottleneck in the advanced-node
IC design process [56] [103].
To mitigate pin accessibility-induced DRCs, several approaches are proposed to improve
pin accessibility through detailed placement (DP) optimization [35] [66] [72] [88] [108] [130]
and standard-cell layout optimization [23] [126]. The authors of [88] perform detailed placement
optimization using a global routing solution as guidance, with pin accessibility modeled only
in the form of pin density. Dynamic programming and deep learning-based detailed placement
optimizations considering each pin’s access are developed in [35] [66] [72] [130]. In [108], the
authors introduce a measurement of inaccessible pins in a cell to optimize detailed placement.
However, these models have limited capability of comprehending design rules in detailed
routing (DR). The works of [23] and [126] have proposed pin accessibility-driven cell layout
optimization frameworks for improving routability at block-level. Recently, the authors of [76]
perform replacement of inaccessible cells with diverse cell layouts in terms of pin locations
and number of access points, in the ECO stage. The applicability of these works is intrinsically
limited due to the lack of holistic consideration for the block-level design steps (e.g., detailed
placement and detailed routing).
35
Figure 2.10. Framework overview.
In this work, we propose CoRe-ECO, a Concurrent Refinement framework which si-

multaneously performs incremental detailed placement and detailed routing, along with cell
replacement, at the ECO stage. To our knowledge, this is the first work to present a concurrent
co-optimization of detailed placement, detailed routing and cell replacement within block-level
P&R. Our main contributions are summarized as follows.
• We propose a concurrent refinement framework which simultaneously performs detailed

placement, detailed routing and cell replacement for each local window (i.e., switchbox)
covering DRC locations, to determine ECOs at post-layout stage. We devise a novel dy-
namic cell allocation (DCA), merging the refinement steps into a single-step optimization.
• CoRe-ECO performs (i) placement adjustment such as horizontal/vertical shifting, hor-

izontal flipping and cell swapping; (ii) pinlength extension with a recommendation of
adopting alternative master cells while maintaining the same functionality; and (iii) routing
36
optimization to seek the best-quality DRC-clean solution.
• CoRe-ECO minimizes perturbation of the given layout by utilizing a satisfiability modulo

theories (SMT) solver, enabling multi-objective optimization.
• We validate the proposed CoRe-ECO framework with various testcases, demonstrating

successful fixing of pin accessibility-induced DRCs through the proposed ECO flow along
with total wirelength optimization.
The remaining sections are organized as follows. Section 2.2.1 describes the proposed
CoRe-ECO framework. Section 2.2.3 discusses our experimental setup and results. Section 2.2.4
concludes the work.
2.2.1 CoRe-ECO Framework
This section introduces an overview of the proposed framework, grid-based P&R archi-
tecture, refinement operations, switchbox generation, perturbation-minimized optimization and
SMT formulation.
Framework Overview
We formulate a conventional (sequential) layout refinement process as a constraint

satisfaction problem (CSP) with variables and constraints to integrate placement adjustment
and routing steps into a single multi-objective optimization problem. We adopt the SMT
solver Z3 [34] to solve the given optimization problem. Figure 2.10 illustrates an overview of
our framework. Given standard-cell library, switchbox and instance slack information,8 our
framework simultaneously obtains the optimal solution that strictly satisfies the constraints
integrated into our novel DCA scheme. Our notations are described in Table 2.4.
8We define instance slack as the worst slack among pins of a given instance.
37
Table 2.4. Notations for the proposed CoRe-ECO framework.
Notation Meaning
T Set of instances in a switchbox.
t t th instance.
lt 0-1 indicator if instance t is flipped.
xt x-axis coordinate of lower-left corner of t.
yt y-axis coordinate of placement row of t.
ytorg Initial y-axis coordinate of placement row of t.
xtorg Initial x-axis coordinate of lower-left corner of t.
wt Width of instance t.
Pt Set of internal pins of instance t.
pti ith pin of instance t.
aext (pti ) Set of extended vertices of pin pti .
V (Vi ) Set of vertices in (ith metal layer of) the routing graph G.
v A vertex with the coordinate (xv , yv , zv ).
a(v) Set of adjacent vertices of v.
ev,u An edge between v and u, u ∈ a(v).
wv,u Weighted cost for metal segment on ev,u .
N Set of multi-pin nets in the given routing box.
n nth multi-pin net.
fmn A two-pin subnet connecting a source and sink, i.e., a commodity.
env,u 0-1 indicator if ev,u is used for n.
fmn (v, u) 0-1 indicator if ev,u is used for commodity fmn .
mv,u 0-1 indicator if there is a metal segment on ev,u .
Cmn (v, u) Capacity variable for ev,u of commodity fmn .
Grid-based Place-and-Route Architecture
We define the grid-based placement and 3-D routing graph composed of four metal
layers (i.e., M1-M4) as shown in Figure 2.11. Cell instances and I/O pins are aligned with
M1 vertical tracks and gate poly of the standard cell. Inspired by [73], we adopt supernodes
to cover the multiple candidates for each pin, either the I/O pin of a standard cell (i.e., PIN ) or
the outer pin of a switchbox (i.e., PEX ). The location of PIN is dynamically determined by the
placement formulation and is associated with the flow formulation for routing through DCA.
PEX interconnects the internal pins inside the switchbox to the outer pins and is located along the
38
Cell Placement Grid
VDD Standard Cell Instance
Origin of Cell Instance
PIN1
1 2
PIN2
Clipped 1 2
M1 Track
Cell Pins on Metal1 Internal Pin
2 Candidates
(Fixed) PIN3 Pins on Metal2
3 3 3 (PIN)
Pins on boundary External Pin

VSS Candidates
of switchbox
(PEX)
Routing Grid
Metal2 (M2) Metal3 (M3) Metal4 (M4)
Figure 2.11. Grid-based place-and-route architecture.
boundary of the switchbox which corresponds to the pre-routed result.

The horizontal routing grid (i.e., M2 and M4) consists of eight tracks per placement row
and the vertical routing grid (i.e., M3) is aligned with the cell placement grid.9 Note that we
focus on M2-M3 layers assuming that M2 and M3 have the same metal pitches, because our
proposed framework targets the correction of the pin accessibility-induced DRCs. Thus, M4
layer only contains VIA34 elements as the external pins for connections to the upper layers.
Refinement Operation
CoRe-ECO uses placement adjustment and pinlength extension as refinement operations

within its adaptive perturbation method.
9We assume an on-grid routing scheme for each routing layer, consistent with sub-7nm multi-patterning
technologies and IC practitioners’ restriction of preferred routing direction per each layer [142]. In this work, we do
not support different pitches for M2, M3 and M4, which are out of scope.
39
M1 Pin M1 Track
VDD VDD
Inaccessible Accessible M2 Pin VIA12
M2 Wire (Pre-Routed)
VDD
VSS VSS
(a) (b)
VDD VDD VSS
VSS VSS VDD

(c) (d) (e)
Figure 2.12. Placement adjustments. (a) Initial placement. (b) Horizontal shifting. (c) Horizontal
flipping. (d) Cell swapping. (e) Vertical shifting.
Placement Adjustment. Figure 2.12 illustrates possible adjustments during detailed placement
to solve the pin accessibility-induced DRC. When the given placement layout in a switchbox
(Figure 2.12(a)) does not have feasible routing solutions due to an inaccessible M1 pin, CoRe-
ECO adjusts the placement of instances in the switchbox by horizontal/vertical shifting, horizontal
flipping and cell swapping as shown in Figures 2.12(b)-(e). Note that cell swapping is only
performed between two adjacent instances.
Pinlength Extension. As a rule, master cells with minimum pinlength are preferred for use
during initial P&R, e.g., to achieve better timing optimization. However, in the ECO stage,
engineers should consider adopting alternative master cells for specific instances, so as to
improve pin accessibility or to achieve the target design specification. While works of 20+ years
ago [9] [11] [92] pursued liquid library approaches, today it is well-understood that a library
cell must be qualified before it is used in a production chip. Thus, our framework suggests
minimum-achievable pinlength extensions needed to fix DRCs: by only extending the metal
segments of I/O pins, we enable engineers to adopt (i.e., swap in) alternative master cells that
maintain the same functionality while minimizing the magnitude of undesired timing impact on
40
VDD
Inaccessible
VDD M1 Pin
M1 Pin (Extended)
M2 Track
M2 Wire
M2 Wire (Pre-Routed)
VSS VSS VIA12
(a) (b)
Figure 2.13. Pinlength extension. (a) Initial placement. (b) Pinlength extension.
each cell. Furthermore, our framework minimizes the effective number of alternative master cells
by the perturbation-minimized optimization described in Section 2.2.2. Note that this simple
pin-extension satisfies all the conditional design rules for generating standard-cell libraries that
are described below in Section 2.2.3. Figure 2.13 illustrates pinlength extension that resolves
a DRC caused by the inaccessible pin (Figure 2.13(a)) by extending the instance’s I/O pins
(Figure 2.13(b)) without detailed placement adjustments. We only apply the pin-extension for
M1 pins because (i) we do not allow routing on M1 layer, and (ii) M2 pins are directly accessible
from BEOL layers.
Adaptive Perturbation. We utilize instance timing slacks to set the perturbation range (i.e., the
range of the vertical and horizontal adjustments) by considering timing margins of each instance.
For the instances with the worst slacks, we fix the placement and the routed wires to prevent
the deterioration of timing characteristics. For the rest of the instances, the applicable range
of perturbation is set by the input parameter settings. Flipping of instances and extension of
pinlengths are allowed for all instances except for the fixed instances.
2.2.2 Switchbox Generation
CoRe-ECO generates a switchbox by extracting the instance, pin, net and obstacle
information from each local window covering DRCs. Figure 2.14(b) visualizes the generated
switchbox representation from the local window depicted in Figure 2.14(a).
Instances. We separate the instances in the refinement region (i.e., blue dotted box) according
41
M2 Short
(a)
I0 E2 E4 E5 I1 I2 I3
E0 P0 P4 P5 P8
E1 P1 P6
P2 P7
P3
E3 E6
(b)
Net Information
Refinement Region Internal Pins Net0 : P0 - E0
Obstacle Region External Pins Net1 : P1 - E1
Net2 : P2 - E2
Adjustable Instance Obstacle Metal Element
Net3 : P3 - P7 - E3
Fixed Instance Net4 : P4 - P8 - E4
Net5 : P5 - E5
Clipped Instance Net6 : P6 - E6
Figure 2.14. Switchbox generation. (a) Local window with one DRC, displayed by the commer-
cial tool [144]. (b) Visualization of the generated switchbox.
42
to the given P&R results. We first fix the placement and the corresponding pre-routed nets of
the instances (i.e., I2) whose timing slacks are less than a predefined worst-slack upper bound.
Then, the instances inside the region are extracted as adjustable instances (i.e., I0, I1) that are
allowed for the refinement operations. The clipped instances (i.e., I3) are partially included in
the refinement region. Therefore, those instances are not adjustable, but their partial pins/nets
are extracted for the routing optimization.
Pins/Nets. The I/O pins in the adjustable and clipped instances are extracted as internal pin
candidates (i.e., P0 - P8). The external pins (i.e., E0 - E6) are extracted along the boundary of the
refinement region if there is a connection from the internal pins to the outside of the refinement
region. The net information defines new interconnections between internal pins and external
pins.
Obstacle Elements. The switchbox has two types of obstacle elements (i.e., gray rectangles).
First, we extract the routed metal elements in the obstacle region (i.e., yellow solid box) and set
those elements as obstacles to check for DRCs on the boundary of the refinement region. Second,
we consider the routed elements inside the refinement region as obstacles if those elements do
not have any connections to the internal/external pins or they are connected to the fixed instances.
Perturbation-Minimized Optimization
The proposed CoRe-ECO has multiple objectives associated with the refinement opera-
tions and routing problems. To honor the given (initial) detailed placement and detailed routing
solution, we minimize the perturbations made by refinement operations as well as the total metal
length. The vertical adjustment (∆V) and horizontal adjustment (∆H) are respectively defined as
the total amount of vertical and horizontal shifts of the adjustable instances as shown in (2.24)
and (2.25). The horizontal flipping (∆F) is defined as the total number of flipped instances as
shown in (2.26). The pinlength extension (∆P) is defined as the sum of extended pinlengths as
shown in (2.27). The routing (ML) is the sum of routed VIA/Metal elements (i.e., VIA12, M2,
VIA23 and M3). Each element has the same weight in the calculation of ML because we separate
43
the objective functions for each type of element as shown in (2.28). CoRe-ECO simultaneously
optimizes these multiple objectives in light of the “lexicographic” order described in (2.29). In
other words, the objectives are optimized according to the priority order given by LexMin; for
each given objective, this effectively induces a single-objective optimization problem under the
constraining condition that optimizes the higher-priority objectives.
Vertical Adjustment (∆V) : ∑ ( yt − ytorg ) (2.24)

t∈T
Horizontal Adjustment (∆H) : ∑ ( xt − xtorg ) (2.25)

t∈T
Horizontal Flipping (∆F) : ∑ lt (2.26)

t∈T

Pinlength Extension (∆P) : ∑ w p,r × m p,r , (2.27)
e p,r ∈E
∀p ∈ Pt , ∀t ∈ T, ∀r ∈ aext (p)
Routing (ML{#VIA12, #M2, #VIA23, #M3}) : ∑ mv,u (2.28)

ev,u ∈E
E=Sets of each VIA12, M2, VIA23, and M3 Element
44
𝑙𝑏𝑟1 ≤ 𝑥𝑡 ≤ 𝑢𝑏𝑟1 − 𝑤𝑡 , if 𝑦𝑡 = 1
𝑟=1 t
Clipped Clipped
s t s Cell 𝑙𝑏𝑟1 𝑢𝑏𝑟1 Cell
𝑟 =0 s
Clipped Clipped
𝑥𝑡 − 𝑤𝑠 ≥ 𝑥𝑠 𝑥𝑡 + 𝑤𝑡 ≤ 𝑥𝑠 Cell 𝑙𝑏𝑟0 𝑢𝑏𝑟0 Cell
𝑙𝑏𝑟0 ≤ 𝑥𝑠 ≤ 𝑢𝑏𝑟0 − 𝑤𝑠 , if 𝑦𝑠 = 0
(a) (b)
Figure 2.15. Placement constraints. (a) Relative positions between two instances in the same
placement row. (b) Boundary condition of each placement row.
LexMin: (a) ∆V, (b) ∆H, (c) ∆F, (d) ∆P,
(e) ML {(1) #VIA12, (2) #M2, (3) #VIA23, (4) #M3} (2.29)
SMT Formulation
Placement Formulation. We utilize the conventional floorplanning approach (i.e., Relative

Positioning Constraint (RPC)) for the placement problem [115]. All instance positions in a
specific placement row can be represented by two RPCs as shown in Figure 2.15(a). At least
one of the two inequalities holds for each pair t̸=s through the SMT expression described in
Algorithm 1. The maximum adjustable column boundary of each instance t is determined by the
placement row yt due to the different composition of clipped instances in each row as shown
in Figure 2.15(b). These geometric conditions determine the position and the flip status of the
instance.
Dynamic Cell Allocation (DCA). Every pin in each instance has its corresponding flow capacity
variable Cmn (p, r) for certain net n and commodity m on the corresponding vertices of the
placement grid, according to the shape and relative position of the pin in the instance as well as
the possible adjustment range of each instance (see Figure 2.16(a)). When locations of instances
are determined by the placement formulation, the flow capacity variables of each instance’s
45
Algorithm 1. Set RPC constraint (instances t, s)
1: if yt = ys then ▷ t and s are on the same placement row
2: if xt ≥xs + ws then
3: xt ≥ xs + ws ; ▷ t is on the right side of s
4: else if xt + wt < xs then
5: xt + wt ≤ xs ; ▷ t is on the left side of s
6: else
7: Unsatisfiable condition;
8: end if
9: end if
VDD
1 1 1 1 1 1 1 1 1 1 1 1 1 1
PIN1
Placement Grid
(a) (Instances)
Clipped 1 1 1 1 1 1
Shift/Flip
1 1 1 1 1 1 1 1 Clipped
(Fixed) (Fixed)
PIN2
2 2 2 2 2 2 2 2 2 2 2 2 2 2
2 2 2 2 2 2 2 2 2
PIN32 2 2 2 2
3 3 3 3 3 3 3 3 3 3 3 3
Instance VSS Flow Capacity
Shifting/Flipping Variables
VDD
PIN1
1
PIN1
DCA Clipped 1 Clipped

(b) (Fixed)
Flow Capacity = 0
(Fixed)
PIN2
(Dynamic Cell Allocation) 2

2
PIN3
3 3
VSS
Routing Pin VDD
Allocation
PIN1
1
Routing Grid
PIN1
Clipped 1 Clipped
(c) (Nets) (Fixed) 2 (Fixed)
PIN2
2
PIN3
3 3
Routing Pins
VSS
Figure 2.16. Dynamic cell allocation (DCA).
pins are conditionally assigned to the corresponding locations according to the placement status
of each instance (i.e., shifted, flipped) as described in Algorithm 2. First, the coordinates of
each pin are determined by the location of each instance and flip status (Lines 1-6). Then, all
capacities Cmn (p, r) outside the range of each pin are assigned to zero (Lines 7-11) as depicted in
Figure 2.16(b).
46
Algorithm 2. Set flow capacity control constraint (Cmn (p, r))
/* x coordinate (resp. placement row) of a routing grid r: xr (resp. yr ) */
/* x coordinate (resp. placement row) of a pin p: x p (resp. y p ) */
/* p is either source or sink of a net n and commodity m */
/* origin’s column, origin’s row, flipping of a instance i: ix ,iy ,i f */
/* x offset from the instance origin of a pin p: ox p (resp. oy p ) */
1: if i f = false then
2: x p = ix + ox p ;
3: else
4: x p = ix − ox p ;
5: end if
6: y p = iy + oy p ;
7: if (xr ̸= x p ) | (yr ̸= y p ) then
8: Cmn (p, r) = 0;
9: else
10: Cmn (p, r) is determined by routing formulation;
11: end if
fmn (v = p, u = r) ≤ Cmn (p, r), ∀r ∈ a(p), ∀r ∈ V0 (2.30)
Equation (2.30) associates the flow variable fmn (v, u) with the flow capacity variable Cmn (p, r).
Each fmn (v, u) is determined by the routing formulation when vertex v is the internal pin p, and
the adjacent vertex u is the adjacent vertex r of p in M1 and M2 (i.e., V1 ,V2 ). This enables our
routing formulation to recognize the feasible sets of r in V1 ,V2 layers as routing pins, as depicted
in Figure 2.16(c).
Pinlength Extension. We generate flow variables for extendable pin candidates to enable
pinlength extension when finding a routable solution, as illustrated in Figure 2.17. The extendable
pins are generated in both up/down directions from the uppermost/lowermost vertices of each
I/O pin. We set different weights for the extendable pin candidates, proportional to their distance
from the nearest I/O pins. These weights are used as the priority in our objective function
(Equation (2.27)) for minimizing the total length of the extended pins.
47
Weight = 1
M1 I/O Pin
Pin candidates of original master cell
Pin candidates extendable by length of 1 (= weight 1)
Pin candidates extendable by length of 2 (= weight 2)
Weight = 2
Figure 2.17. Weighted extendable pin candidates.
Routing Grid Metal Via
MAR=3 EOL=3 VR=1
MAR=2 EOL=2
VR=1.5
MAR=1 EOL=1
(a) (b) (c)
Figure 2.18. Grid-based conditional design rules: (a) MAR, (b) EOL and (c) VR.
Routing Formulation. We use multi-commodity network flow and conditional design rules
to formulate the detailed routing problem, following the same principles as [100]. The flow
formulation finds a routing path between the source and the sink for each commodity. Specifically,
refined constraints for commodity flow conservation and vertex exclusiveness in uni-directional
edges are implemented in our framework to reduce the search space of the routing formulation.
The conditional design rules work as constraints to route using design-rule violation-free paths.
CoRe-ECO implements three fundamental grid-based design rules,10 namely, Minimum Area
(MAR), End-of-Line Spacing (EOL) and Via Rule (VR), as illustrated in Figure 2.18. MAR
(Figure 2.18(a)) defines the minimum number of grids that should be covered by the metal
segments. EOL (Figure 2.18(b)) defines the minimum number of grids between two metal
segments. VR (Figure 2.18(c)) defines the minimum distance (in L2 norm) between vias.
10 Inthis work, we assume sub-7nm technologies that are based on Extreme Ultraviolet (EUV) lithography as in
the previous work [24]. However, our framework is applicable to additional multi-pattern-aware design rules, such
as Parallel Run Length (PRL) and Step Height Rule (SHR).
48
Table 2.5. Standard-cell architectures.
Cell architecture Design constraint

#Fin #Routing tracks Cell height Design rule Pin accessibility
3 6 8T EUV-Loose (EL) MPO3
3 6 8T EUV-Tight (ET) MPO3
MPO2
EUV-Loose (EL)
MPO3
2 4 6T
MPO2
EUV-Tight (ET)
MPO3
We have implemented the proposed CoRe-ECO framework in Perl/SMT-LIB 2.0 and

validated on a Linux workstation with Intel (R) Xeon E5-2560L 1.8GHz processor and 128GB
memory. The SMT Solver Z3 (version 4.8.5) [34] is used to produce the optimized solution.
Experimental Setup
Standard-Cell Library Preparation. Using an SMT-based cell layout automation [169], we

prepare six types of standard-cell libraries with various cell architecture and design constraints
presented in Table 2.5. We adopt two design rule sets that comprise combinations of specific
design rule settings, as follows. EUV-Loose (EL) consists of MAR/EOL/VR = 1/1/1. EUV-Tight
(ET) consists of MAR/EOL/VR = 1/2/1, inspired by [24]. We also generate two different types
of cell libraries ensuring at least two and three I/O pin access points (i.e., MPO2 and MPO3).
Then, for design enablement, we convert the primitive layout solutions of the SMT to LEF
format. We assume contacted poly pitch (CPP), metal pitch (MP), and cell height of 40, 40 and
280nm, respectively.11 We also generate three additional LEFs that have cells with pinlengths
extended by 1, 2 and 3 grids, respectively, consistent with the cell height of the corresponding
standard-cell library.
Place-and-Route (P&R). We validate our framework by using four open-source RTL designs
AES, JPEG, LDPC [161] and IBEX [160]. We utilize M2-M7 layers as back-end-of-line (BEOL).
11 Since the layouts are fully grid-based, we consider the CPP and MP as pitches of grids.
49
Table 2.6. Experimental statistics. RT = the number of routing tracks, DRSet = Design rule set, Impr./Incr. = Improvement/Increment
ratio over ECO routing, #V/#H/#F/#P = the number of refined instances by vertical shifting/horizontal shifting/horizontal flip-
ping/pinlength extension.
Proposed CoRe-ECO refinement
Design Cell library ECO routing
#Remaining DRCs WL #Refinements #ECO Runtime
Name #Nets #Cells RT DRSet MPO #DRCs WL(µm) #DRCs Impr. WL(µm) Incr. #V #H #F #P round (h)
13,958 13,694 6T ET 3 40 46,272.16 2 95.0% 46,262.73 -0.020% 2 17 5 36 3 9.2
13,912 13,648 4T ET 2 74 45,968.76 13 82.4% 45,958.58 -0.022% 3 32 31 60 5 9.9
AES 13,938 13,674 4T EL 3 41 45,797.96 0 100.0% 45,781.98 -0.035% 2 19 12 28 4 4.3
13,802 13,538 4T EL 2 146 45,035.54 56 61.6% 45,025.73 -0.022% 8 51 23 86 5 16
13,867 13,603 4T ET 3 153 44,548.12 44 71.2% 44,528.05 -0.045% 25 104 45 79 4 18.2
70,543 70,518 6T ET 3 79 144,905.98 30 62.0% 144,892.66 -0.009% 0 12 15 51 3 13.2
50
71,491 71,466 4T EL 2 155 134,901.91 73 52.9% 134,893.09 -0.007% 3 30 14 73 4 25.3
JPEG 71,177 71,152 4T EL 3 37 133,558.22 18 51.4% 133,555.19 -0.002% 2 14 8 18 3 2.8
70,932 70,907 4T ET 2 198 132,739.50 132 33.3% 132,738.24 -0.001% 4 22 12 49 3 41.4
70,008 69,983 4T ET 3 68 136,852.58 42 38.2% 136,849.64 -0.002% 1 13 5 12 2 8.7
57,133 55,081 6T EL 3 21 732,917.12 8 61.9% 732,908.83 -0.001% 2 5 3 7 2 1.3
LDPC 57,138 55,086 4T ET 2 12 737,957.04 7 41.7% 737,955.47 0.000% 0 3 1 1 1 0.24
57,106 55,054 4T EL 3 90 751,812.13 53 41.1% 751,797.69 -0.002% 7 19 7 9 2 5.3
15,540 12,225 4T EL 2 166 48,734.35 56 66.3% 48,730.00 -0.009% 12 40 16 80 5 15.8
15,432 12,117 4T EL 3 62 48,426.72 6 90.3% 48,425.94 -0.002% 3 11 9 8 3 4.8
IBEX
15,502 12,187 4T ET 2 141 47,042.10 80 43.3% 47,045.31 0.007% 3 18 10 48 4 13.9
15,179 11,864 4T ET 3 93 49,346.74 32 65.6% 49,367.76 0.043% 0 9 14 13 3 5.1
Average 92.7 195695.11 38.4 58.6% 195689.23 -0.003% 4.5 24.6 13.5 38.7 3.3 11.5
Figure 2.19. Overall ECO flow using CoRe-ECO framework.
We assume power/ground pins on the M1 layer for the initial detailed routing. However, since
our proposed framework targets grid-based architecture and correction of the pin accessibility-
induced DRCs, we focus on M2-M3 layers assuming that M2 and M3 layers have the same
metal pitches. Two commercial tools [144] [172] are used to generate the initial P&R layouts
and to execute the following ECO routing. In the commercial tool, we perform 20 iterations
of ECO routing until the commercial tool is unable to further reduce the number of DRCs (i.e.,
#DRCs) for most of the benchmark cases. The blue lines in Figure 2.20 show the example trends
of #DRCs through the ECO routing iterations for four representative cases from Table 2.6 (i.e.,
cases in bold). Note that we compare our work with the results of ECO routing because we are
not able to fairly compare our work with the previous works, [35] [66] [72] [76] [88] [108] [130],
due to (i) the different target design stage (i.e., detailed placement optimization vs. ECO) and (ii)
the lack of exact experimental settings.
51
Figure 2.20. Example trends of the number of DRCs by the ECO routing and the proposed
CoRe-ECO iterations.
Setting up the Perturbation Range. We set the 1% of worst slack cells and the routed nets
connected to those cells as fixed instances and obstacle elements, respectively. For the rest of
the instances, we set the perturbation range of the vertical and horizontal adjustments to two
placement rows and eight poly pitches, respectively.
Design of Experiments
Figure 2.19 illustrates an overview of the ECO flow utilizing our CoRe-ECO framework.
Given a cell library and initial detailed P&R result, the new ECO round starts with converting
this layout information to a pinLayout format for the proposed framework in the LEF/DEF
Conversion step. Then, if there exist any remaining DRCs in the non-overlapping regions, we
52
rip up the region and generate a switchbox representation in the Switchbox Generation step.
Note that the switchboxes in the same ECO round cannot overlap because our framework could
change the P&R in each switchbox; the framework also refers to the horizontal/vertical obstacle
regions for checking of design rules on the boundary of each switchbox. Given the switchbox
representation, we generate an SMT code and solve the problem through the CoRe-ECO SMT
Code Generation and SMT Solving steps. We iterate these refinement steps until we find a
routable solution or there are no remaining DRCs or feasible switchboxes in non-overlapping
regions. After the iterations, in the SMT Solutions to DEF Conversion step, we apply the
DRC-clean solutions to the original DEF and generate a revised DEF for the next ECO round or
publish as the final ECO result.
The ECO flow described above is fully automated, and each sequence (i.e., switchbox
generation to SMT solving) can be executed in parallel through the multi-threaded operation. In
this work, up to 24 threads are used for all testcases. For each DRC, our framework examines
multiple switchboxes of various sizes (i.e., 10 - 25 vertical tracks and 1 - 5 placement rows)
for several relative locations to the target DRC. The size of a routable switchbox for each DRC
varies according to the existing P&R results, routing congestion, and locations of neighboring
DRCs. To minimize the perturbation of the placement as well as the runtime of the SMT Solving
step, our framework increases the size of the switchbox from the minimum (i.e., 10 vertical
tracks × 1 row) to the maximum (i.e., 25 vertical tracks × 5 rows) in the Switchbox Generation
step until it finds a routable solution or fails.
Experimental Results
Statistics on the Proposed ECO Flow. Table 2.6 summarizes the experimental statistics of the
proposed ECO flow for benchmark cases which consist of four base design circuits synthesized
with various cell libraries described in Table 2.5. In Table 2.5, #VTrack, #Row, #DRC, #Net and
#Pin denote the number of vertical tracks, placement rows, DRCs, Nets and Pins in Switchbox,
respectively. #TotalInst., #Adj.Inst. and #FixedInst. denote the number of total, adjustable and
53
Table 2.7. Detailed experimental results of the proposed ECO flow (AES, 4T, ET, MPO2, 74
DRCs).
Switchbox #Cell refinement
ECO SMT
Index #Total #Adj. #Fixed
round #Vtrack #Row #DRC #Net #Pin #V #H #F #P runtime(s)
Inst. Inst. Inst.
0 18 2 1 5 3 2 15 33 0 2 0 1 11.0
1 19 1 2 4 3 1 9 22 0 0 0 2 1.3
2 18 2 1 4 3 1 13 29 0 1 1 1 8.5
3 14 1 1 3 2 1 6 17 0 0 0 1 0.6
4 14 1 2 3 2 1 6 15 0 0 0 0 0.6
5 21 2 1 7 3 4 15 43 0 0 0 3 25.4
6 17 1 1 3 3 0 7 17 0 2 0 1 3.4
7 15 2 1 4 2 2 9 21 0 1 0 0 28.8
8 14 1 1 3 2 1 8 17 0 0 0 1 0.7
9 14 1 1 2 0 2 6 14 0 0 0 0 0.4
10 18 2 1 8 5 3 16 41 0 0 3 3 11.8
11 1st 16 1 1 2 1 1 6 13 0 0 0 1 0.6
12 15 1 2 4 2 2 6 16 0 0 1 2 0.7
13 22 2 3 9 6 3 19 42 0 3 2 5 20.2
14 16 1 1 3 3 0 8 18 0 2 1 2 1.4
15 14 1 1 2 2 0 5 11 0 1 0 0 0.4
16 14 1 1 2 1 1 4 12 0 0 0 0 0.5
17 14 1 1 2 2 0 4 10 0 0 1 0 0.4
18 14 1 1 2 1 1 4 9 0 1 0 0 0.4
19 14 1 1 2 2 0 6 14 0 0 0 0 1.0
20 18 2 1 6 3 3 16 38 0 0 1 2 5.7
21 14 1 1 3 2 1 6 14 0 2 1 0 0.5
22 16 1 1 3 2 1 7 16 0 1 2 2 0.8
23 18 1 1 4 3 1 10 25 0 0 1 1 2.2
24 14 1 1 3 2 1 7 15 0 0 0 1 0.7
25 14 1 1 2 1 1 3 7 0 0 0 0 0.3
26 15 2 1 5 2 3 12 29 0 0 0 0 13.7
2nd
27 14 1 1 2 1 1 7 18 0 0 1 1 0.8
28 14 1 1 2 1 1 5 11 0 1 0 0 0.4
29 19 2 1 7 6 1 15 42 0 1 1 3 54.6
30 16 2 1 5 4 1 16 34 0 1 0 0 5.6
31 29 2 1 10 8 2 27 63 0 1 1 5 588.9
32 18 5 2 13 11 2 34 77 0 0 3 2 859.9
33 17 2 7 7 5 2 14 38 2 3 2 2 6.3
34 19 3 3 10 7 3 24 54 0 1 1 4 109.1
3rd
35 20 3 1 9 8 1 25 67 0 1 4 4 1465.8
36 14 3 3 7 5 2 16 41 0 0 0 1 10.8
37 17 2 1 5 2 3 18 39 0 1 0 2 8.3
38 18 5 3 13 7 6 33 71 1 2 2 3 1841.8
39 14 1 2 2 2 0 5 12 0 2 1 1 0.6
4th
40 21 2 1 9 6 3 16 43 0 0 1 2 30.7
41 5th 16 2 1 7 3 4 12 28 0 2 0 1 6.5
Average 16.6 1.7 1.5 5.0 3.3 1.6 11.9 28.5 0.1 0.8 0.7 1.4 122.2
Total 61 208 139 69 500 1196 3 32 31 60 5131.84
fixed or clipped instances in Switchbox, respectively. #V, #H, #F and #P denote the number of
refined instances by Vertical Shifting, Horizontal Shifting, Horizontal Flipping and PinLength
54
Extension, respectively. Column “ECO routing” represents the total number of DRCs (i.e.,
#DRCs) in target layers (i.e., M2-M3) and the total wirelength (i.e., WL) after the 20 iterations
of ECO routing. The target DRCs mainly include “Cut Spacing” on M1-M2 layers and “Metal
End-of-Line Spacing”, and “Metal Short” on M2-M3 layers. Our CoRe-ECO framework reduces
the remaining DRCs after ECO routing by 58.6% on average, with reductions ranging from
33.3% to 100.0%. Figure 2.20 shows the trend of #DRCs versus iterations of ECO routing
(i.e., the blue line), along with the following CoRe-ECO flow (i.e., the orange line) for four
representative cases of each base benchmark circuit from Table 2.6 (i.e., cases in bold). The
figure demonstrates that our framework can further improve the routability with the concurrent
cell refinements and the routing optimization. The reduction of the average total wirelength
by 0.003% shows that our framework has successfully minimized the wirelength despite the
refinement of cell placement and pinlength extension. We observe that the number of ECO
rounds and the total runtime depend on #DRCs and the benchmark circuit configurations. For all
benchmark cases, CoRE-ECO performs 3.3 ECO rounds on average with an average runtime of
11.5 hours.
Table 2.7 presents the detailed refinement results of the AES benchmark circuit with
4T/ET/MPO2 cell library and 74 DRCs. A total of five ECO rounds with 42 switchboxes
have been performed to resolve 61 out of all 74 DRCs. The average number of vertical tracks,
placement rows, and DRCs in the switchboxes are 16.6, 1.1, and 1.5, respectively. Each
switchbox includes 5.0 total/3.3 adjustable/1.6 fixed or clipped instances and 11.9 nets/28.5
pins on average. Through the five rounds of ECO flow, 126 out of 139 adjustable instances
have been perturbed in the placement or the length of pins. 6 out of 42 ECO cases have been
fixed without any perturbation. And 7 and 7 cases require the extension of pinlengths or the
re-placement of instances, respectively. The remaining 22 cases are routable by only changing
both instance placement and pinlengths. The average runtime per switchbox is less than 3
minutes and switchboxes up to 18×34 vertical/horizontal tracks, 7 adjustable instances, 33 nets
and 71 pins (i.e., Index 38) have been solved within 31 minutes. Figure 2.21 shows the reduction
55
Figure 2.21. DRC reductions by CoRe-ECO rounds for AES (4T/ET/MPO2).
of DRCs in full-chip layouts by multiple ECO rounds utilizing CoRe-ECO framework, displayed
by a commercial tool [144]. The yellow circles indicate regions with DRCs.
Example Refinement Operations. Figure 2.22 shows an example of refinement operations in
our proposed ECO framework. Figure 2.22(a) depicts a switchbox of index 38 case in Table 2.7.
The switchbox consists of 7 adjustable / 6 clipped (i.e. fixed) cell instances with 3 ‘M3 Short’
DRCs in 18×34 vertical/horizontal tracks. Figure 2.22(b) illustrates the DRC-clean solution with
the refinement operations (i.e., placement adjustment and pinlength extension) and the optimized
routing in terms of the metal length. Note that the elements in gray color represent the obstacles
inside the switchbox and that M1 I/O pins are not displayed in Figure 2.22(b). The pre-routed
56
Figure 2.22. Example of refinement operations in the proposed ECO flow (Index 38 case in
Table 2.7). (a) Switchbox with three DRCs. (b) Routable solution with placement adjustments
and pinlength extensions.
wires, that (i) are connected to the fixed instances or (ii) have no internal connection inside the
switchbox or (iii) exist outside the switchbox, are regarded as obstacles.
• Placement Adjustment. The placement of the instance I2 in Figure 2.22(b) has been
adjusted from the placement row 0 to row 1 and horizontally shifted from the vertical track
12 to track 10. The instance I0 has shifted in the same placement row from the vertical
track 8 to track 10. Instances I0 and I5 have been flipped on the same placement locations.
57
• Pinlength Extension. The I/O pins of the instance I1, I3 and I5 in Figure 2.22(b) have
been extended by 1-2 grids to maximize the pin accessibility. In the proposed ECO flow,
the respective master cell of each of these instances is replaced with the additional master
cell with extended pinlengths, in the SMT Solutions to DEF Conversion stage.
2.2.4 Conclusion
We have described a novel concurrent refinement framework for the automated ECO
flow. Our framework provides simultaneous and perturbation-minimized refinements of detailed
placement-, detailed routing- and cell-optimized layout solutions to address the DRCs during
the ECO stage. By ripping up and refining a local window of the whole layout design, CoRe-
ECO is capable of achieving a DRC-clean layout solution. We have demonstrated that our
framework successfully resolves an average of 58.6% (range: 33.3% to 100.0%) of remaining
post-ECO route DRCs on M1-M3 layers, across a range of benchmark circuits with various cell
architectures, with no adverse effect on total routed wirelength (average of 0.003% reduction).
2.3 Acknowledgments
Chapter 2 contains a reprint of Hamed Fatemi, Andrew B. Kahng, Minsoo Kim and
Jose Pineda de Gyvez, “Optimal Bounded-Skew Steiner Trees to Minimize Maximum k-Active
Dynamic Power”, Proc. International Workshop on System-Level Interconnect Problems and
Pathfinding, 2020; and Chung-Kuan Cheng, Andrew B. Kahng, Ilgweon Kang, Minsoo Kim,
Daeyeal Lee, Bill Lin, Dongwon Park and Mingyu Woo, “CoRe-ECO: Concurrent Refinement
of Detailed Place-and-Route for an Efficient ECO Automation”, Proc. IEEE International
Conference on Computer Design, 2021. The dissertation author is a main contributor to, and a
primary author of, each of these papers.
I would like to thank my coauthors, Professor Chung-Kuan Cheng, Dr. Hamed Fatemi,
Professor Andrew B. Kahng, Dr. Ilgweon Kang, Dr. Daeyeal Lee, Professor Bill Lin, Dr.
Dongwon Park, Dr. Jose Pineda de Gyvez and Mingyu Woo.
58
Chapter 3
Technology-Aware Physical Design Method-

ology
This chapter presents two technology-aware physical design methodologies. The first
work addresses local layout effects (LLE) in physical design. In a standard cell-based design, a
diffusion break (DB) isolates neighboring devices, and the type of DB (single or double diffusion
break) used in the standard cells can significantly impact device performance and leakage. The
2nd DB effect is a type of LLE in which the distance to the second-closest DB affects device
performance. We develop 2nd DB-aware leakage optimization and detailed placement heuristics.
The second work addresses the high resistance of interconnects which causes severe supply
voltage (IR) drop problems at advanced nodes. To solve this problem, pre- or post-placed power
staples are inserted in pin-access layers to connect adjacent power rails and reduce power delivery
network (PDN) resistance, at the cost of reduced routing flexibility, or reduced power staple
insertion opportunity. We propose dynamic programming-based single- and double-row detailed
placement optimizations to maximize the power staple insertion in a post-placement flow. We
further propose metaheuristics to improve the quality of results.
59
3.1 Diffusion Break-Aware Leakage Power Optimization
and Detailed Placement in Sub-10nm VLSI
With aggressive lateral scaling of advanced CMOS technology, many challenges of

standard-cell architectures, floorplan, placement, routing and timing signoff have been introduced
into leading-edge IC design. Design today is challenged by numerous complex front-end-
of-line (FEOL) and back-end-of-line (BEOL) design rules and layout restrictions. Moreover,
aggressive area reduction via the standard-cell architecture leads to a breakdown of the traditional
assumptions of composability and independence of cell-based layout. Notably, in sub-10nm
VLSI, a given cell instance’s timing and power can be strongly affected by the cell’s neighbors
in the placement; this is commonly referred to as local layout effect (LLE). Since standard-cell
library models are provided for the cell itself, without regard to its neighbors in the placement,
LLE induces performance modeling errors and/or added margins in signoff. Today’s advanced-
node design and signoff technologies must now carefully take LLEs into account.
New Standard-Cell Architecture with Diffusion Breaks. A diffusion break (DB) isolates two
adjacent devices and is one of several critical aspects related to LLEs. For example, Slide 44
of the DAC-2018 tutorial [61], referring to such sources as [4], notes how continuous diffusion
and diffusion breaks are an important lever for shrinking standard-cell area – at the cost of
more complex design rules and LLEs. In modern sub-10nm enablements, there are two types of
diffusion breaks: single diffusion break (SDB) and double diffusion break (DDB). SDBs occupy
a smaller space than DDBs do, but incur greater LLE. On the other hand, while DDBs require
more area and incur more leakage than SDBs, they make devices faster and are more immune
to layout context. In this work, we ignore the 2nd DB effect for DDB cells due to its negligible
impact on leakage and timing. We study two types of standard cells, i.e., SDB cells and DDB
cells. An SDB (resp. DDB) cell has an SDB (resp. a DDB) at both its left and right edges. In the
following, we call a design with only SDB cells as Type-I, and a design with both SDB and DDB
cells as Type-II.
60
Figure 3.1. Examples illustrating DB types and placement constraints. (a) Layout with four
SDB cells (Type-I). (b) Mixed-DB layout with three SDB cells plus one DDB cell (Type-II).
Figure 3.1 shows two legal placements with SDB and DDB cells. SDB and DDB cells
can be placed within a single block as long as: (i) SDB (resp. DDB) cells all align to SDB (resp.
DDB) grids, where the SDB grid has a half contacted poly pitch (CPP) offset from the DDB
grid; and (ii) there is a process-specific minimum spacing requirement (SpacingSD in Figure 3.1)
between neighboring SDB and DDB cells. In Figure 3.1, for a device in a green oval, D1 is the
distance to the first placed SDB, and D2 is the distance between the first placed SDB and the
second placed SDB. That is, D2 refers to the “2nd DB distance of an SDB cell”, and we use
the term “2nd DB effect” to indicate the threshold voltage (Vt) shift due to D2 for an SDB cell.
Note that the Vt shift due to the 2nd DB effect causes a timing path with fixed cell instances
and routing to be impacted by the neighbors of its cell instances. Since an SDB cell can have a
second placed SDB on each of its left and right sides, in this work, we use the distance to the
farther of the two placed 2nd DBs as the D2 of a cell.
61
Design Impact from the 2nd DB Effect. Table 3.1 shows results of a motivating analysis
that demonstrates how much the 2nd DB effect changes timing results and leakage power of
designs.12 We use two designs, AES and VGA, from OpenCores [161] and a reference flow of
Cadence Innovus Implementation System v17.1 [144] to generate post-P&R designs. We use
80% initial utilization and 0.35ns target clock period. We note that filler cell insertion is a critical
step that determines 2nd DB distances of cells and the impact of the 2nd DB effect. Filler cells
are inserted into empty spaces after place-and-route (P&R), typically starting with larger-width
filler cells and finishing with smaller-width filler cells so as to greedily minimize instance count.
However, filler cells with large width can create large 2nd DB distances for any neighboring
non-filler cells. So, in order to see the design impact of the 2nd DB effect, we compare two
filler cell insertion strategies: largest to smallest filler (Standard) and small filler only (Small).
As seen in Table 3.1, when we perform 2nd DB-aware analysis for designs which come from a
commercial P&R tool, 2nd DB-aware leakage is larger than 2nd DB-unaware leakage by 8.8% to
144.4%. In other words, actual leakage can be much larger than the leakage reported by signoff
analysis that does not take into account the 2nd DB effect. Further, after 2nd DB-aware timing
analysis, setup worst timing slacks are improved or similar, and hold worst timing slacks are not
changed, due to small 2nd DB impacts for hold corner. Overall, this preliminary study shows
that there is substantial model-hardware miscorrelation in terms of leakage power because the
2nd DB effect is not comprehended in cell library models. Since the physical location of the 2nd
DB matters when tracing timing impacts of optimizations, the physical layout design and signoff
flows should be aware of the 2nd DB effect to avoid this model-hardware miscorrelation.
In what follows, we focus on mitigation of analysis error – in particular, underestimation
of leakage – that is caused by unawareness of the 2nd DB effect. We assume a place-and-route
methodology that inserts only Small fillers, to a priori minimize the leakage impact of the 2nd
DB effect. We only consider setup timing slacks during our optimization.
This Work. In this work, we study design impacts caused by the 2nd DB effect and propose
12 Details of our modeling of 2nd DB effect are given in Section 3.1.2.
62
Table 3.1. Preliminary study of the 2nd DB effect. Design nomenclature specifies design name
(AES or VGA), filler insertion strategy (Small (sm) or Standard (std)), and design type (Type-I
or Type-II). WS denotes the worst slack of a design.
Design Hold WS (ns) Setup WS (ns) Leakage (mW) (∆%)

AES-sm-I 0.108 / 0.108 -0.056 / -0.047 0.387 / 0.742 (91.7%)
AES-sm-II 0.113 / 0.113 -0.055 / -0.055 0.419 / 0.456 (8.8%)
VGA-sm-I 0.080 / 0.113 0.016 / 0.022 1.486 / 3.186 (114.4%)
VGA-sm-II 0.086 / 0.086 -0.002 / -0.002 1.523 / 2.439 (60.1%)
AES-std-I 0.108 / 0.108 -0.056 / -0.045 0.387 / 0.909 (134.9%)
AES-std-II 0.113 / 0.113 -0.055 / -0.054 0.419 / 0.474 (13.1%)
VGA-std-I 0.080 / 0.080 0.016 / 0.024 1.486 / 3.632 (144.4%)
VGA-std-II 0.086 / 0.086 -0.002 / -0.001 1.523 / 2.640 (73.3%)
Figure 3.2. Target of our work. We seek to minimize actual, 2nd DB-aware leakage power with
no timing degradation.
a 2nd DB-aware leakage optimization that uses cell relocation, gate sizing, Vt swapping and
DB (Type) swapping to mitigate the P&R tool’s lack of awareness of 2nd DB effect. Our main
contributions are summarized as follows.
• We study design impacts of the 2nd DB effect for designs using only SDB cells (Type-I)
and both SDB and DDB cells (Type-II).
• We propose a 2nd DB-aware leakage optimization and placement methodology that uses re-
location, gate sizing, Vt swapping and DB swapping. Our algorithm honors the placement
grid for each DB and a specified spacing constraint (SpacingSD in Figure 3.1) between
SDB and DDB.
63
• We achieve 80% leakage recovery (i.e., of the leakage increment seen with correct, 2nd
DB-aware analysis) on average without changing design performance.
Figure 3.2 conceptually illustrates the target of our work. We aim to recover leakage
increments caused by the 2nd DB effect without any worsening of negative timing slacks.
The remainder of this chapter is organized as follows. Section 3.1.1 describes previous
related works. Section 3.1.2 presents our problem statement, leakage model and optimization
approach. Section 3.1.3 describes our experimental setup, key experiments and results. We
conclude in Section 3.1.4.
3.1.1 Related Work
We categorize relevant previous works as: (i) gate sizing and (ii) local layout effect and
placement.
Gate Sizing. Traditional gate sizing problems have been studied for many years. Objectives of
the problems are to find the gate width and threshold voltage for each cell so that a circuit can
achieve optimized timing, power, and area as a result. Gupta et al. [44] describe a sensitivity-
based gate sizing algorithm to reduce leakage power. Hu et al. [57] and Kahng et al. [65]
propose global timing recovery and leakage power reduction with a sensitivity-guided greedy
sizing algorithm. Wei et al. [122] minimize total power using gate sizing and Vt assignment.
Luo et al. [90] propose a combined methodology with placement and gate sizing. The work
of [90] minimizes power by sequential placement, sizing and Vt swapping stages with a slack
management scheme. The works of [62] [93] suggest heuristics for co-optimization of sizing
and placement with minimum implant area constraints.
Local Layout Effect and Placement. In sub-10nm, various works address LLEs from a standard-
cell perspective. Yang et al. [128] describe LLEs of 2nd DB and gate-cut stress in 10nm high
performance mobile SOCs. Zhao et al. [134] introduce gate-cut stress induced LLEs on 14nm
FinFET, and Xie et al. [125] introduce SDBs and DDBs in actual FinFET devices. At the
same time, many recent works aim to mitigate LLEs for detailed placement. Ha et al. [46]
64
Figure 3.3. Overview of our overall optimization flow. The red box indicates steps that we
implement. A commercial P&R tool is used for all other steps in the flow.
introduce a pre-placement methodology to mitigate LLEs from a long SDB created by vertically
stacked SDB cells and a 2nd DB-aware timing analysis scheme. Berthelon et al. [6] propose a
design and technology co-optimization with strain-induced LLEs along with asymmetric and
non-rectangular active area. Han et al. [50] propose a detailed placement method to minimize
the number of abutments between cells, and Du et al. [37] deal with new placement constraints
of a 16nm technology. Han et al. [48] [52] also develop a detailed placement algorithm to reduce
the number of “steps”, i.e., diffusion height differences, between adjacent cells. Overall, our
work is distinguished from the previous works in that (i) we handle both sizing and placement
with constraints from SDB and DDB, and (ii) we mitigate leakage increments caused by the 2nd
DB effect, which is introduced as a recent type of LLE.
3.1.2 2nd DB-Aware Leakage Optimization and Placement
In this section, we first describe our problem statement and the timing and leakage model
for the 2nd DB effect. We then describe our methodology for detailed placement and leakage
65
Figure 3.4. Normalized derating values for (a) leakage power and (b) cell delay according to the
2nd DB distance.
optimization. Figure 3.3 shows the overall flow of our optimization. The input of the flow is
an optimized post-route design database produced by a commercial P&R tool. Our leakage
optimization is performed imax iterations, and is followed by incremental routing. After routing,
Vt swapping with fixed cell location recovers negative timing slacks caused by routing changes.
This incremental routing and Vt swapping is performed by a commercial P&R tool.
Problem Statement
Given a post-P&R design, our goal is to swap cells and perturb placement so as to
minimize leakage with 2nd DB awareness.
Input: A post-P&R design database from a commercial tool.
Output: An optimized design with 2nd DB awareness.
Objective: To minimize leakage power of the design.
Do: Relocation, gate sizing, Vt swapping and DB swapping.
Constraints: No total negative slack (TNS) degradation, no cell overlap, grid-based
placement for each DB, and a spacing constraint between SDB and DDB.
Modeling for the 2nd DB Effect
Delay and leakage power of devices are changed by the 2nd DB effect. We introduce
our model for 2nd DB using derating values of Vt shift from the work of [128] and our col-
laborator [149]. Figure 3.4 shows the normalized derating values for leakage power and delay.
66
Figure 3.5. Examples of 2nd DB distance change due to cell relocation and sizing. The red color
of D2C2 indicates the 2nd DB distance of C2 and the green color of D2C1 indicates the 2nd DB
distance of C1.
According to [149], we use the following sets of process, voltage, and temperature conditions:
(SS, 0.72V, -40◦C) for setup corner and (FF, 0.88V, 125◦C) for hold corner. Since the Vt shifts
for PMOS and NMOS are different, we assume that the averaged Vt shift for cells is proportional
to 2nd DB distance for SS and FF [149]. Advised by [149], we calculate the derating values
for leakage power and delay of cells by assuming that leakage power exponentially increases
with 2nd DB distance. In our library, we assume that a given DDB cell is 2CPP wider than the
corresponding SDB cell, and that the 2nd DB effect maxes out at the maximum width of standard
cells in our library. Furthermore, we follow design rules for the spacing constraint between
neighboring SDB and DDB cells. Our derated libraries are created using a 2CPP-step of 2nd DB
distance. We do not have a 2nd DB distance of 1CPP because the minimum cell width is 2CPP.
And, we also round down odd-valued 2nd DB distances to even values because libraries with a
step size of 1CPP incur more than 3× peak memory usage during our optimization.13
13When the 2nd DB-unaware leakage is 0.225mW for the AES Type-I design (60% utilization, 0.5ns clock period),
the 2nd DB-aware leakage with the 1CPP step size is 0.319mW, and that with the 2CPP step size is 0.295mW; our
optimization results (leakage recovery) are 99.0% and 97.1%, respectively.
67
Notation Meaning
mk .c Cell for the kth candidate.
mk .m Method for the kth candidate.
mk .s f Sensitivity for the kth candidate.
mk .∆leakage ∆leakage for the kth candidate.
∆leakage ∆leakage of a design.
M Set of candidates.
c.cl Left neighboring cell of cell c.
c.cr Right neighboring cell of cell c.
c.2nddb 2nd DB distance of cell c.
Wmin Minimum cell width.
Algorithm 3. 2nd DB-aware relocation algorithm.

Procedure: Relocation()
Inputs: a netlist D, a placement of D
Outputs: an optimized netlist
1: M ← 0/ ; k ← 0
2: for cell c in the netlist do
3: if cell c is neither a DDB cell nor a flip-flop then
4: if cell c is abutted with c.cl and whitespace on right side of cell c ≥ Wmin then
5: mk .c ← c ; mk .m ← move to the right by Wmin
6: update c.2nddb, c.cl .2nddb, c.cr .2nddb
7: mk .∆leakage ← ∆leakage ; M ← M ∪ mk ; k ← k + 1
8: end if
9: if cell c is abutted with c.cr and whitespace on left side of cell c ≥ Wmin then
10: mk .c ← c ; mk .m ← move to the left by Wmin
11: update c.2nddb, c.cl .2nddb, c.cr .2nddb
12: mk .∆leakage ← ∆leakage ; M ← M ∪ mk ; k ← k + 1
13: end if
14: end if
15: end for
16: while M ̸= 0/ do
17: Pick mk with the maximum mk .∆leakage in M
18: Commit mk ; M ← M\mk
19: if TNS degrades then
20: Revert
21: end if
22: end while
68
2nd DB-Aware Relocation
We first describe our sensitivity-based method to relocate cells with no TNS degradation.
Cell location is critical for the 2nd DB effect. Figure 3.5 illustrates how 2nd DB distance varies
due to cell relocation and sizing. In this figure, D2C2 denotes the 2nd DB distance of cell C2
and D2C1 denotes the 2nd DB distance of cell C1. If C2 is relocated to the right by 2CPP, then
the 2nd DB distances of both C1 and C2 decrease to 2CPP. If C2 is sized down and still abuts
C1, then only D2C1 decreases. Therefore, we can reduce the 2nd DB distance by cell relocation,
and recover leakage power. Algorithm 3 shows this relocation flow (Table 3.2 gives notations
that we use). Line 1 initializes an empty set of candidate moves. In Lines 2-15, we calculate
candidate moves per cell and the sensitivity per move. Here, we only consider moves of SDB
combinational cells, as shown in Lines 2 and 3. In Lines 4-8, for each cell, we add a candidate
right move (relocated to the right) and calculate the sensitivity (leakage recovery due to the
move) if the cell has no whitespace on the left but whitespace on the right. Similarly, in Lines
9-13, we add a candidate left move and calculate the sensitivity. In Lines 16-21, we greedily
commit the candidate move with the highest sensitivity as long as there is no TNS degradation
and the placement is legal.
2nd DB-Aware Sizing
Algorithm 4 describes our sensitivity-based methodology for gate sizing, Vt swapping

and DB swapping. Similar to the methodology of 2nd DB-aware relocation, we greedily perform
either gate sizing, Vt swapping or DB swapping for each cell according to a pre-sorted list.
Line 2 initializes an empty set of candidate moves. In Lines 3-20, we first populate the list
with all candidate moves per cell, i.e., swapping to a higher Vt, downsizing and DB swapping.
∆leakage/∆slack is used as our sensitivity function and ∆slack denotes the timing slack differ-
ence for the swapped cell. Here, we only consider swaps of combinational cells, as shown in
Line 4. In Lines 5-9, we add a candidate and perform the sensitivity calculations for downsizing
if the cell is downsizable. Similarly, Lines 10-13 and Lines 14-17 add a candidate and calculate
69
Algorithm 4. 2nd DB-aware leakage optimization.
Procedure: ReduceLeak()
Input: a netlist D, a placement of D
Output: an optimized netlist
1: Relocation(D)
2: M ← 0/ ; k ← 0
3: for cell c in the netlist D do
4: if cell c is not a flip-flop then
5: if cell c is downsizable then
6: mk .c ← c ; mk .m ← downsize
7: update c.cl .2nddb, c.cr .2nddb
8: mk .s f ← ∆leakage/∆slack ; M ← M ∪ mk ; k ← k + 1
9: end if
10: if cell c is not a highest V t then
11: mk .c ← c ; mk .m ← downV t
13: end if
14: if cell c.cl or c.cr is a different DB type from cell c then
15: mk .c ← c ; mk .m ← changeDB
16: update c.cl .2nddb, c.cr .2nddb
18: end if
19: end if
20: end for
21: while M ̸= 0/ do
22: Pick mk with the maximum mk .s f in M
23: Commit mk ; M ← M\mk
24: if TNS degrades then
25: Revert
26: end if
27: end while
the sensitivity for Vt swapping and DB swapping, respectively. 2nd DB distances are calculated
for downsizing and DB swapping, as shown in Lines 7 and 16. In Lines 21-27, we then greedily
commit the candidate move with the highest sensitivity as long as there is no TNS degradation.
70
Table 3.3. Experimental results for all the design blocks.
Design information 2nd DB-unaware 2nd DB-aware Our results Runtime
WS TNS lbase WS TNS lactual (∆%) WS TNS lopt (∆%) LeakOpt Routing Total
Design #Inst. Type Recovery
(ns) (ns) (mW) (ns) (ns) (mW) (ns) (ns) (mW) (s) (s) (s)
13806 Type-I 0.011 0.000 0.225 0.012 0.000 0.295 (31%) 0.001 0.000 0.227 (1%) 97.1% 207 819 1026
AES
13279 Type-II 0.002 0.000 0.182 0.002 0.000 0.190 (4%) -0.001 -0.002 0.180 (-1%) >100% 228 740 968
71
12380 Type-I 0.004 0.000 0.227 0.005 0.000 0.272 (20%) -0.001 -0.003 0.245 (8%) 60.0% 257 902 1159
MPEG
12250 Type-II 0.004 0.000 0.228 0.004 0.000 0.264 (16%) 0.002 0.000 0.237 (4%) 75.0% 261 872 1133
47061 Type-I 0.002 0.000 0.683 0.002 0.000 0.867 (27%) 0.001 0.000 0.745 (9%) 66.3% 1820 6911 8731
JPEG
39422 Type-II 0.003 0.000 0.697 0.003 0.000 0.765 (10%) 0.002 0.000 0.714 (2%) 75.0% 2039 7093 9132
67491 Type-I 0.011 0.000 1.203 0.012 0.000 1.786 (48%) 0.001 0.000 1.295 (8%) 84.2% 3819 29183 33002
VGA
67546 Type-II 0.015 0.000 1.230 0.015 0.000 1.449 (18%) -0.002 -0.010 1.256 (2%) 88.1% 3998 30799 34797
Experimental Setup
We implement our heuristics in C++ with OpenAccess 2.2.43 [167] to support LEF/DEF
[158]. We perform our experiments in a commercial 14nm FinFET technology with 9-track
triple-Vt libraries. We apply our optimization to four design blocks, AES, MPEG, JPEG and
VGA, from OpenCores [161]. We use two types of designs: Type-I has only SDB cells and
Type-II has both SDB and DDB cells. We vary the target clock period from 0.3 to 0.5ns with a
step size of 0.05ns, and we vary the block utilization from 60% to 80% with a step size of 5%. For
each data point, we denoise the experiments using three runs, by varying the target clock period
by ±1ps. We report the median value of the leakage recovery. We use Synopsys Design Compiler
L-2016.03-SP4 [171] for synthesis and Cadence Innovus Implementation System v17.1 [144] for
P&R, with leakage optimization in the highest effort mode. All the timing results in this section
are reported by OpenSTA [164]. Due to different timing results between the commercial P&R
tool and OpenSTA, we use a correlation methodology from [96] to correlate the P&R tool to
OpenSTA. All experiments are performed with eight threads on a 2.6GHz Intel Xeon server.
We measure how much recovery we achieve by our flow, as shown in Equation (3.1).
lopt − lbase
Recovery = 1 − (3.1)
lactual − lbase
We use lbase to denote the leakage value reported by the P&R tool at post-P&R stage, which is
2nd DB-oblivious. We use lactual to denote the 2nd DB-aware leakage value at post-P&R stage.
We use lopt to denote the leakage value after our optimization. Thus, Recovery is the leakage
recovery. Note that according to the metric Recovery, it is possible to have >100% leakage
recovery. For example, if the 2nd DB-aware analysis increases leakage from 0.182mW (lbase ) to
0.190mW (lactual ), and our optimization reduces leakage to 0.180mW (lopt ), then we say that
Recovery is 125%.
72
Design Space Exploration
In this section, we explore the sensitivity of results to design space choices. We explore
three knobs for the experimental setup: (i) 2CPP padding for all flip-flops, (ii) DB swapping and
(iii) multiple optimization iterations. We apply a minimum space of padding between flip-flops,
because our optimization does not touch flip-flops. Flip-flops can be abutted with other cells
but we require at least 2CPP space between flip-flops. We perform four sets of experiments to
validate our experimental setup:
• Baseline: A design with 2CPP padding for flip-flops.
• Expt. 1: A design without 2CPP padding for flip-flops.
• Expt. 2: A design without DB swapping.
• Expt. 3: A design with four iterations (imax = 4).
• Expt. 4: Sensitivity to timer setting (tolerance).
For the baseline, we use 2CPP padding for flip-flops and execute two iterations including
DB swapping. Flip-flops are wide and they contribute a large portion of the 2nd DB leakage
increase to neighboring cells. Expt. 1 shows that leakage recovery is degraded by 29.7% without
2CPP padding for flip-flops. In Expt. 2, we compare our baseline to the optimization with
DB swapping and show that 2.2% leakage recovery can be achieved with the addition of DB
swapping. In Expt. 3, we perform our optimization in four iterations and compare the result
to the baseline, where we only have two iterations of optimization. Four iterations give more
leakage recovery in exchange for longer runtime. In Expt. 4, we explore the sensitivity to
timer tolerance. The tolerance for the timer denotes a minimum percentage change in delay that
causes propagated delays to be recomputed during incremental timing updates. We sweep the
tolerance from 0% to 1% with a 0.1% step. As shown in Table 3.5, a design with 0.8% tolerance
shows 67% runtime improvement and the same leakage recovery compared to a design with 0%
73
Figure 3.6. Runtime analysis for AES design. We use Type-II, 60% initial utilization and 0.5ns
clock period. 1st loop denotes the first iteration of our optimization and 2nd loop denotes the
second iteration. Sizing includes sensitivity calculations and cell moves for gate sizing, Vt
swapping and DB swapping.
tolerance. Based on these results, we use 2CPP padding for flip-flops as our default P&R setup,
two iterations of optimization including DB swapping, and 0.8% tolerance in incremental timing
updates in our results.
Main Results
We conduct two overarching experiments to show results of our flow. The first experiment
shows leakage optimization outcomes over four design blocks. The second experiment further
studies the sensitivity of leakage recovery to target clock period and initial utilization of the
design.
Leakage Optimization. The design information and experimental results are summarized in
Table 3.3. The table reports (i) the number of instances in a design (#Inst.), (ii) design type as
defined in Figure 3.1 (Type), (iii) worst setup slack (WS), (iv) total negative setup slack (TNS),
(v) percentage of leakage increase from 2nd DB-unaware analysis (∆%), (vi) runtime of leakage
optimization including relocation and sizing (LeakOpt), and (vii) runtime of incremental routing
by a commercial P&R tool (Routing). We use four design blocks – AES, MPEG, JPEG and VGA
– with both Type-I and Type-II designs. Going from 2nd DB-unaware to 2nd DB-aware analysis,
the leakage is increased by up to 43%. After our leakage optimization flow with incremental
routing and timing recovery performed by the commercial tool, we achieve 80% leakage recovery
on average, with negligible timing degradation. Figure 3.6 shows the runtime breakdown in our
74
Table 3.4. Experimental results for design space exploration. We use AES Type-II design with
60% initial utilization and 0.5ns target clock period.
2nd DB-unaware 2nd DB-aware Our results

Expt. TNS lbase TNS lactual (∆%) TNS lopt (∆%)
Recovery
(ns) (mW) (ns) (mW) (ns) (mW)
Baseline 0.000 0.182 0.000 0.190 (4%) -0.002 0.180 (-1%) >100%
Expt. 1 0.000 0.178 0.000 0.198 (11%) -0.002 0.186 (4%) 60.0%
Expt. 2 0.000 0.182 0.000 0.190 (4%) -0.002 0.183 (1%) 87.5%
Expt. 3 0.000 0.182 0.000 0.190 (4%) -0.003 0.178 (-2%) >100%
flow. We can see that 24% of runtime is spent on gate sizing, Vt swapping and DB swapping,
including sensitivity calculations. 76% of runtime is spent on the incremental routing by the
commercial P&R tool. Next, we verify the robustness of our optimization with different clock
period and utilization.
Sensitivity to Target Clock Period and Utilization. Leakage increases due to the 2nd DB effect
and optimization results vary with initial utilization and target clock period. We sweep initial
utilization and target clock period for each design so as to study design impacts and what our
proposed methods can achieve. Figures 3.7(a) and (b) show leakage increments due to the 2nd
DB effect. Figures 3.7(c) and (d) show percentage recovery of the leakage after our optimization.
In this experiment, 70% initial utilization is used for target clock period sweep, and 0.4ns clock
period is used for initial utilization sweep. As target clock period decreases, the Type-I case
has larger leakage increments caused by the 2nd DB effect because cells are closely placed.14
However, there are smaller changes for the Type-II case because DDB cells mitigate the 2nd
DB effect. On the other hand, as initial utilization increases, both the Type-I and Type-II cases
have more or similar effects. Our proposed optimization, across the range of clock period and
utilization values, achieves 80% leakage recovery on average. For the testcases, the recovery
seen in Figures 3.7(c) and (d) is stable across clock periods and utilizations.
14With a tight timing constraint, there are more large cells and cells are more closely placed to reduce the net
delay, resulting in larger 2nd effect.
75
Figure 3.7. Studies of leakage power with varying target clock period and initial utilization.
Leakage increment from the 2nd DB effect vs. (a) target clock period and (b) initial utilization.
Leakage recovery by our optimization vs. (c) target clock period and (d) initial utilization.
3.1.4 Conclusion
In this work, we study the 2nd DB effect, which is one of the critical LLEs for physical
design in sub-10nm technologies, along with mixed-DB design that can have both SDB and
DDB cells. The 2nd DB effect is a new challenge to the physical design flow, and designers
must take this effect into account in order to mitigate model-hardware miscorrelation. We
propose new heuristics to recover leakage power increments caused by the 2nd DB effect,
including the use of 2nd DB-aware relocation, gate sizing, Vt swapping and DB swapping while
satisfying placement constraints. Our work achieves 80% leakage recovery on average using the
proposed flow. Directions for future work include: (i) detailed placement considering inter-row
minimum implant width and spacing constraints due to the mix of SDB and DDB cells; (ii) more
comprehensive study of sensitivity functions for improved leakage recovery considering multiple
corners; and (iii) support of mixed-DB within one cell, e.g., PMOS has SDB and NMOS has
DDB within one standard cell.
76
Table 3.5. Experimental results for sensitivity to the tolerance. The 2nd DB-aware leakage power
of the design is 0.190mW. Reloc., 1st loop and 2nd loop denote runtime of the relocation step,
and of the first and the second iterations of the optimization, respectively.
1st loop 2nd loop Our results

Tolerance Reloc. Sizing Reloc. Sizing LeakOpt WS Leak
(s) (s) (s) (s) (s) (ns) (mW)
0% 23 320 22 316 681 -0.001 0.180
0.1% 23 152 21 149 345 -0.001 0.180
0.2% 23 142 22 140 327 -0.001 0.180
0.3% 23 137 22 132 314 -0.001 0.180
0.4% 23 134 22 130 309 -0.001 0.180
0.5% 23 128 21 125 297 -0.001 0.180
0.6% 23 119 22 117 281 -0.001 0.180
0.7% 23 107 22 103 255 -0.001 0.180
0.8% 23 94 21 90 228 -0.001 0.180
0.9% 23 90 22 88 223 -0.001 0.184
1.0% 23 87 22 84 216 -0.001 0.187
3.2 Detailed Placement for IR Drop Mitigation by Power

Staple Insertion in Sub-10nm VLSI
Distribution of power (VDD) and ground (VSS) through the Power Delivery Network
(PDN) is extremely challenging in modern VLSI design. Back-end-of-line (BEOL) resistance
increases dramatically in sub-10nm VLSI [7] [129]; this increases delay and requires additional
buffers to meet timing requirements. The resulting routing increases capacitive loads and power
consumption, which causes greater supply voltage (IR) drop, and requires a denser PDN layout –
which causes more congestion again. In sub-10nm technologies, due to this vicious cycle, adding
more to the power mesh is not always the best mitigation of IR drop.
Insertion of power staples. Traditional PDN structures have power mesh on higher metal layers
and power rails on one or two lower metal layers, with stacked vias in between. Insertion of
power staples is a new technique for improving PDN robustness in sub-10nm technologies.
Power staples are short pieces of metal connecting two or more adjacent (i.e., consecutive)
VDD or VSS rails, to mitigate the IR drop. Figure 3.8 illustrates power staple insertions. In
77
Figure 3.8. Illustration of power staple insertions.
this example, adjacent M2 rails (VDD-VDD, or VSS-VSS) are connected by power staples in
M1. Since each power staple goes across at least two cell rows, vertical track availability in the
context of standard-cell pins and pre-routes is crucial to achieve sufficient power staple insertion.
Present-day limitations and our approach. There are two basic types of power staple insertion
flows: pre-placement and post-placement. A pre-placement flow inserts power staples at fixed
intervals before placement. This strategy can achieve good IR drop, but creates regular hard
placement blockages, and thus affects the placement flexibility. Alternatively, a post-placement
flow inserts power staples opportunistically after placement, wherever long empty vertical
segments (spanning at least two cell rows) are available. However, this approach is inherently
limited by the placement solution: especially for high-utilization designs where empty tracks are
usually not aligned between adjacent rows, the post-placement flow will result in fewer power
78
staples and worse IR drop than the pre-placement flow.15
In this work, we present new single-row and double-row dynamic programming (DP)-
based detailed placement optimizations to improve power staple insertion in post-placement
flows. Our contributions are summarized as follows.
• We propose a single-row dynamic programming-based approach to maximize a given

objective function which comprehends the benefit from placing power staples. The
proposed algorithm preserves the original cell ordering within each row.
• We propose a double- (multi)-row dynamic programming-based approach to maximize the

objective function. The multi-row algorithm can more aggressively optimize the placement
while keeping cells in their original rows and preserving row ordering.
• We propose a heuristic weighting function that balances the number of power staples for
VDD and VSS.
• We perform extensive studies on the sensitivity and scalability of our algorithms, and
show up to 13.2% IR drop reduction compared to a post-placement flow without our
optimization.
3.2.1 Related Work
Power staples and IR drop-aware placement. Power staple insertion is introduced in several
works. [94] proposes a design-technology co-optimization (DTCO) framework with power
staple insertion to mitigate IR drop. [131] introduces a “jumper” connection between vertically
separated power rails to mitigate both dynamic IR drop and electromigration. [85] shows cell
architectures using power staples, with both pre- and post-placement methodologies. Other works
address IR drop during placement, e.g., [67] and [71] use analytic placement. [31] calculates
15This is not an apples-to-apples comparison: the pre-placement flow inherently opts for PDN robustness over
layout density, and cannot achieve very high utilizations.
79
Notations Meaning
C Set of cells in a window of the initial placement.
ck kth cell in the left-to-right ordered initial placement.
xk Initial x coordinate of the cell ck . We define x0 = 0 for convenience.
[−x∆ , x∆ ] Horizontal displacement range of the cells.
wk Width of the cell ck . We define w0 = 0 for convenience.
[1, xmax ] Valid sites for cell placement, starting from 1.
d[i][ j] The maximum value of the benefit function of the staples that can be
placed when the cell ci is at coordinate xci + j. i ranges from 0 to n
and j ranges from −x∆ to x∆ .
Additional notations for multi-row DP
crk The kth cell, in left-to-right order, in row r.
wrk Width of the kth cell in the left-to-right order in row r.
D[i] [s1i ][l1 ] [s2i ][l2 ] The maximum value of the benefit function of the staples that can be
placed from coordinate 1 to coordinate i, knowing that the nearest
cell on the first row is cs1 i at coordinate i − l1 and the nearest cell on
the second row is cs2 i at coordinate i − l2 . Note that sri = 0 means
that no cell in row r is placed.
current and voltage drop based on an admittance matrix, and proposes a cost function to guide
placement perturbation.
Graph- and DP-based placement. Graph- and DP-based optimizations are widely used in
detailed placement targeting various objectives. [68] proposes a series of methods in ordered
single-row placement using a shortest-path algorithm to minimize perturbation or wirelength.
[48] describes an optimal single-row and double-row DP optimization to address the neighbor
diffusion effect. Their multi-row DP supports double-height cells, with cell relocation, reordering
and flipping. [87] describes an ordered double-row placement with support of multi-height cells.
A heuristic chain move algorithm is used to further improve the wirelength.
In summary, the works of [48] [68] [87] propose graph or dynamic programming models
targeting various problems in detailed placement. They also support various cell movements and
cell heights, which are not necessarily needed to optimize power staple insertion. Our work is
distinguished in that (i) we formulate a single-row and double-row dynamic programming-based
approach to maximize benefit functions that include the number and length of power staples; (ii)
80
our double-row DP algorithm makes progress site by site instead of cell by cell as in [48], which
allows precise calculation of power staple benefits (e.g., empty tracks available across multiple
cell rows) per track; and (iii) we show the benefits of our optimization according to both #staples
as well as IR drop in a post-placement power staple insertion flow.
Algorithm 5. Single-row optimization.

1: Initialize for all legal (i, j)
2: d[i][ j] ← −∞, d[0][ j] ← 0
3: for i = 0 to |C| do
4: for j = −x∆ to x∆ do
5: for all ( j′ ) ∈ getNext(i, j) do
6: i′ = i + 1
7: t ← d[i][ j] + getBene f it(xci + wci + j, xci′ + wc′i + j′ − 1) - α · disp( j′ )
8: d[i′ ][ j′ ] ← max(d[i′ ][ j′ ],t)
9: end for
10: end for
11: end for
12: f inalBene f it ← max(d[|C|][ j])
13: Return f inalBene f it
14: Procedure getNext
15: Input: cell i and displacement j of cell ci
16: Initialize nextList ← 0/
17: for j′ = −x∆ to x∆ do
18: i′ = i + 1
19: if inbound(xci′ + j′ ) and xci′ + j′ ≥ xci + j + wci then
20: nextList ← j′
21: end if
22: end for
23: Return nextList
3.2.2 Our Approach

Single-Row Optimization
Single-Row Optimization Problem: Given an initial legalized single-row placement and benefit
table of power staples, perturb the placement to maximize power staple benefits to upper and
lower rows while preserving the original cell ordering and placement legality.
For single-row dynamic programming, we place one cell at a time until all cells have
81
been placed. Table 3.6 shows notations that we use in our formulation. For each cell ck , cell
index k is that cell’s position in the row, from left-to-right. Given a cell set C, the leftmost cell is
c1 and the rightmost is c|C| .
For each cell, xck is the location of cell ck in the original placement. The displacement
range [−x∆ , x∆ ] means that we cannot place a cell more than x∆ sites away from the original
location. Thus, cell ck can only be placed in the interval [xck − x∆ , xck + x∆ ].
We use a 2D array d[i][ j] to represent the best-to-now solution (benefit) when we have
placed cell ci at location xi + j. Thus, we obtain the best solution when we have placed all |C|
cells. The array is of size (|C| + 1) × (2 ∗ x∆ + 1) where we initialize all d[0][ j] to have an initial
benefit of zero. Since cell ordering is preserved, the best result can be found by simply finding
the most beneficial solution across all d[|C|][ j].
Algorithm 5 describes our dynamic programming optimization. Lines 1-2 initialize all
DP array entries to have a benefit of negative infinity, except that all d[0][ j] are initialized to
have an initial benefit of zero. Lines 3-8 describe the main DP algorithm. The algorithm places
each cell one at a time, until all cells are placed. From each current placement of i cells with cell
ci at xi + j, we try to place cell ci+1 at all legal sites j′ . Function getNext(i, j) gets results for all
legal sites j′ so that cell ci+1 does not overlap with any previous cells. The benefit is updated
incrementally using the getBene f it function. To encourage smaller displacement for each cell
relative to its initial placement location, we subtract a displacement cost with a displacement
factor α. d[i′ ][ j′ ] is updated whenever the current solution is better in terms of t. Line 12 gets
the best single-row placement solution across all d[|C|][ j]. We also store a pointer for each entry
of the DP array so that the algorithm can trace back from d[i′ ][ j′ ] to d[i][ j], from which d[i′ ][ j′ ]
gets updated. In procedure getNext of Algorithm 5, Line 16 initializes the legal location for cell
ci′ to be an empty list. Lines 17-20 add each available legal displacement j to the list.
The procedure getBene f it takes the left and right coordinates of cell ci′ and incrementally
calculates the new power staples within this range. Since power staples should span at least two
row heights, a simple benefit table could be as follows: a power staple length of one row height
82
gets a benefit of zero; a power staple length between two and ten row heights gets a benefit of
one (i.e., the power staple’s length is extended by the two row-height stapling increment in the
current cell row); and a power staple length larger than 10 row heights does not get any benefit
(i.e., we do not encourage power staples longer than 10 row heights). Further discussion of the
benefit table is given in Section 3.2.3.
Double-Row Optimization
We present our problem statement and double-row (extendable to multi-row) dynamic

programming-based detailed placement.
Double-Row Optimization Problem: Given an initial legalized double-row placement and
benefit table of power staples, perturb the placement to maximize power staple benefits to upper
and lower rows without inter-row relocation while preserving the original cell ordering and
placement legality.
Unlike our single-row dynamic programming algorithm, our double-row dynamic pro-
gramming algorithm makes progress site by site instead of cell by cell. We use an array
D[i][s1 ][l1 ][s2 ][l2 ] to represent the solution and benefit up to site i. Recall that in single-row
optimization, d[i][ j] represents a solution with (i − 1) cells placed on the left of xi + j, and cell ci
placed at xi + j. Here, a valid solution up to site i may have cells placed on the left of site i, or
crossing site i. s1 (resp. s2 ) indicates the last placed cell cs1 (resp. cs2 ) in row 1 (resp. row 2),
and l1 (resp. l2 ) indicates the distance between the left boundary of cs1 (resp. cs2 ) and site i.
Figure 3.9 gives an example. The long vertical line indicates the current site i, and we
have placed one cell in row 1, and two cells in row 2. From the figure, the left boundary of the
last placed cell in row 1 has a distance of four to site i, and the left boundary of the second placed
cell in row 2 has a distance of two to site i. Thus, such a solution is represented by DP array
entry D[i][1][4][2][2].
To progress from D[i][s1 ][l1 ][s2 ][l2 ] to D[i + 1], we must: (i) choose whether to place new
cells in the two rows, and thus increase l1 and/or l2 ; and (ii) update the benefit incrementally by
83
Algorithm 6. Double-row optimization.
1: Initialize for all (i, s1 , l1 , s2 , l2 )
2: D[i][s1 ][l1 ][s2 ][l2 ] ← −∞
3: D[0][s1 ][l1 ][s2 ][l2 ] ← 0
4: for i = 0 to xmax do
5: for all (s1 , l1 , s2 , l2 ) do
6: for all (i′ , s′1 , l1′ , s′2 , l2′ ) ∈ getNext(i, s1 , l1 , s2 , l2 ) do
7: t ← D[i][s1 ][l1 ][s2 ][l2 ] + getBene f it(i′ , s′1 , l1′ , s′2 , l2′ ) − α · disp(s1 , s′1 , l1′ , s2 , s′2 , l2′ )
8: D[i′ ][s′1 ][l1′ ][s′2 ][l2′ ] ← (max(D[i′ ][s′1 ][l1′ ][s′2 ][l2′ ],t)
9: end for
10: end for
11: end for
12: f inalBene f it ← max(d[xmax ][|c1 |][l1 ][|c2 |][l2 ])
13: Return f inalBene f it
14: Procedure getNext
15: Input: i, s1 , l1 , s2 , l2
16: Initialize nextList ← 0/
17: i′ ← i + 1
18: Only place a cell on the first row
19: s′1 ← s1 + 1, l1′ ← 0, s′2 ← s2 , l2′ ← l2 + 1
20: if legalPlace(s1 , l1 , s′1 ) then
21: nextList.add ← (i′ , s′1 , l1′ , s′2 , l2′ )
22: end if
23: Only place a cell on the second row
24: s′1 ← s1 , l1′ ← l1 + 1, s′2 ← s2 + 1, l2′ ← 0
25: if legalPlace(s2 , l2 , s′2 ) then
27: end if
28: Place cells on both rows
29: s′1 ← s1 + 1, l1′ ← 0, s′2 ← s2 + 1, l2′ ← 0
30: if legalPlace(s1 , l1 , s′1 ) and legalPlace(s2 , l2 , s′2 ) then
32: end if
33: Place no new cells
34: s′1 ← s1 , l1′ ← l1 + 1, s′2 ← s2 , l2′ ← l2 + 1
36: Return nextList
84
adding the new power staples at site i + 1 and subtracting any displacement cost of the new cells.
Note that since we go over all sites, assuming we make progress from site i to i + 1, we can limit
the new cell placement (with its left boundary) to be exactly at site i + 1 without sacrificing the
solution space. A new cell placement at sites other than i + 1, e.g., i′ + 1, will be handled when
we make progress from i′ to i′ + 1.
Algorithm 6 describes our double-row optimization. Lines 1-3 initialize the DP array
entry to have a negative infinity benefit, except that all D[0] entries have an initial benefit of zero.
For the current DP entry D[i][s1 ][l1 ][s2 ][l2 ] in Line 5, Lines 6-8 update the corresponding DP
entry at site i′ = i + 1 for all legal (i′ , s′1 , l1′ , s′2 , l2′ ). The displacement function checks whether
there are newly placed cells, and will only count the displacement for the newly placed cells.
Then, the final solution is obtained by finding the most beneficial solution when we go over all
sites, and all cells. In Algorithm 6, procedure getNext generates all legal (i′ , s′1 , l1′ , s′2 , l2′ ). From
the current DP entry, Lines 18-21 generate all legal (i′ , s′1 , l1′ , s′2 , l2′ ) when we place a cell only on
the first row; similarly, Lines 23-26, 28-31, and 33-35 generate all legal (i′ , s′1 , l1′ , s′2 , l2′ ) when we
place a cell only on the second row, on both rows, or on no rows, respectively.
For dynamic programming optimization of more than two rows, we can just add DP
array dimensions [sk ][lk ] for each added row. However, the added dimensions may grow the total
runtime and memory footprint beyond practical usage. In this work, we use double-row dynamic
programming for optimization.
The size of the DP array for n rows is O(xmax (smax (2 ∗ x∆ + wmax ))n ), where wmax is the
maximum cell width and smax is the maximum value for dimension sk . Experiments using our
implementation reveal runtimes that are more obviously dependent on the number of getNext()
function calls, as opposed to the size of the DP array. Our experiments also show that the number
of getNext() function calls mainly depends on the number of cells in the optimization window
and the displacement range x∆ . This points to tuning of our DP implementation as a necessary
future improvement.
85
Figure 3.9. The case represented by D[i][1][4][2][2].
Overall Flow
Figure 3.10 shows the overall optimization flow. Given a post-P&R database, we first
perform our dynamic programming-based detailed placement optimization to maximize the
potential number of power staples, and then perform incremental routing and Vt swapping to fix
routing and timing degradation after our optimization. Next, we perform power staple insertion
to complete the PDN and perform a vectorless dynamic IR drop analysis to show the results. We
note that our optimization and power staple insertion could also be performed in both post-place
and post-CTS stages. However, to compare the QoR with, and without, our optimization using a
single database, in this work we only perform our optimization in the post-route stage, as shown
in Figure 3.10.
We implement our dynamic programming in C++ with OpenAccess 2.2.43 [167] to

support LEF/DEF [158]. We perform experiments in a 7nm FinFET technology with multi-
height triple-Vt libraries from a leading technology consortium.16 We apply our optimization
16We use the process, voltage, and temperature conditions (TT, 0.65V, 25◦C) for both place-and-route and IR-drop
analysis. We build the power mesh on M7 to M9 with 0.5µm width, 15.5µm pitch and 7µm offset, and M2 power
rails. We use M1 and V12 for power staples, and M1 is not allowed for routing.
86
Figure 3.10. Overall flow of our work. The red box indicates steps that we implement. Commer-
cial tools [144] [148] are used for all other steps.
to Arm Cortex-M0 and three design blocks (AES, JPEG and MPEG) from OpenCores [161].
Design information is summarized in Table 3.7.17
In this section, we first investigate sensitivity and scalability with respect to the displace-
ment range. Second, we study effects of weighting factors, i.e., displacement factor and benefit
table. Third, we propose a heuristic benefit table that is able to balance VDD and VSS staples.
Fourth, we show experimental results across different design blocks and utilizations, with IR drop
analysis. Last, we compare three examples of metaheuristics that can be implemented around
our basic DP optimization. In the following experiments, we use our double-row optimization by
default, and we show the comparison with single-row optimization.
Scalability/Sensitivity Study
Since the displacement range (x∆ ) determines the solution space, and also the size of the
DP table data structure, we first study the sensitivity of #staples to the displacement range, and
17We synthesize designs using Synopsys Design Compiler L-2016.03-SP4 [171], and perform place-and-route
using Cadence Innovus Implementation System v16.2 [144]. We perform vectorless dynamic IR drop analysis using
Cadence Voltus IC power integrity solution [148]. Our dynamic programming-based optimizations are performed in
a single thread on a 2.6GHz Intel Xeon server.
87
Table 3.7. Design information.
Design #Inst Clkp

M0 ∼10K 500ps
AES ∼12K 500ps
MPEG ∼14K 500ps
JPEG ∼54K 500ps
Figure 3.11. (a) Sensitivity of #staples to displacement range x∆ and (b) sensitivity of runtime to
displacement range (x∆ ).
the scalability of our algorithm when we increase the displacement range. In this experiment, we
sweep x∆ from 0 to 10 with a step of 1. To show the stability of our optimization with respect to
block utilization, we use design block AES with four target utilizations: 0.60, 0.65, 0.70 and
0.75. In this experiment, since weighting factors do not affect the scalability, they have a default
value of 0. The benefit table simply shows the total length of power staples that can be inserted.
Figure 3.11(a) shows the number of power staples vs. displacement range. The x-axis is
the displacement range, and the y-axis is #staples. We can see that for values of displacement
range > 5, there are diminishing returns in terms of #staples. Also, design blocks with a lower
utilization tend to saturate with larger #staples, as one would expect. These data motivate future
efforts to obtain a more detailed and accurate understanding of the relationship between target
88
utilization and maximum IR drop, e.g., for use in early design exploration.18 Note that the
saturation of #staples versus displacement range depends on the library used. The dotted lines in
Figure 3.11(a) show the (normalized) result when all widths are doubled (x-coordinates, and cell
widths and pin locations in each library cell). With the wider cells, increasing the displacement
range continues to give staple insertion benefits up to larger values of x∆ .
Figure 3.11(b) shows the runtime vs. displacement range. The run with x∆ = 9 consumes
4× runtime compared to the run with x∆ = 5, but only results in minor increase in #staples.
Utilization generally does not affect the runtime, showing the robustness of our optimization.
Thus, to balance the solution quality and runtime, we choose x∆ = 5 in all the following
experiments.
Another study examines whether there is any benefit to breaking the optimization into
multiple, more “gradual” phases. Table 3.8 shows how number of staples changes when we run
two rounds of optimizations with total displacement range (x∆ ) = 6. The outcome is worst when
we perform two phases of optimization, each with displacement range (x∆ ) = 3. The apparent
takeaway is that it is better to consume available displacement range “in one shot”.
Study of Weighting Factors
Our next experiment studies the impact of the displacement factor (α). We sweep the
displacement factor with value 0, and from 10−3.5 to 10−0.5 , with a multiplier of 100.5 . We still
use the same design block AES with four different utilizations. Figure 3.12(a) shows the #staples
vs. displacement factor and Figure 3.12(b) shows the total cell displacement vs. displacement
factor.
From Figure 3.12, with a non-zero weight, total displacement decreases sharply while
#staples does not change much. This indicates that we can preserve most of #staples while
18As our experiments show, there is less available increment in the number of placed staples when the utilization
is higher. Decreasing (resp. increasing) utilization reduces current density while increasing (resp. decreasing)
potential staple insertion. I.e., there is a compounding of IR drop mitigation (resp., degradation) effects, which is
further compounded by changes to gate sizing and voltage-induced timing margins. Deriving a principled model of
this dynamic is an open challenge.
89
Figure 3.12. Sensitivities of (a) #staples and (b) total cell displacement, to the displacement
factor (α).
Table 3.8. Total #staples placed when we apply two rounds of optimization with total dis-
placement range (x∆ ) = 6. Each row represents design AES with utilization between 0.60 and
0.75.
x∆
Util
6 1+5 2+4 3+3 4+2 5+1
0.60 36974 36857 36634 36493 36611 36748
0.65 32389 32240 32079 31936 32054 32199
0.70 28624 28540 28362 28252 28287 28432
0.75 25090 25939 24867 24722 24780 24915
reducing total cell displacement by using a non-zero α. A smaller total cell displacement is
beneficial to minimize perturbation of the input layout and thus we use α = 0.01 in all the
following experiments. Note that this choice is based on our experiments, and that a different α
may be preferred in a different technology node or design enablement.
Study of Benefit Table
We now discuss our investigations into the benefit table. First, we study the tradeoff
between short and long staples, i.e., whether we should encourage fewer but longer staples,
or more but shorter staples. Since our optimization changes the placement, to make a fair
90
Figure 3.13. Design examples with different lengths of power staples. (a) Power staple length =
2 row heights. (b) Power staple length = 10 row heights.
comparison and investigate the tradeoff, we use the AES design with pre-placed power staples at
fixed intervals (20CPP) and perform IR drop analysis.19 Each run uses a different staple length,
from 2 to 10 row heights, with a step size of 2 row heights. We only use VDD power staples in
this study.
As shown in Figure 3.13, the total length of all power staples is kept essentially constant
across different configurations. We also ensure that the initial input placement is identical across
all runs, with power staple tracks always free from pins, blockages and pre-routes. As shown in
Table 3.9, the IR drop is not sensitive to the staple length.
Table 3.9. IR drop vs. staple length.
Staple length 2 4 6 8 10
Worst IR drop (mV) 67.9 67.8 67.7 67.7 67.6
19 CPP = contacted poly pitch. In this work, we use a technology enablement where 1CPP = 42nm.
91
Now we propose our heuristic benefit table setting to improve the balance between #VDD
and VSS staples. Since the double-row dynamic programming is performed every two rows
in a window, without overlapping (e.g., optimize 1st and 2nd rows, then 3rd and 4th rows), the
optimization is always aligned with either VDD rails or with VSS rails, resulting in potentially
biased power staples (e.g., staples of one rail are systematically more likely to occur than those
of the other rail). To overcome this issue, we propose a reconfiguration of the benefit table so that
a single-height staple gets a non-zero benefit (β ), with all other staple lengths getting a benefit of
one (up to ten row heights). Such single-height staples may get extended in the next optimization
window so that there is a chance to re-balance VDD/VSS staples. We have experimentally swept
β from 0 to 1.0, with a step size of 0.1, with results shown in Figure 3.14(a). As β increases,
VDD/VSS staples become balanced without sacrificing the total #staples. We use β = 0.7 in all
following experiments.
Metaheuristics
In this section, we compare three strategies: (i) single-row optimization (SR); (ii) double-
row optimization with overlap (Meta-1); and (iii) two-pass double-row optimization without
overlap (Meta-2). In (ii), we perform the double-row optimization every two rows with one row
overlapped (e.g., we optimize row 1 and row 2, then row 2 and row 3, then row 3 and row 4, etc.).
In (iii), the first pass starts at the odd rows, and then the second pass starts at the even rows (e.g.,
we optimize row 1 and row 2, row 3 and row 4, etc. in the first pass, and row 2 and row 3, etc. in
the second pass).
Therefore, effectively each row is covered once in SR, and twice in Meta-1 and Meta-2.
We show the experimental results in Figure 3.14(b). Compared to Baseline, where there is
no detailed placement optimization, the methods SR, Meta-1 and Meta-2 respectively achieve
20.1%, 22.3% and 22.7% more inserted staples, averaged over all four target utilizations. Thus,
we use Meta-2 for our main experimental results.20
20 Note that there is generally no benefit from making multiple passes over the rows of the layout with the same
total length of staples. Our ongoing efforts seek richer combinations of metaheuristics to further improve results.
92
Table 3.10. Experimental results using double-row optimization.
#Staples IR drop (mV,Avg) IR drop (mV,Worst) ∆WNS Runtime

Design Util
Init Final (∆%) Init Final (∆%) Init Final (∆%) (ns) (s)
60% 29109 33423 (+14.8%) 29 28 (-3.4%) 76 66 (-13.2%) -0.009 2666
65% 25295 29254 (+15.7%) 28 28 (0.0%) 58 60 (+3.4%) -0.016 2249.5
M0
70% 22233 25628 (+15.3%) 29 29 (0.0%) 77 76 (-1.3%) -0.008 2450
75% 19700 22889 (+16.2%) 29 29 (0.0%) 80 80 (0.0%) -0.017 2058
60% 31355 37813 (+20.6%) 23 23 (0.0%) 51 47 (-5.6%) -0.001 2542.5
65% 27165 33164 (+22.1%) 22 22 (0.0%) 54 51 (-5.6%) -0.001 2127
AES
93
70% 23571 29129 (+23.6%) 22 22 (0.0%) 63 62 (-1.6%) -0.001 3142
75% 20496 25530 (+24.6%) 22 22 (0.0%) 71 68 (-4.2%) -0.006 2520
60% 75781 80951 (+6.8%) 24 24 (0.0%) 64 60 (-6.3%) 0.000 3700
65% 67836 72685 (+7.1%) 24 24 (0.0%) 57 60 (+5.3%) 0.000 4088.5
MPEG
70% 61152 65362 (+6.9%) 25 25 (0.0%) 62 59 (-4.8%) 0.000 4410
75% 55426 58928 (+6.3%) 30 30 (0.0%) 81 79 (-2.5%) -0.002 3709
60% 217467 244333 (+12.4%) 28 27 (-3.6%) 88 84 (-4.5%) -0.001 16679
65% 193089 218483 (+13.2%) 26 27 (+3.8%) 79 76 (-3.8%) 0.000 13869.5
JPEG
70% 171857 195781 (+13.9%) 28 28 (0.0%) 69 70 (+1.4%) -0.007 15069
75% 152805 174618 (+14.3%) 29 28 (-3.4%) 73 68 (-6.8%) 0.000 12563.5
Figure 3.14. (a) Number of staples (total, VDD and VSS) that can be placed, versus balance
factor (β ) (AES with utilization 0.60) and (b) number of staples that can be placed, versus
utilization by different strategies (AES with utilization between 0.60 and 0.75).
Main Results
We execute our Meta-2 optimization using all design blocks with the aforementioned
parameter settings, and report #staples, average and worst IR drop before and after optimization,
along with ∆ worst negative slack (WNS) and runtime, at four utilizations per design block.
Table 3.10 shows both the initial and post-optimization values of metrics (IR drop values are
in units of mV). Figure 3.15 shows vectorless dynamic IR drop heatmaps for AES with 0.60
utilization.
We observe that our optimization can increase #staples by anywhere from 6.2% to
24.6%, and reduce IR drop by up to 13.2%, compared to post-placement staple insertion without
our optimization. Furthermore, ∆WNS is similar before and after optimization, showing the
effectiveness and robustness of our optimization. On the downside, IR drop can increase after
our optimization (e.g., the case of M0 with 0.65 utilization); this is because the design’s current
map changes with placement perturbation and timing recovery steps, even as we add more
power staples. Hence, directions for improvement include better control of the commercial
P&R tool, and/or implementing our own legalization and ECO capabilities. We also note that
94
Figure 3.15. Heatmaps of vectorless dynamic IR drop for AES.
current runtimes are long, reflecting unoptimized implementation that pays heavily for getNext()
function calls.
3.2.4 Conclusion
We have presented novel DP-based single- and double-row detailed placement optimiza-
tions with metaheuristics to maximize power staple insertion in a post-placement flow. We
perform extensive studies on the scalability and sensitivity to parameters, impact of the benefit
table, as well as balance of VDD/VSS staples. Compared to the traditional post-placement
flow, we achieve up to 13.2% (10mV) reduction in IR drop, with almost no WNS degradation
compared to a pre-placement flow. Directions for future research include (i) better exploration of
the benefit table; (ii) understanding in greater detail the dynamics of the relationship between
power staples and IR drop; (iii) further exploration of metaheuristics, (iv) runtime speedup
techniques, and (v) exploration of power mesh structures with different width, pitch and offset.
95
3.3 Acknowledgments
Chapter 3 contains a reprint of Sunik Heo, Andrew B. Kahng, Minsoo Kim and Lutong
Wang, “Diffusion Break-Aware Leakage Power Optimization and Detailed Placement in Sub-
10nm VLSI”, Proc. Asia and South Pacific Design Automation Conference, 2019; and Sunik Heo,
Andrew B. Kahng, Minsoo Kim and Lutong Wang and Chutong Yang, “Detailed Placement for
IR Drop Mitigation by Power Staple Insertion in Sub-10nm VLSI”, Proc. Design, Automation
and Test in Europe, 2019. The dissertation author is a main contributor to, and a primary author
of, each of these papers.
I would like to thank my coauthors, Sunik Heo, Professor Andrew B. Kahng, Dr. Lutong
Wang and Chutong Yang.
96
Chapter 4
Design-Technology Co-Optimization
This chapter presents three design-technology co-optimization methodologies. The first

work presents a systematic framework for routability assessments, which can systematically
evaluate a measure of intrinsic routability, Kth , across both technology and design choices. We
focus on routability since it is a critical factor in the scaling of area and cost. The experimental
studies demonstrate the assessment of routability impacts for advanced-node technology and
design options. We demonstrate a machine learning (ML)-based Kth prediction flow to reduce
runtime, disk space and commercial tool licenses needed to implement our framework. The
second work presents a routability study that uses the framework with 3nm technology configu-
rations. We demonstrate that block-level area scaling is reversed from cell-level scaling in cell
height <150nm regime, using an ML-assisted DTCO methodology to optimize block-level area
accounting for routability. The third work presents a systematic framework for design-technology
pathfinding with improved design enablements. The goal of our work is to build a “configurable”
and open-sourced process design kits (PDK) generation with promising scaling boosters, such as
backside PDN and buried power rails, and study PPAC given technology and design parameters.
In addition, we propose our ML-based parameter tuning flow and clustering-based cell width-
regularized placements to generate more realistic artificial designs. We conduct experiments to
present PPAC evaluation capabilities of the framework with our predictive technology and we
achieve comparable results with studies from industry.
97
4.1 PROBE2.0: A Systematic Framework for Routability
Assessment from Technology to Design in Advanced
Nodes
Due to the slowdown of dimension scaling relative to the trend of the traditional Moore’s
Law, scaling boosters, such as backside power delivery networks (BSPDN) and buried power
rails (BPR), are introduced at advanced technology nodes. Since scaling boosters play an
important role to optimize power, performance, area and cost (PPAC) of IC products in advanced
technologies, accurate and fast evaluations and predictions of PPAC are critical at an early stage
of technology development. Also, use of scaling boosters makes evaluations and predictions
of technology more difficult since they introduce a large number of knobs that simultaneously
contribute to PPAC improvement.
Design-Technology Co-Optimization (DTCO) is now a well-known key element to
develop advanced technology nodes and designs in modern VLSI chip design. Today’s DTCO
spans assessment and co-optimization across almost all components of semiconductor technology
and design enablement. As described in Figure 4.1, the DTCO process consists of three stages:
Technology, Design Enablement, and Design. First, the technology stage includes modeling and
simulation methodologies for process and device technology. Second, the design enablement
stage includes creation of required process design kits (PDK) for the ensuing design stage; these
include device models, standard-cell libraries, routing technology files and interconnect parasitic
(RC) models. Last, the design stage includes logic synthesis and place-and-route (P&R) based
on the generated PDKs from the design enablement stage.
To evaluate and predict technology and design at advanced nodes, all three stages must be
correctly performed, and PDKs must be generated from technology and design enablement stages.
However, the DTCO process is not simple: feedback from design stage to technology stage
takes weeks to months of turnaround time, along with immense engineering efforts. Also, based
on the design feedback, additional PDKs may need to be generated at the design enablement
98
Figure 4.1. The DTCO process consists of three main stages: Technology, Design Enablement,
and Design. Figure is redrawn from [174].
stage, which requires additional weeks to months. In order to reduce the turnaround time and
maximize the benefit of the DTCO process, a fast and accurate DTCO methodology is needed to
assess PPAC with reasonable turnaround time, and to more precisely guide multi-million dollar
decisions at an early stage of technology development.
Density is the overarching metric for enablement of system-level benefits through scal-
ing [123], and directly determines the Area and Cost aspects of PPAC. In our work, we focus
on routability implications of technology choices, since routability and density are intimately
tied together. The challenge of routability arises as back-end-of-line (BEOL, i.e., metal layers)
technology fails to scale down in step with front-end-of-line (FEOL, i.e., device layers). Also,
cell heights of standard-cell architectures are a crucial lever for density scaling [109], but small
standard-cell heights challenge area routing and pin accessibility. Furthermore, high BEOL
resistances require denser power delivery networks (PDN), which occupy more routing resources
and harm routability.
Contributions of This Work. The goal of the work described in this section is to enable faster
and more comprehensive evaluation of technology options, early in the technology development
process. In particular, we describe a framework for systematic assessment of routability across
the combined space of technology options and design enablement options.
99
Many measures of routability have been developed and applied over the past decades.
These span the use of congestion maps, metrics of pin accessibility, machine learning-based
congestion predictors, and other techniques, as we review in Section 4.1.1 below. However,
these previous methods to assess routability do not solve two root causes of the long feedback
loop in DTCO (Figure 4.1). The first root cause is that efficient simultaneous exploration of
technology and design options in DTCO is blocked by the effort and expense of the design
enablement stage. Producing layouts and characterizations for standard-cell libraries requires an
enormous amount of engineering cost and time, due to complex constraints such as transistor-
level placement, in-cell routing, and pin accessibility. Today’s DTCO relies on limited, heuristic
layout synthesis (e.g., manual layout of 15-60 key cells) to assess a given set of technology
options. The second root cause is that routability assessment methods have mostly focused on
assessing implementations of designs (e.g., to predict routability of the placement of a particular
netlist), rather than assessing design enablements.
Our present work attacks both of the above-mentioned root causes of long design feedback
loops in DTCO. To do this, we build on two threads of recent works: (1) automatic standard-cell
layout generation using Satisfiability Modulo Theories solvers [26] [82], and (2) assessment of
intrinsic routability of BEOL stack options via the Kth metric [64]. We review these works in
Section 4.1.1. As demonstrated below, our framework is able to provide assessments of intrinsic
routability across a range of technology and design parameters reflecting sub-7nm technologies.
Our main contributions are summarized as follows.
• We describe a systematic and complete framework to evaluate routability across key

parameters of technology and design. Our framework is generalizable and flexible; it
enables rapid evaluation of hundreds of technology and design enablement options within
hours or days, providing a valuable tool for early technology development.
• We propose a pin shape selection strategy based on the remaining pin access (RPA) [108],
along with a top-metal-only pin shape selection strategy, at our design enablement stage.
100
We also extend methods of [26] [82] to automatically produce more realistic standard-cell
libraries (LEF format [158]) in terms of power and ground pins, contacted poly pitches
(CPP) and metal pitches.
• We extend the method of [64] to assess routability across configurations of technology and
design, rather than only BEOL stack options. We study both cell-level and design-level
routability, and show advantages of using knight’s tour-based artificial netlist topology
generation in cell-level routability assessment. (A knight’s tour is a sequence of knight’s
moves in a chessboard that visits each square exactly once. Section 4.1.3 below explains
its use in our methodology.)
• We achieve seamless integrations with commercial P&R tooling, via automated generation
of power-ground hookups in cell layouts, and routing technology files to reflect modified
design rules.
• We demonstrate accurate learning-based Kth prediction that reduces runtime, disk storage
and tool license overheads of our framework.
Our work is organized as follows. Section 4.1.1 reviews related previous works. Sec-
tion 4.1.2 gives an overview of our framework, along with the parameterizations and other details
of how we generate standard-cell layouts and design enablements. Section 4.1.3 describes our
methodologies for routability assessments and learning-based Kth prediction. Section 4.1.4
shows our experimental setup, key experiments and results. We give conclusions and directions
of future work in Section 4.1.5.
4.1.1 Related Work
We now review relevant previous works. Broadly, these can be categorized into works on
(1) standard-cell layout generation and (2) routability assessment. Our review includes the works
of [26] [82] and of [64] which have provided a basis for our present work.
101
Standard-Cell Layout Generation
Automatic Standard-Cell Layout Generation. Standard-cell layout synthesis can help library
design teams explore cell architectures with holistic consideration of transistor placement, in-cell
routing, complicated design rules, and pin accessibility. The methods of [45] [136] provide
co-optimization of transistor placement and in-cell routing, but do not consider such aspects
as multi-patterning design rules that are seen in advanced technology nodes. [59] [83] [118]
propose standard-cell layout automation frameworks for sub-7nm technologies, but these works
incorporate multiple heuristic approaches with no guarantees of optimality. Park et al. [82] unify
transistor-level placement and routing with dynamic pin allocation, and apply a satisfiability
modulo theories (SMT) solver to achieve optimal layout solutions.
Pin Accessibility-Aware Standard-Cell Layout. One of the most difficult design features
for standard-cell layout generation is the pin accessibility, which is challenged by the limited
number of tracks and complicated design rules. The works of [105] [127] define metrics for
pin accessibility within their objectives for standard-cell layout optimization. Seo et al. [108]
propose the remaining pin access (RPA) metric to capture pin access interference from access
points of neighboring pins. Cheng et al. [26] devise “at-least-k” Boolean constraints that
guarantee a minimum number of pin openings (access points) per pin in the cell layout.
The SMT-based Standard-Cell Layout Generation of [26] [82]. Our present work builds on
the SMT-based parametric standard-cell generation framework of [26] [82]. This framework
takes in three main inputs. (1) Cell architectures: number of routing tracks and transistor
fins, and track pitches. (2) Netlist: component connectivity of library cell. (3) Design rules:
parametric conditional design rules depending on cell architecture. Given these inputs, a cell
layout is produced that is optimal with respect to cell area, M2 track use (routability) and routed
metal length, in lexicographic order of these criteria. A unifying dynamic pin allocation (DPA)
constraint integrates additional design constraints such as transistor placement, in-cell routing,
conditional design rules and pin accessibility constraints. This yields a constraint satisfaction
102
formulation that produces an optimized cell layout via a single multi-objective optimization.
We observe that these previous works do not provide necessary enablements of commer-
cial standard-cell P&R, such as LEF [158] generation, PDN generation, and routing technology
file generation. Nor do these works support control parameters for standard-cell layout generation
that are relevant in sub-7nm nodes. Below we describe a complete framework to support both
standard-cell layout generation and associated P&R enablement. Our layout generation uses
RPA-based and top-metal-only pin shape selection methods that improve pin accessibility.
Routability Assessment
Routability is a hard constraint in the modern (fixed-die) place-and-route context. Thus, many
previous works have studied routability-driven placement, as well as ripup-and-reroute methods
in global routing. For example, [25] [54] [75] [86] all propose routability-driven placement
based on congestion maps derived from early trial or global routing. Pin accessibility of a given
standard-cell instance also affects routability. The above-mentioned work of [108] describes pin
accessibility-aware detailed placement based on the RPA metric. However, in this discussion we
do not focus on placement and routing optimizations, but rather on methods for assessment of
routability.
Routability Analysis and Prediction. Tseng et al. [117] propose a systematic framework with
P&R tools to check routability, aiming to improve placement outcomes. The authors propose
standard cell-level and placement-level routability scores to generate cell spacing constraints.
[49] proposes an optimal ILP-based detailed router and evaluate feasibility (routability) of routing
clips based on an ILP solver. The authors of [73] [99] [100] use Boolean satisfiability (SAT)
to analyze routability under conditional design rules. [99] [100] furthermore extract minimal
unsatisfiable subsets to diagnose bottlenecks when designs are proven to be unroutable.
Several recent works propose machine learning (ML)-based routability predictions. Zhuo
et al. [135] propose a new routability prediction model based on supervised learning in placement.
The works of [21] [124] predict routability based on convolutional neural networks (CNN)
103
and support vector machines (SVM), respectively. Chan et al. [13] also propose SVM-based
routability prediction, but aim to evaluate routability for various BEOL stack options.
The PROBE Routability Measurement Utility of [64]. Our present work builds on the
“PROBE” framework of [64], which gives a measure of inherent routability of BEOL stack
options. PROBE begins with a placement solution that is easy to route – e.g., a regular mesh
placement of a mesh netlist topology. PROBE then iteratively swaps the locations of random
pairs of neighboring placed cells, progressively “tangling” the placed netlist until the routing fails
with more than some threshold number of post-route DRC violations. The number of random
neighbor cell swaps performed, normalized to the number of instances in a design, is denoted by
K. The number of swaps beyond which routing fails is denoted as the K threshold (Kth ), and
captures intrinsic routing capacity (e.g., of a given BEOL stack).
Figure 4.2(a) shows the scope of PROBE. Given a placement, a set of BEOL stack
options {B1 , B2 , ..., Bi } can be ranked in terms of routability. The framework supports two types
of placements, shown in Figure 4.3. Mesh-like placements do not reflect any specific design;
they consist of an array of instances of a given 2-input or 3-input cell. Connections are made
between neighbors, inducing a near-meshlike netlist topology. Cell width-regularized placements
are design-specific, and are produced by commercial P&R tools. However, the standard cells in
the placements are all given the same cell width to avoid illegal placements after neighbor cell
swaps.
We observe that [64] is applied only to BEOL stack options, and does not cover the
rich space of FEOL technology and design enablement options. Moreover, the near-meshlike
topology can produce only a limited range of routed wirelength and Rent parameter values
that may not match realistic values. The framework we describe below supports DTCO with
routability assessment across technology and design enablement options. We use a knight’s
tour-based construction that can better reflect actual design attributes.
104
Figure 4.2. Overall flows for PROBE [64] and PROBE2.0. (a) PROBE only evaluates BEOL
stack options. (b) The PROBE2.0 flow includes standard-cell layout generation [26] [82], design
enablement and routability assessment. (c) The PROBE2.0 flow uses a trained learning-based
model to predict Kth .
4.1.2 The PROBE2.0 Framework
Figure 4.2(b) shows our PROBE2.0 framework. It takes technology and design pa-
rameters as primary inputs, and consists of three major stages. (1) The standard-cell layout
generation stage is based on input technology parameters, and is performed using an extension
of an SMT-based standard-cell layout generation [26] [82]. It produces purely grid-based pin
locations and cell boundaries. (2) The design enablement stage begins with the generated
standard-cell layouts, and is also performed according to the input technology parameters. Design
enablement generates LEF [158], Liberty [159] and routing technology files. LEF file generation
converts the primitive form of layouts to LEF format. The conversion considers real-world
constraints for the stability of standard-cell characteristics. (3) The routability assessment stage
uses a knight’s tour-based topology as well as open-source designs with the PROBE approach
[64]. Also, Figure 4.2(c) shows how the PROBE2.0 flow can be realized with learning-based Kth
prediction, where a trained machine learning model enables more efficient routability assessment.
The remainder of this section describes aspects of standard-cell architecture, technology
and design parameters, and design enablement in the PROBE2.0 framework. Our routability
assessment and learning-based Kth prediction flow are explained in Section 4.1.3.
105
Figure 4.3. Placements for PROBE [64]. (a) A mesh-like placement based on a 2-input cell. (b)
A cell width-regularized placement. The orange-striped cells are considered to be neighbors of
the blue-striped cell. The blue-striped cell is swapped with a randomly-selected neighbor.
Standard-Cell Architecture
Figure 4.4 shows a grid-based standard-cell architecture and technology parameters of

standard cells, as used in our work. We follow the 7nm standard-cell architectures in [32] [119]
to generate the grid-based P&R graph with four layers TS/PC, M0, M1, and M2. TS/PC and M0
layers are included in FEOL layers and M1 and M2 layers are included in BEOL layers. Next,
we give detailed definitions of the eight technology parameters and five design parameters that
PROBE2.0 supports as user inputs.
Definitions of Technology Parameters
Technology parameters include various options for process technology as well as design
enablement.
(1) Fin: The number of fins for devices of standard cells. We use 2 and 3 for Fin [32] [119].
(2) CPP: Contacted poly pitch for standard cells. We use 48 and 54nm for CPP [109] [119].
(3) MP: Metal pitch for M2 and M3 routing layers. We use 24, 32 and 40nm for MP [109] [119].
(4) RT: The number of available M2 routing tracks for standard cells. We use 4, 5 and 6 for
RT [32] [119].
106
Figure 4.4. A grid-based standard-cell architecture and technology parameters for standard cells.
(5) PGpin: Pin types for power and ground of standard cells. We support three types of power
and ground pins: M1 [119], M1+M2 [17], and BPR [107]. M1 denotes power and ground pins
on the M1 layer. M1+M2 denotes power and ground pins on both M1 and M2 layers. BPR has
no power and ground pins on BEOL layers. M1 and M2 power and ground pins have width equal
to twice the minimum width on Mx (M1, M2 and M3) layers. (Thus, since minimum width and
minimum spacing are each equal to half the metal pitch MP, the power and ground pins have
widths equal to the Mx pitch.) Also, in enablement of BPR, the width of M1 power pins is the
same as the minimum width of Mx. This is because commercial P&R tools require power and
ground pins to be connected to PDN. Note that the minimum-width M1 power and ground pins
do not affect routability since the minimum routing layer is M2.
(6) CH: Cell height of standard cells, expressed as a number of M2 routing track (T) pitches. For
example, if M2 routing track pitch is 40nm and the cell height is 240nm, then CH is 6. We use
and refer to standard cells with 5T, 6T, 7T and 8T [32] [119] cell heights. Note that CH depends
on the combination of RT and PGpin. For example, if RT is 4 and PGpin is M1, then CH is 6. If
RT is 4 and PGpin is BPR, then CH is 5.
107
(7) MPO: The number of minimum pin openings (access points) per pin. For example, if MPO is
2, every pin must have at least two access points for the generated standard-cell layout. We use
the values of 2 and 3 for MPO [26] [82].
(8) DR: Design rule sets. In this work, all design rules are defined based on routing grids.
We assume that our technologies are based on Extreme Ultraviolet (EUV) lithography. We
define three fundamental grid-based design rules for Mx layers, DR-MAR, DR-EOL and DR-
VR [26] [82]; we also define two design rule sets, namely, EUV-Loose (EL) and EUV-Tight
(ET).
Figure 4.5 illustrates the three design rules. (i) DR-MAR denotes minimum area rules,
as shown in Figure 4.5(a). When a metal shape occupies only one grid point and DR-MAR is
1, this violates the DR-MAR rule. I.e., a metal shape must be long enough to occupy at least
two routing grid points. (ii) DR-EOL denotes end-of-line rules, as shown in Figures 4.5(b) and
(c). When edges of two co-linear metal shapes are placed next to each other and DR-EOL is
1, this violates the DR-EOL rule. I.e., there must be at least one unoccupied routing grid point
between edges of metal shapes. When DR-EOL is 2, there must be at least two empty routing
grid points. (iii) DR-VR denotes via restriction rules, as shown in Figure 4.5(d). The figure
shows the prohibited locations for other vias, relative to the placement of a given via. When
DR-VR is 1, only four neighbors are blocked by the DR-VR rule.
Our two design rule sets each comprise combinations of specific design rule settings, as
follows. EUV-Loose (EL) consists of DR-MAR = 1, DR-EOL = 1, and DR-VR = 1. EUV-Tight
(ET) consists of DR-MAR = 1, DR-EOL = 2, and DR-VR = 1).
Last, we note that in this work, we assume metal enclosures of vias are 10nm in a
preferred direction and 0nm in a non-preferred direction. Also, many practical design rules can
be framed using our grid-based design rules. For example, rules for end-to-end spacing and
minimum enclosures can be captured with the DR-EOL rule. We note that a wide range of via
rules, including center, edge and corner spacing rules, can be captured with the DR-VR rule.
Based on the cell architecture of Figure 4.4 and the above technology parameters, our
108
Figure 4.5. DR-MAR, DR-EOL and DR-VR design rules.
Table 4.1. Standard-cell architectures in our experiments.
Fin RT PGpin CH
M1/M1+M2 5
2 4
BPR 6
M1/M1+M2 6
3 5
BPR 7
M1/M1+M2 7
3 6
BPR 8
standard-cell layout synthesis framework can generate layouts for various cell architectures. In
particular, our studies use six types of cell architectures with the combinations of Fin, RT and
PGpin and CH shown in Table 4.1.
Definitions of Design Parameters
Our framework uses the following design parameters.

(1) BEOL: Metal stack options. We define 9M, 10M, 11M and 13M BEOL stack options based
on scaling down from a commercial 14nm technology. Recall that Mx (1X layer) pitch, i.e., MP
above, is a technology parameter that we can vary in routability exploration. To scale down a
14nm BEOL technology to sub-7nm technologies, we define 2X, 3.2X, 9X and 18X layer pitches
109
based on 40nm as the 1X pitch; this reflects advanced-node stacks as well as considerations
such as litho/cost “cliff” (∼80nm pitch limit for single-exposure 193i patterning). We calculate
the routing resource of [64] for each BEOL stack option. The routing resource is defined as
∑b (1/pitchb ) where b denotes a metal layer and pitchb is the pitch of b. Essentially, this sums
available routing tracks over all routing layers. Table 4.2 summarizes layer counts per pitch and
routing resource (R) for the BEOL stack options that we study.
(2) PDN: Power delivery network options. These include traditional PDN options with different
layers and pitch, and backside PDN as a scaling booster for advanced technology. We define four
PDN options: Backside, Sparse, Middle and Dense. Table 4.3 shows the detailed information of
these PDN options.
(3) Tool: Commercial tools to perform P&R [144] [172], referred to only as Tool A and Tool B to
comply with vendor license agreements.
(4) Util: Placement utilization (0.6, 0.7, 0.8).
(5) Design: Designs studied in routability assessment. We use knight’s tour-induced artifi-
cial topologies, along with four open-source designs (AES, LDPC, JPEG, VGA) from Open-
Cores [161]. The respective instance counts of AES, LDPC, JPEG and VGA are approximately
13K, 56K, 69K and 72K.
Table 4.4 summarizes the technology and design parameters that we use in our experi-
ments. Note that in our framework, the parameter list is flexible and readily extendable. This
enables accommodation of new technology requirements or new scaling booster options. For
example, sets of smaller CPP and MP values, including with non-unit “gear ratio” values such
as 2:1 or 3:2 (e.g., CPP relative to vertical M1 pitch), can be evaluated with the PROBE2.0
framework. This serves a real DTCO and technology pathfinding problem in industry for sub-
3nm technologies, through the end of the lateral scaling roadmap. Also, we can easily add new
designs and/or PDN strategies as design parameters, for richer assessments.
110
Table 4.2. Four BEOL stack options. R denotes routing resource.
BEOL 1X 2X 3.2X 9X 18X R

9M 3 4 NA NA 2 104.4
10M 3 4 2 NA 1 117.8
11M 3 4 2 NA 2 120.1
13M 3 2 4 2 2 116.3
Table 4.3. Details of PDN options. All numbers are pitches in units of µm for each layer. “P”
indicates that we use a given layer only for PDN at maximum area density, and do not allow the
layer to be used for signal routing. The width of M5 stripes is 0.96µm and the width of M6/M8
stripes is 1.296µm.
PDN 9M 10M 11M 13M

Backside NA NA NA NA
M5 10 M5 10 M5 10 M5 10
M6 10 M6 10 M6 10 M8 10
Sparse
M8 P M9 P M10 P M12 P
M9 P M10 P M11 P M13 P
M5 20 M5 20 M5 20 M5 20
M6 20 M6 20 M6 20 M8 20
Middle
M8 P M9 P M10 P M12 P
M9 P M10 P M11 P M13 P
M5 40 M5 40 M5 40 M5 40
M6 20 M6 40 M6 40 M8 40
Dense
M8 P M9 P M10 P M12 P
M9 P M10 P M11 P M13 P
Design Enablement
Our design enablement produces ready-to-use standard libraries and required inputs for
P&R. We generate LEF format from the primitive layout produced by SMT-based cell layout
generation. Layouts are fully grid-based, with CPP and MP technology parameters defining the
grid pitches. Importantly, we propose two pin shape selection schemes: RPA-based pin shape
selection and top-metal only pin shape selection.
RPA-Based Pin Shape Selection. The SMT-generated cell layouts can have multiple distinct
111
Table 4.4. Technology and design parameters in our experiments.
Type Parameter Option

Fin 2, 3
CPP 54, 48
MP 24, 32, 40
RT 4, 5, 6
Technology
PGpin BPR, M1, M1+M2
CH 5, 6, 7, 8
MPO 2, 3
DR EUV-Loose, EUV-Tight
BEOL 9M, 10M, 11M, 13M
PDN Backside, Sparse, Middle, Dense
Design Tool Tool A, Tool B
Util 0.6, 0.7, 0.8
Design A knight’s tour, AES, LDPC, JPEG, VGA
pin shapes for a single pin. Access to such pins must be carefully handled to avoid instability of
timing and power models of the standard cells. Figure 4.6(a) shows the initial SMT-generated
layout of an OAI21 X1 cell. In the figure, pin ZN has two M1 shapes, ZNa and ZNb . When
connections are made to different pin shapes (i.e., ZNa or ZNb ), at least one of the cases will not
match the cell’s characterized timing and power model. For example, when output pin ZN is
connected through the M1 pin shape ZNa , the cell delay is 10ps, but when the connection is made
through ZNb , the cell delay is 8ps. This instability with respect to the cell timing/power model
is unacceptable in modern design enablements. Therefore, when a pin has multiple candidate
shapes, our framework chooses one of these shapes to use in the standard-cell layout that is
produced.
Given our focus on routability, we apply pin shape selection based on the Remaining
Pin Access (RPA) pin accessibility metric [108]. Figure 4.6 shows an example of our pin shape
selection. Accessibility of a given pin is affected by other pins within a distance dint . We set
dint as 1.0 for the EL design rule set, and 2.0 for the ET design rule set. We then calculate RPA
values for each pin shape. From [108], pin access points are defined as a p,m where p is the
112
corresponding pin (i.e., single pin shape) and m is the position (index) in terms of M2 routing
tracks. A (p) denotes the set of access points for a given pin p. Then, N (a p,m ) denotes the set
of neighboring pin access points that (1) do not belong to p and (2) are within distance dint of
p on metal layer m. For example, A (A2) = {aA2,1 , aA2,2 } and N (aA2,2 ) = {aA1,2 , aZNb ,2 } in
Figure 4.6(a). The used pin access (UPA) of pin p is defined as
UPA(p) = ∑ ∑ 1/|A (p(a j ))| (4.1)

ai ∈A (p) a j ∈N (ai )
Finally, RPA of pin p is defined as
RPA(p) = |A (p)| −UPA(p) (4.2)
For example, RPA(A2) = 2.0 − (0.5 + 0.5) = 1.0 when we apply Equations (4.1) and (4.2) to
the example of Figure 4.6(a). In the same figure, the RPA values are 2.0 and 1.0 for ZNa and
ZNb , respectively. Since the RPA value of ZNa is larger than that of ZNb , we choose the ZNa pin
shape for ZN and make ZNb into an obstacle on M1, as shown in Figure 4.6(b). We empirically
confirm the benefits of this RPA-based pin shape selection for routability, as measured by the Kth
metric, in Section 4.1.4.
Top-Metal-Only Pin Shape Selection. When standard cells have multiple pin shapes on
multiple metal layers, we incorporate a top-metal-only pin shape selection step. In the example
of Figure 4.7, the initial layout of the OAI21 X1 cell has three pin shapes for the ZN pin; two on
M1 (and V1) and one on M2. In this multiple-layer situation, we do not calculate RPA values for
the pin shapes. Instead, we choose the M2 pin shape for ZN, and the two M1 pin shapes (and
V1) become obstacles. This methodology brings several benefits. (1) Replacing M1 (and V1) pin
shapes with obstacles benefits overall pin accessibility by reducing the complexity of pin access:
P&R tools can solve the pin access problem more easily and with fewer DRCs. (2) As with
the above-described pin shape selection, we avoid instability of timing and power with respect
113
dint dint
aA2,1
aA2,2
A2 B A2 B
1.0 1.5
ZNa A1 ZNb ZN A1 OBS
2.0 0.83 1.0
(a) (b)
Figure 4.6. Example of pin shape selection in a standard cell (2Fin, 5RT, 2MPO and EUV-Loose).
(a) Initial layout of OAI21 X1 from SMT-based layout generation. (b) The OAI21 X1 layout
after RPA-based pin shape selection.
A2 OBS A1
OBS
OBS M1
B V1
M2
ZN
Figure 4.7. Example of standard-cell layout with top-metal-only pin shape selection. We show
an initial layout of AOI21 X1 (2Fin, 5RT, 2MPO and EUV-Loose). In this case, the ZN pin has
only the M2 pin shape, and the M1 and V1 shapes become OBS.
to characterized models. (3) Our top-metal-only pin shape selection mitigates susceptibility to
electromigration (EM) by removing M1 pin shapes (e.g., Posser et al. [101] show a cell-internal
signal EM problem in accordance with output pin position). (4) Last, changing M1 (and V1) pin
shapes into obstacles improves accessibility of other neighboring M1 pins.
Power and Ground Pin Generation. To make SMT-produced layouts usable by P&R tools,
power and ground pins must be added. Figure 4.8 shows how we define power and ground pins
for standard cells. For BPR, the width of power and ground pins is equal to the minimum M1
114
M2 Track Cell boundary M1 M2
VDD VDD
VDD
A2 B A2 B A2 B
ZNa A1 ZNb ZNa A1 ZNb ZNa A1 ZNb
VSS (a) (b) (c)

VSS VSS
Figure 4.8. Power and ground pin (PGpin) examples: (a) BPR, (b) M1 and (c) M1+M2. The
signal pin shapes (gray) are the same regardless of PGpin.
width. For M1 and M1+M2, it is double this width. M2 routing grids of standard cells with BPR
have an offset by half of the M2 track pitch. The figure also shows how the height of standard
cells is determined by the RT and PGpin parameters.
Liberty and Technology File Generation. The remaining parts of our design enablement
generate Liberty and technology files. We use dummy Liberty files to avoid any potential errors
in the commercial P&R tools. Our technology file generation takes as inputs the technology
parameters CPP, MP and DR, and converts design rules into corresponding file formats (e.g.,
LEF and routing technology files) to enable use of commercial P&R tools.
4.1.3 Routability Assessment and Learning-Based Kth Prediction
We now describe how we perform both cell-level and design-level routability assessment
using the PROBE2.0 framework. Cell-level routability assessment aims to evaluate the intrinsic
routability of an individual standard cell’s layout. Design-level routability assessment uses a real
design testcase to evaluate the intrinsic routability of a design enablement and/or methodology.
115
Recall from Section 4.1.1 that the PROBE approach [64] starts from an initial placement of an
initial netlist, then progressively increases K, the amount of “tangling”, until the router exceeds
a given DRC limit at the threshold Kth . The Kth value is specific to a given library along with
technology and design parameters. To find Kth , multiple P&R runs are launched with various
values of K, as illustrated in Figure 4.2(b).
Cell-Level Routability Assessment Based on Knight’s Tour
In advanced technologies, pin accessibility of individual standard-cell masters has become

a critical challenge with decreasing cell heights. We therefore propose cell-level routability
assessment to evaluate routability at the individual cell level. The previous work of [64] uses
baseline mesh-like placements to rank BEOL stacks by routability. These induce nearly square-
mesh netlist topologies constructed using a single cell master from a fixed library. By contrast, our
goal is to assess routability for a given choice of technology and design parameters, leveraging our
standard-cell generation capability. The mesh-like placements used in [64] have key drawbacks:
(i) they do not closely reflect the wirelength and Rent parameter attributes of real design testcases,
and (ii) their construction encompasses only 2- and 3-input cells. In this work, we overcome
these drawbacks by using a knight’s tour to induce the initial placed netlist that is “tangled” to
find Kth .
A knight’s tour is a sequence of knight’s moves that visits each square on a chessboard
exactly once. A knight moves from its current square to a square that is two by one (or one
by two) squares away, in any direction. In our approach, a square in an M by N chessboard
corresponds to a placed cell in an initial placement of M by N instances of a given master. A
knight’s tour ordering induces connections – i.e., a netlist – within this placement. More precisely,
for a given layout region with M rows and N columns of placed instances, and a k-input cell
master, we create a knight’s tour-based topology as follows.
(1) In the layout region with M rows by N columns of instances, we generate a knight’s tour
ordering. For this, we use the well-known Warnsdorff-rule algorithm [139].
116
Figure 4.9. Example of knight’s tour-based connections in a placement of 2-input cells. Numbers
in red circles indicate the knight’s tour ordering.
(2) An instance of the k-input cell master that is at position i in the knight’s tour is assigned
fanins from cells at positions i − 1 to i − k in the knight’s tour ordering. For a given 2-input cell
i, there will be a fanin connection from the output of cell i − 1 to the first input of cell i, and a
fanin connection from the output of cell i − 2 to the second input of cell i.
(3) We end up with an initial placement (“chessboard”) having M rows and N columns of
instances, and an artificial netlist having connections induced by the knight’s tour as described
above.
Figure 4.9 shows an example of a knight’s tour topology with 2-input cells. The red
arrows indicate the connections between cell i and cell i − 1. The green arrows indicate the
connections between cell i and cell i − 2.
Our studies confirm improved flexibility and correspondences to real design netlists
when we apply the knight’s tour-based approach. Table 4.5 shows placed wirelengths of the
AES design, and of mesh-like placements [64] and knight’s tour-induced netlists. INV X1,
NAND2 X1, NAND3 X1 and NAND4 X1 cell masters are used, with default technology and
design parameters as specified in Section 4.1.4. Particularly when based on the NAND3 X1
master, the knight’s tour topologies can give more realistic wirelengths than mesh-like placements.
117
Table 4.5. Placed design information for AES, mesh-like placements and knight’s tour based
topologies. Mesh-like placements and knight’s tours have 115 × 115 placed instances. Area
denotes total instance area in the designs. Area is in units of µm2 and wirelength (WL) is in
units of µm.
Design Cell Area #Insts WL

AES - 1088 13221 46880
INV X1 Not supported
NAND2 X1 799 13225 11001
Mesh
NAND3 X1 999 13225 13932
NAND4 X1 Not supported
INV X1 599 13225 10264
NAND2 X1 799 13225 23610
Knight’s Tour
NAND3 X1 999 13225 41103
NAND4 X1 1199 13225 62251
(As noted above, the mesh-like placement construction does not support use of 1- or 4-input cell
masters.)
We also note that mesh-like placements have a simple topology wherein connections
are only between neighbor cells in the initial placement. This inflexibility can weaken corre-
spondences to real-world designs. Notably, the Rent parameter p of the (2-dimensional) mesh
topology is 0.5 [47] [81] according to Rent’s rule. We have studied achievable Rent parameters
for mesh-like placements and knight’s tour-based topologies by using RentCon [165]. Table 4.6
shows that Rent parameters of real circuits (AES, LDPC, JPEG, VGA) range from 0.617 to 0.891.
On the other hand, the mesh-like placements p values from 0.500 to 0.585, while our proposed
knight’s tour-based placements have p values from 0.676 to 0.820. Again, the knight’s tour
construction can more closely reflect real circuits. Based on these studies, we use the knight’s
tour-based topology construction in our cell-level routability assessment.
Design-Level Routability Assessment
To perform design-level routability assessments, we begin with open-source RTL de-

signs from OpenCores [161] and use commercial logic synthesis to obtain gate-level netlists.
118
Table 4.6. Rent parameters for mesh-like and knight’s tour-based placed netlists. We use
rectangle sampling-based methods I and II for the evaluation method, Types I and II for the pin
counting method, and geometric mean for the averaging method in [165].
Source Design Rent parameter

AES [0.726, 0.832]
JPEG [0.617, 0.813]
OpenCores [161]
VGA [0.656, 0.879]
LDPC [0.812, 0.891]
Mesh (NAND2 X1) [0.530, 0.535]
PROBE [64]
Mesh (NAND3 X1) [0.500, 0.585]
Knight’s tour (NAND2 X1) [0.676, 0.761]
PROBE2.0
Knight’s tour (NAND3 X1) [0.708, 0.820]
Following [64], we evaluate design-level routability based on cell width-regularized placements

(Figure 4.3(b)) of these netlists. For each standard-cell library, we modify LEF such that all
combinational cells are bloated to have the same width as the maximum-width cell among all
38 combinational cells in the library. (Our designs also instantiate D flip-flops, which have the
largest cell width in our generated standard-cell libraries. We fix the locations of flip-flops in
designs after placement and do not swap them.)
Cell width regularization is proposed in the work of [64]. Its purpose is to avoid illegal
placements (with cell overlap) when neighbor cells are swapped during “tangling”. Without
the cell width regularization, the neighbor-cell swap operations might cause cell overlap due to
different cell widths. Of course, the standard P&R flow does not use width-regularized cells,
which raises the question of whether use of bloated cells could lead to misleading design-level
routability assessments. Our experimental study of Kth obtained with bloated cells, versus
maximum achievable utilization obtained with unbloated cells, is reported and offers some
reassurance in this regard. Aside from this, we use bloated cells for two main reasons. (1)
Our cell-level routability assessment focuses on intrinsic routability of cell layouts without any
bloating of cell widths. (2) Despite the extra widths of cells, we can still assess pin accessibility
for cells. Bloated cells still retain their original order and shapes of pins. So, the intrinsic
119
routability for standard cells can still be measured. Below, design-level routability assessments
with various technology and design parameters are reported.
Learning-Based Kth Prediction
Although Kth provides a useful means of routability assessment, it has three potential
logistic disadvantages: large runtimes, large data footprints, and high consumption of commercial
EDA licenses. First, as we perform neighbor-swaps to increase routing difficulty for a given
placement, the increasing number of DRCs leads to large runtimes. This is because P&R tools
perform increasingly expensive search-and-repair iterations in the detailed router. Furthermore,
although we set a maximum runtime per each P&R run,21 our overall runtime burden is still
large, since the total number of tool runs itself is large. For example, when we perform P&R with
K = 1 to 30 to obtain Kth , and there are just 100 sets of standard-cell libraries to be evaluated, up
to 3000 P&R runs are required for this assessment. Second, large amounts of disk space may
be used by our framework. For example, just 3000 runs of P&R with the VGA design occupy
around 300GB of disk space in our experiments. Third, in many contexts the number of available
commercial P&R tool licenses will also be limited.
To mitigate these potential disadvantages, we have developed learning-based Kth pre-
diction. That is, to reduce the number of tool runs (thus saving runtime, disk space and license
usage), we predict Kth via machine learning (ML) techniques, without performing all P&R runs.
Figures 4.2(b) and (c) show PROBE2.0 flows without, and with, Kth prediction. The flow without
Kth prediction in Figure 4.2(b) requires generation steps and multiple P&R runs. On the other
hand, the flow with prediction in Figure 4.2(c) finds Kth using only inference with a trained
ML model. We use 12 input features in our ML modeling, based on the technology and design
parameters in Section 4.1.2. Table 4.7 summarizes the types, names, and data types of the input
features that we use in the Kth prediction. From the original set of technology parameters, we
21We assume that runs with #DRCs under/around the threshold are finished within the maximum runtime. We set
maximum runtimes based on the instance counts of the designs. In this work, we use maximum runtimes of two
hours for a knight’s tour and AES, and three hours for LDPC, JPEG and VGA.
120
Table 4.7. Input features of the Kth prediction.
Type Name Data Notes

CPP Integer
MP Integer
PGpin Category BPR, M1, M1+M2
Technology
CH Integer
MPO Integer
DR Category EL, ET
BEOL Integer
PDN Category Backside, Sparse, Middle, Dense
Tool Category Tool A, Tool B
Design
Util Float [0.0, 1.0]
Inst Integer
Net Integer
omit Fin and RT from our feature list since Fin and RT are derivable from PGpin and CH. From
the original set of design parameters, we omit “Design” and add the number of instances (Inst)
and the number of nets (Net) as design features in our model.
We present our experimental results for the ML-based Kth prediction. Importantly, our
ML models are trained in less than an hour using approximately 1968 data points (80% out of
a total of 2460 data points); inference then requires only seconds for any technology-design
parameter combination. By contrast, collecting a single Kth value for a given technology-design
parameter combination requires not only the generation of standard-cell layouts and design
enablements, but also up to 30 P&R runs that each consume significant average and maximum
tool runtimes of 0.4 and 3.0 hours, respectively. Thus, runtime overhead is 12 (resp. 90) hours
on average (resp. at most) per Kth value. In light of such large runtimes, the primary goal of our
ML-based Kth prediction is to reduce the number of required P&R runs by predicting unseen
data (that is, Kth values) with reasonably small errors. In Section 4.1.4 below, we also perform
an experiment to show the tradeoff of model accuracy versus training data overheads.
121
Table 4.8. Default parameters in our experiments.
Design parameter Technology parameter

Name Value Name Value
Fin 3 BEOL 10M
CPP 54 PDN Middle
MP 40 Tool A
RT 5 Util 0.7
PGpin M1 Design AES
CH 7 - -
MPO 2 - -
DR EL - -
We demonstrate routability assessment capability of PROBE2.0 via five main experi-

ments.
• Expt. 1: Cell-level assessment with a knight’s tour.
• Expt. 2: Design-level assessment with technology parameters.
• Expt. 3: Design-level assessment with design parameters.
• Expt. 4: Achievable utilization study with Kth results.
• Expt. 5: Learning-based Kth prediction.
We first show routability assessments using the PROBE2.0 framework. Expt. 1 demon-
strates evaluation of cell-level routability using our knight’s tour-based assessment. Expts. 2 and
3 perform design-level routability assessments with various technology and design parameters.
Next, Expt. 4 studies achievable utilization in relation to the Kth routability metric, across a
range of generated standard-cell libraries. Last, Expt. 5 demonstrates how learning-based Kth
prediction can mitigate potential logistical overheads of our framework.
122
Table 4.9. List of 39 standard cells per generated library.
Cell list Size

Inverter (INV), Buffer (BUF) X1, X2, X4, X8
2-input AND/OR/NAND/NOR (AND2/OR2/NAND2/NOR2) X1, X2
4-input NAND/NOR (NAND4/NOR4) X1, X2
2-1 AND-OR-Inverter (AOI21), 2-2 AND-OR-Inverter (AOI22) X1, X2
2-1 OR-AND-Inverter (OAI21), 2-2 OR-AND-Inverter (OAI22) X1, X2
2-input MUX/XOR (MUX2/XOR2), D flip-flop (DFF) X1
Experimental Setup
We generate 39 cells (38 combinational and 1 sequential) for each standard-cell library
that we study, as listed in Table 4.9. We name cells according to functionality, number of inputs,
and size. For example, INV X1 is an X1-sized inverter, and AND3 X2 is an X2-sized 3-input
AND gate. Our SMT-based standard-cell layout generation tool is open-sourced in [169].
Our experiments are based on a commercial 14nm technology. We modify design rules
of Mx (M1, M2 and M3) layers (and of vias above) in routing technology files to enable scaling
to sub-7nm technologies. We draw FEOL and BEOL (Mx) layer parameters from [32] [119].
Layer parameters and rules for M4 and above are drawn from the 14nm technology. For example,
we choose the 9M (9-layer) BEOL stack option from the 14nm technology, and modify rules
for the Mx layers (M1, M2, M3). We use existing design rules for the upper layers (M4 to M9).
Mx layers are constrained to use unidirectional and min-width routing for standard-cell layout
generation and P&R. We also assume that M1 is unavailable for signal routing.
All design rules are described in LEF format and routing technology file format [172]
for use by commercial P&R tools. SMT-generated standard-cell layouts are converted into LEF
format. For cell-level routability assessments, we use a knight’s tour with 50 by 50 instances,
constructed using an open-sourced implementation of the Warnsdorff-rule algorithm [139]. For
design-level assessments, we perform synthesis by using a commercial tool [171] and P&R by
123
Figure 4.10. Kth results from knight’s tour-based cell-level routability assessments across various
track heights and cell masters. 15 cell masters from 6T, 7T and 8T standard-cell libraries are
used in this experiment.
using two commercial P&R tools [144] [172].

We also use GNU parallel [152] to perform multiple evaluations in parallel. We assume
that our designs with #DRCs less than the threshold finish within three hours. Setting the
maximum runtime in GNU parallel reduces the total runtime of the framework. We use 500
DRCs as the DRC threshold to determine Kth . (The work of PROBE [64] used a threshold of
150 DRCs. We set the higher threshold of 500 DRCs since SMT-generated standard-cell layouts
might not be as well-optimized for pin accessibility as mature libraries from industry.) Finally,
we set default parameters as shown in Table 4.8. In all of our experiments, when not otherwise
specified, we use these pre-defined default parameters.
Expt. 1: Cell-Level Assessment with a Knight’s Tour
In this experiment, we perform knight’s tour-based cell-level routability assessments. To

show Kth of standard cells with various cell heights, we draw cells from 6T, 7T and 8T libraries.
Among the library cells, we choose 15 combinational cells that have X1 size and more than one
input pins.
Figure 4.10 shows Kth results from the cell-level routability assessment. From the results,
we observe larger Kth as cell height increases. Also, by examining Kth per cell, we can see which
cell types are routing-critical. For example, 4-input cells have relatively small values of Kth
according to this experiment. Thus, at an early stage of technology development, cell-specific
124
Figure 4.11. Cell-level routability assessments showing #DRCs versus K with different pin
shape selection methods. RPA-Best (B) and RPA-Worst (W) pin shape selection methods are
shown. Solid lines are for cells with RPA-Best, and dashed lines are for cells with RPA-Worst.
Kth is a promising indicator to determine which cells need to be improved in terms of routability.
We also show the benefit of our RPA-based pin selection. We generate standard-cell
layouts using two pin shape selection methods, RPA-Best and RPA-Worst. RPA-Best denotes
our proposed RPA-based pin shape selection which chooses the pin shape with the largest RPA
value. To show that RPA-Best brings meaningful benefits, we also create standard-cell libraries
by choosing a pin shape per pin that has smallest RPA value; this yields the RPA-Worst method.
We study six cells from 6T libraries and two cells from 7T libraries. We perform cell-level
routability assessments on RPA-Best and RPA-Worst versions of these standard cells.
Figure 4.11 plots the number of DRCs versus K, for the eight standard cells with different
pin shape selection methods. Also, Table 4.10 shows the Kth values for each standard cell. Seven
standard cells (out of eight cells) with RPA-Best show larger Kth than those with RPA-Worst.
One cell (OAI X2) shows the same Kth for both methods. We see that our RPA-based pin shape
selection helps to improve the routability of standard cells.
125
Table 4.10. Kth results with RPA-Best and RPA-Worst pin shape selection methods. This table
shows the Kth values corresponding to the data shown in Figure 4.11.
Kth
CH Cell
RPA-Best RPA-Worst
AOI22 X2 11 0
NAND3 X2 18 16
NAND4 X2 11 10
6T
NOR3 X2 21 20
NOR4 X2 14 13
OAI21 X1 2 2
AOI21 X2 28 27
7T
OAI21 X2 28 27
Expt. 2: Design-Level Assessment with Technology Parameters
In this experiment, we perform design-level routability experiments with various tech-

nology parameters. The goal is to show the routability impact of technology and standard-cell
architecture options. Figure 4.12 shows Kth results that highlight the individual impacts of five
technology parameters. We also experimentally confirm the ability of our framework to assess
various design rule options. To study design rule choices, we define two scenarios, DR1 and
DR2. DR1 is a scenario which uses CPP = 54 and the EL (EUV-Loose) design rule set. DR2 uses
CPP = 48 and the ET (EUV-Tight) design rule set. We choose these two design rule scenarios so
that they are more likely to induce different standard-cell architectures.
From the results in Figure 4.12, we make five observations, as follows.
(1) As cell heights (or routing tracks) increase, the corresponding Kth values also increase.
(2) Standard cells with M1 PGpin (and the same RT) have better routability than those with BPR
or M1+M2 PGpin. Also, standard cells with BPR and M1+M2 show similar Kth values. This
reflects two phenomena. First, cells with BPR have one track less cell height than corresponding
non-BPR cells. Second, cells with M1+M2 have worse routability than those with M1. This
is because the M2 power and ground pins in standard cells with M1+M2 will block routing
resources over the cells.
126
Figure 4.12. Design-level assessments with various technology parameters.
(3) Smaller pitches of Mx layers bring lower Kth values, despite smaller pitches nominally
providing more routing resources (total number of wiring tracks) on Mx layers.
(4) Minimum pin openings of our standard-cell layouts have a clear impact on routability.
Standard cells with 3 MPO show better routability than those with 2 MPO.
(5) For the DR cases as shown in Figure 4.12, DR2 has smaller Kth than DR1, implying that the
tight design rule (DR-EOL in the scenario DR2) has a harmful impact on routability.
With respect to the fifth observation, we note that while DR2 brings smaller Kth , standard-
cell layouts with 48nm CPP (DR2) have around 11% less area than layouts with 54nm CPP
(DR1). We explore the relationship between area (utilization) and Kth . Finally, Table 4.11 gives
additional details of Kth results for all the possible combinations with various Fin, MP, RT, PGpin
and MPO in our experiments. (Data shown in Figure 4.12 is a subset of the table data.)
Expt. 3: Design-Level Assessment with Design Parameters
In this experiment, we study routability impacts of design parameter options. Figure 4.13
shows the Kth results from this study, which spans BEOL, PDN, Tool, Util and Design parameters.
We make the following observations.
(1) According to the Kth values with the various BEOL stack options, 10M has better routability
127
Table 4.11. Kth results with various Fin, MP, RT, PGpin and MPO.
Kth
Fin RT MP MPO
BPR M1 M1+M2
2 2 11 4
24
3 8 11 6
2 10 15 9
2 4 32
3 11 16 10
2 10 20 13
40
3 13 23 16
2 4 16 10
24
3 6 19 11
2 17 27 17
3 5 32
3 19 28 18
2 21 30 21
40
3 24 32 24
2 17 28 19
24
3 19 27 19
2 27 44 25
3 6 32
3 28 45 27
2 32 44 34
40
3 33 45 36
than 9M, and 11M has better routability than 10M. Also, 10M and 11M have better routability
than 13M. As shown in Table 4.2, the routing resources of 10M and 11M are larger than those of
13M. This illustrates how overall routing capacity can worsen even if the number of metal layers
is increased. However, the additional layers bring other benefits. For example, we may implement
a denser, more robust PDN resulting in better design performance due to less resistivity of wide
metal layers.
(2) As fewer layer resources are used for PDN, the results show better routability.
(3) The Kth results from Tool A dominate those of Tool B. In this sense, our framework also has
the ability to evaluate placers and routers, as does the previous work of [64].
(4) As Util increases, the Kth values decrease since routing with denser placements is more
difficult.
128
Figure 4.13. Kth results of design-level routability assessments with various design parameters.
(5) The base designs used in the design-level routability assessments also affect the results. In
particular, designs with larger instance counts have smaller Kth . This is expected: for a given
amount of “tangling” (K), and all else being equal, a larger design’s placement is expected to
have more routing #DRCs. ([64] gives a probabilistic analysis of routing hotspots and routing
failure according to the number of instances.)
Expt. 4: Achievable Utilization versus Kth
In production IC design, maximizing the utilization of die area is a common objective

toward achieving the best PPAC. Also, the maximum achievable utilization is a first-order
indicator for routability, even as it captures the A, C dimensions of PPAC. In this experiment,
we explore the relationship between achievable utilization (in design implementations using
unbloated cells) and Kth (in design-level routability assessment using cell width-regularized
libraries). Such a study could inform the future use of Kth to predict die utilization and area.
We define achievable utilization as the highest utilization for which #DRCs is smaller than the
pre-defined DRC threshold.
In our study of achievable utilization, we use two designs, AES and LDPC. For AES, we
129
1 2MPO_BPR 6T
2MPO_M1 7T
0.9 2MPO_M2
3MPO_BPR
0.8 AES
3MPO_M1 5T
Achievable Utilization 0.7 3MPO_M2
Util LDPC design
0.4
0.6
0.5
LDPC 0.3
0.4
0.3
0.2 0.2 Kth
3 8 13 18
0.1
0 10 20 30 40 50
Kth
Figure 4.14. Comparison between Kth and achievable utilization for the AES and LDPC designs.
collect achievable utilizations from 0.60 to 0.99, with step size of 0.01. For LDPC, which is a
notoriously difficult to route testcase, we collect achievable utilizations from 0.20 to 0.40, again
with step size of 0.01. To extract Kth , we set Util as 0.7 for AES and 0.2 for LDPC. Figure 4.14
shows our Kth results and the corresponding achievable utilizations. The plot shows that Kth
values and achievable utilizations are sensibly related for the given design parameters. The data
also show how proper combination of technology and design parameter choices can potentially
bring routing-friendliness and high utilization: large cell heights, large cell areas, sparse PDN,
ample routing resources, and/or relaxed design rules all lead to routability with maximum
utilization. Especially, in the enlarged view of LDPC design plot, we observe similar tendencies
as seen in Figure 4.12. For example, the cells with M1 PGpin have better Kth compared to
cells with M2 and BPR. Also, both Kth values and achievable utilizations gradually increase
with cell heights. The cells with 2MPO and 3MPO have similar Kth values and achievable
utilizations. This experiment hints that Kth could be useful in a predictor of design-specific
achievable utilization. We leave this possibility for future work.
130
Expt. 5: Learning-Based Kth Prediction
We now show results from our learning-based Kth prediction. We compile a dataset of
2460 Kth values corresponding to a wide range of settings for the various features described in
Table 4.4. We use the open-source AutoML package [154] to predict Kth . AutoML has a hyper-
parameter tuning ability on various models including gradient boosting machine (GBM) [41],
XGBoost [20], distributed random forest (DRF), extremely randomized tree (XRT) [42], deep
learning (DL) and generalized linear model (GLM) [97]. AutoML also suggests combined
models that outperform a single model.
We down-sample and up-sample our data since the data are imbalanced in terms of Kth
distributions.22 We use 60 as a target sample number to generate balanced data. We randomly
partition our augmented (i.e., after sampling schemes are applied) dataset as 80% used for
training and 20% used for testing. We perform all experiments using an Intel Xeon Gold 6148
2.40GHz server (80 threads) with 256GB RAM.
We use the 3.30.0.6 version of AutoML to train our models, with default input parameter
settings for AutoML. The default settings of AutoML include nfolds = 5 for cross-validation,
leaderboard frame = testing set, and sort metric = mean residual deviance to rank the trained
regression models. Details of these settings are found at [154]. We set the training time limit as
one hour with 80 threads. The suggested models from AutoML and the corresponding trained
results are described in Table 4.12. Note that Deviance denotes mean residual deviance, RMSE
denotes root mean squared error, and MAE denotes mean absolute error.
Figure 4.15 shows comparisons of golden Kth (i.e., ground truth data which comes
from actual results obtained from our framework) versus predicted Kth using the model Stacke-
dEnsemble BestOfFamily. The StackedEnsemble BestOfFamily contains six models: GBM,
XGBoost, DRF, XRT, DL and GLM. AutoML generates a stacked ensemble model with GLM
22When generated standard cells are not routing-friendly, routing for designs that use those cells might be
infeasible. This causes the observed distribution of Kth values to be biased toward zero. The down-sampling and
up-sampling schemes help us to avoid over-fitting in the prediction of Kth .
131
Table 4.12. Results on training data of Top-5 ML models by using AutoML. The three GBM
models (GBM 1, GBM 2 and GBM 3) have different configurations for hyperparameters. The
training runtime is limited to one hour with 80 CPU threads.
Model Deviance RMSE MAE

StackedEnsemble BestOfFamily 0.733 0.857 0.353
XGBoost 0.865 0.930 0.300
GBM 1 0.896 0.947 0.374
GBM 2 0.947 0.973 0.350
GBM 3 0.950 0.975 0.440
Table 4.13. Results of DL modeling with AutoML. DL 1h (resp. DL 24h) denotes a DL model
with a 1-hour (resp. 24-hour) runtime limit for training.
Model Deviance RMSE MAE

DL 1h 6.275 2.505 1.789
DL 24h 1.469 1.212 0.615
as meta learner, based on those six models. We show max error and avg error metrics, which
are the maximum and average difference between golden and predicted Kth , respectively. Fig-
ure 4.15(c) shows a balanced error distribution for the testing set. To further understand our Kth
prediction, we extract feature importance from the XGBoost model. We do this for XGBoost
since StackedEnsemble BestOfFamily consists of six models and does not support such a feature
extraction. The extracted feature importance is visualized in Figure 4.15(d). We conclude that
Table 4.14. Model accuracy results with different amounts of training data (TD), i.e., 20, 40, 60
and 80% of the total data (2460 Kth values), with a one-hour limit and 80 threads. Testing data
are fixed for an apples-to-apples comparison. ME and AE respectively denote maximum and
average error of Kth prediction.
TD ME AE Deviance RMSE MAE

20% 14.696 1.155 4.916 2.217 1.155
40% 8.131 0.610 1.702 1.304 0.610
60% 8.553 0.392 1.006 1.003 0.392
80% 8.459 0.353 0.733 0.857 0.353
132
Kth is mostly affected by Inst, CH and BEOL.
From the results in Table 4.12, we can infer that boosting ML models (XGBoost and
GBM) outperform DL models in our experiment. (Put another way: no DL model made it into
the AutoML Top-5 models.) However, given the recent intense interest in deep learning, we
also explicitly study prediction with DL models. Table 4.13 shows the DL model results with
AutoML, for the same training set as in Table 4.12. In this experiment, we train the DL models
with two maximum runtimes for training. DL 1h denotes a DL model obtained with the same
1-hour training runtime limit used for Table 4.12. DL 24h denotes a DL model with 24-hour
training runtime limit. Even with the longer training time, DL does not surpass the other ML
models in Table 4.12. (The recent study [39] provides insights on boosting models outperforming
DL models.)
As noted in Section 4.1.3, a primary benefit of learning-based Kth prediction (Fig-
ure 4.2(c)) is that it can save up to 90 hours of P&R tool runtime with each predicted value. To
characterize the tradeoff of model accuracy versus training data overheads, we further conduct an
experiment with different numbers of training data. Since obtaining a single Kth requires 30 P&R
runs (averaging 0.4 hours each) in our experiment, obtaining a total of 2460 Kth values requires
approximately 15.4 days of runtime. We have studied Kth prediction with various numbers of
training data, and with fixed testing data. Table 4.14 shows results with different amounts of
training data (∼20%, ∼40%, ∼60% and ∼80% of the total dataset size). Generating 20% of the
total data as training data reduces schedule overhead to approximately 3.1 days, and leads to
1.155 average Kth error in predicting the held-out 20% test data. As expected, more training data
leads to better model accuracy (smaller errors) at the cost of schedule overhead.
4.1.5 Conclusion
We have proposed a novel framework to evaluate routability impacts of advanced-node

scaling options, spanning across technology, design enablement, and design parameters. Several
options for scaling boosters are assessed in terms of their routability impacts. Our framework is
133
Figure 4.15. Comparison between golden Kth and predicted Kth and extracted feature importance;
(a) training data, (b) testing data, (c) error distribution on testing data with kernel density
estimation (KDE) plot and (d) extracted feature importance.
well-matched to the needs of DTCO, particularly for early stages of technology development.
Crucially, our framework can reduce the timelines for design enablement and design imple-
mentation that limit today’s DTCO methodologies. Also, our framework can flexibly support
additional technology and design options.
We integrate a powerful SMT-based standard-cell layout generation capability. Optimal
layout solutions are obtained via a unified constraint satisfaction formulation that spans technol-
ogy and cell architecture parameters. We also build upon the previous routability assessment
framework of [64].
Our new framework provides automatic generation of cell libraries and collaterals (LEF,
134
Liberty, routing technology files) for commercial P&R tooling. We propose RPA-based and top-
metal-only pin shape selection to improve routability of our generated standard-cell libraries. We
validate these aspects of our methodology using a novel cell-based routability assessment with
knight’s tour-based topology generation. We also perform design-based routability assessment
using open-source testcases, and show correlations of the Kth routability metric to achievable
utilizations in P&R. Furthermore, we propose learning-based Kth prediction to reduce runtimes
and required disk space, and to mitigate P&R tool license overheads. Finally, experimental
studies confirm the capability of our framework to produce routability assessments across a large
range of technology and design parameters.
Open directions for future research include the following.
• Our framework focuses on the Area and Cost dimensions of PPAC. This has value in early
technology development, especially since density is the dominant driver for foundry tech-
nology and design enablement [123]. However, future work should broaden assessments
to include power and performance evaluations, along lines presented in Section 4.3 below.
Automation and/or prediction of library characterizations is a related challenge for future
research.
• Our framework today covers a number of technology options, design rules and standard-
cell architectures. However, extensions to support other technology options and/or design
rules for advanced technologies may be desirable. For example, a target technology might
require self-aligned double patterning (SADP) design rules and bidirectional routing. Also,
new standard-cell architectures might be required, such as multi-height standard cells and
rectilinear pin shapes.
• Extending our framework to incorporate an open-source P&R tool, such as [2] [176], may
help to scale exploration and turnaround times beyond the limits of available commercial
P&R tool licenses. In this context, learning to map routability and other PPAC-related
assessments between implementation tool chains will be a valuable future contribution.
135
4.2 A Novel Framework for DTCO: Fast and Automatic
Routability Assessment with Machine Learning for
Sub-3nm Technology Options
Design-technology co-optimization (DTCO) is an essential, high-value element of tech-

nology enablement. However, DTCO has proved to be a very expensive process; large turnaround
times and engineering efforts are needed to develop standard-cell libraries and perform compre-
hensive block implementation experiments. Turnaround time of DTCO at advanced nodes is
weeks to months with hundreds of engineers. Meanwhile, standard cells have been optimized
to aim for minimum cell area (cell height) for decades. However, cell-level area scaling does
not necessarily result in block-level area scaling at advanced nodes. As standard-cell area is
scaled down, local routing and pin accessibility of standard cells become significantly difficult.
Various scaling boosters, such as buried power rails (BPR) [102], backside PDN and supervias,
are proposed to improve power, performance, area and cost (PPAC). Scaling boosters also have
an impact on routability. Due to significant routability impacts of standard-cell architectures and
scaling boosters, block-level area must be estimated at an early stage of technology development.
Recently, DTCO methods for technology definition have been proposed in many aspects.
[12] proposes a machine learning (ML)-based approach to find optimal combinations of design,
technology and other ingredients for high-performance CPU designs. However, there is no
automatic design enablement stage (i.e., standard-cell library generation), and the methodology
requires four to six weeks of turnaround time. [133] proposes an ML-based modeling framework
for devices. The framework can generate compact models of novel devices without prior
knowledge. On the other hand, the work does not consider block-level evaluations.
In this work, we apply the PROBE2.0 framework [24] (Section 4.1 of this thesis) to eval-
uate various technology options and configurations for sub-3nm technology nodes. We generate
448 unique standard-cell architectures, combining cell height (CH, 80∼150nm), contacted poly
pitch (CPP, 39∼57nm), metal pitch (MP, 16∼30nm) and use/non-use of buried power rails
136
Figure 4.16. (a) Overall flow of PROBE2.0 [24] and (b) machine learning-based Kth prediction.
(BPR). We study achievable block-level area with automatically generated standard-cell libraries
across four main axes: (i) available M2 routing tracks (RT) on standard cells, (ii) use of BPR,
(iii) gear ratios for MP to CPP, and (iv) different available M2 RT with same CH. We also apply
ML-assisted routability prediction to expedite assessment.
4.2.1 DTCO Framework and Configuration for Sub-3nm Nodes
In this section, we summarize PROBE2.0 [24], the framework used for routability
assessment in DTCO, along with the configurations for sub-3nm standard-cell libraries that we
study. As shown in Figure 4.16(a), the PROBE2.0 framework includes automatic standard-cell
layout generation and design enablement generation, and (cell- and design-level) routability
assessments. The basic idea of PROBE2.0 is to measure inherent routability, represented by
Kth metric, through neighbor-swap operations applied to canonical placements. Figure 4.17
illustrates the neighbor-swap operation. For a given cell, we randomly choose a neighboring
cell and swap the locations of the cell pair. Neighbor-swaps progressively increase routing
difficulty by “tangling” the placement. We let K denote the number of neighbor-swap operations,
normalized to the number of instances (standard-cells) in the design. As K increases, the number
of post-routing design rule check violations (#DRCs) likely increases; we define Kth as the
137
Figure 4.17. An example of neighbor-swap operation. The blue-colored cell is swapped with
one of the red-colored neighboring cells.
maximum K for which #DRCs is less than a prescribed threshold. PROBE2.0 advances over the
previous PROBE [64] with a satisfiability modulo theory (SMT)-based “automatic” standard-cell
layout generation [26] [82] and an ML-based Kth prediction, as illustrated in Figure 4.16(b).
In this work, we generate nine standard-cell libraries to study routability and achievable
block-level cell density at sub-3nm technology nodes. Table 4.16 shows sets of parameters for
the nine libraries, as follows. (i) Lib{1,2,5,6} have 4 RT while Lib{3,4,7,8} have 5 RT. (ii)
Lib{1,3,5,7,9} have BPR for power and ground pins while Lib{2,4,6,8} have M1 pins. (iii)
Lib{1,2,3,4} have a 1:2 ratio for MP to CPP (20nm MP and 40nm CPP) and Lib{5,6,7,8} have a
2:3 ratio for MP to CPP (26nm MP and 39nm CPP). (iv) Lib{3,9} have 120nm CH but Lib3 has
5 RT and Lib9 has four tracks according to MP of Lib{3,9}. Figure 4.18 illustrates OAI21 X1
gates in the nine libraries.
Table 4.15. Technology and design parameters in this work. We use all the parameter types
defined in [24].
Param. type Name Option Name Option

Fin 2, 3 PGpin BPR, M1
CPP 39, 40 CH 5, 6, 7
Technology
MP 20, 24, 26 MPO 3
RT 4, 5 DR EUV-Loose
BEOL 13M Tool Tool A
Design PDN Sparce Util 0.6
Design Knight’s tour, AES, LDPC, JPEG, VGA
138
Table 4.16. Technology parameters for the nine standard-cell libraries which we study in this
work.
Name Fin CPP MP RT PGpin CH (T) CH (nm)

Lib1 2 40 20 4 BPR 5T 100
Lib2 2 40 20 4 M1 6T 120
Lib3 3 40 20 5 BPR 6T 120
Lib4 3 40 20 5 M1 7T 140
Lib5 2 39 26 4 BPR 5T 130
Lib6 2 39 26 4 M1 6T 156
Lib7 3 39 26 5 BPR 6T 156
Lib8 3 39 26 5 M1 7T 182
Lib9 2 40 24 4 BPR 5T 120
Figure 4.18. Standard-cell layout of the nine libraries in Table 4.16. OAI21 X1 of Lib1 to Lib9
are illustrated in (a) to (i), respectively. We show a LEF view (cell boundary and pin) of layout
of OAI21 X1 per library.
4.2.2 Routability Assessment and Block-Level Area Case Study
We perform cell- and design-level routability assessments to evaluate generated standard

cells. Cell-level assessment is used to assess inherent routability for each cell in a standard-cell
library. Design-level assessment, i.e., at the level of an entire place-and-route block, is used to
evaluate a set of standard cells in the same library. We also perform case study of cell-level and
139
Figure 4.19. Results for cell-level routability assessments. (a) #DRC vs. Kth plot for NAND2 X1
and NAND3 X1. (b) #DRC vs. Kth plot for OAI21 X1 and OAI22 X1. (c) Kth values for 15 cell
types per four standard-cell libraries.
achievable block-level area. Cell height of a standard-cell library is linearly related to cell-level
area. However, cell-level area benefit does not always result in block-level area benefit, due to
routability effects. Our case study shows how we can evaluate technology options with respect
to block-level area density using the PROBE2.0 framework.
Figure 4.19 shows results for the cell-level routability assessments. We study 15 standard
cells with size X1, in each of four libraries (Lib{1,2,3,4}). Results in Figure 4.19(c) show that
standard cells with M1 have better routability than those with BPR, and standard cells with five
tracks have better routability than those with four tracks. Further, standard cells with larger
numbers of input pins have worse routability than those with smaller numbers of input pins.
Figures 4.19(a) and (b) show #DRCs vs. K plots for NAND and OAI gates, respectively.
Figure 4.20 shows results for design-level routability assessments with four designs, AES,
LDPC, JPEG and VGA. Figure 4.20(b) shows Kth metric per standard-cell library per design
while Figure 4.20(a) shows #DRCs vs. K plots for the AES design. Figure 4.21 shows results
140
Figure 4.20. Design-level routability assessment for the nine standard-cell libraries. (a) #DRC
vs. Kth plot for AES design. (b) Kth values for AES, JPEG and VGA designs. We use 0.6 Util
for AES, JPEG and VGA and 0.15 Util for LDPC design.
Figure 4.21. Achievable density and block-area case study. (a) Plot for Kth vs. achievable
utilization for AES and LDPC designs. (b) Achievable block-area study with the nine libraries
and AES design. The arrows show block-level area and cell height differences in the case study,
as also shown in Table 4.17.
of our achievable utilization study for AES and LDPC designs. We observe strong positive
correlation between Kth and achievable utilization.
In addition, we perform case study for block-level area benefits across four main axes of
technology options.
• Case 1: Available M2 routing tracks (RT) (Lib1 vs. Lib3).
141
Table 4.17. Results for achievable block-level area case study, expressed as overhead of Lib B
relative to Lib A, i.e., (AreaB − AreaA )/AreaA × 100%
Lib A Lib B Cell-level Block-level

Case
Name CH Name CH area overhead area overhead
Case 1 Lib1 100nm Lib3 120nm 20% -16%
Case 2 Lib1 100nm Lib2 120nm 20% -3%
Case 3 Lib1 100nm Lib5 130nm 30% 45%
Case 4 Lib3 120nm Lib9 120nm 0% 46%
• Case 2: Use of buried power rails (Lib1 vs. Lib2).
• Case 3: Gear ratios for MP to CPP (Lib1 vs. Lib5).
• Case 4: Different RT with the same CH (Lib3 vs. Lib9).
In Case 1, Lib3 has 20% larger cell area than Lib1. However, based on the achievable
utilization (0.71 and 0.92 for Lib1 and Lib3, respectively), the achievable block-level area with
Lib1 is 16% larger than with Lib3. It is important to note that in this case, reduction of RT brings
less cell-level area, but the block-level area actually increases. In Case 2, Lib2 has M1 PGpin
while Lib1 has BPR, so that Lib2 also has 20% larger cell area than Lib1. Applying the same
calculation as with Case 1, we obtain that the achievable area with Lib2 is 3% less than with Lib1.
Thus, in Case 2, the cell-level area benefit from use of BPR is not reflected as a design-level
area benefit. In Case 3, Lib1 has a 1:2 gear ratio for MP to CPP and Lib5 has a 2:3 gear ratio.
Since Lib5 has 26nm MP, Lib5 has 30% more cell area than Lib1. But, block-level area of Lib5
is 45% larger than that of Lib1 since the respective Lib1 and Lib5 achievable utilizations are
0.71 and 0.62. Last, in Case 4, Lib9 has the same CH (120nm) as Lib3, but MP of Lib9 is 24nm
instead of 20nm. Larger MP brings benefits of reduced resistance, but Lib3 incurs 46% area
penalty at design-level. Table 4.17 summarizes our case study. As shown in Table 4.17 and
Figure 4.21(b), block-level area no longer changes linearly with CH at advanced nodes. Even
when CH increases, block-level area can decrease.
142
Figure 4.22. Comparison between golden Kth and predicted Kth . (a) Results for 128 training data
sampled by LHS. (b) Results for 320 testing data. (c) Error distribution on the testing dataset
with kernel density estimation (KDE) plot.
4.2.3 Machine Learning (ML)-Assisted Routability Assessment
PROBE2.0 [24] uses ML-based Kth prediction to reduce the number of place-and-route
(P&R) implementation runs needed for accurate routability assessment. However, large training
sets (e.g., 80% of libraries) are used. In this work, we apply Latin hypercube sampling (LHS) [95]
to select an asymptotically more efficient training set for ML-based Kth prediction. In a case
study, we create 448 standard-cell libraries as follows: (i) CPP = 39, 42, 45, 48, 51, 54, 57nm;
(ii) MP = 16, 18, 20, 22, 24, 26, 28, 30nm for M1 and M2; (iii) 4 and 5 RT; (iv) BPR and M1
PGpin; and (v) M3 pitch:CPP ratios of 1:2 and 2:3. We choose training sets of 64, 128 and 256
data points (i.e., standard-cell library and Kth ) by LHS and use all remaining (resp. 384, 320,
192) data points for model testing. Figure 4.22 shows comparisons for golden Kth values (which
are extracted from P&R experiments) and predicted Kth values from our ML-based prediction.
Figures 4.22(a) and (b) show the golden and predicted Kth comparisons when 128 and 320 data
points are used for training and testing, respectively. Figure 4.22(c) shows the error distribution
for the testing dataset. For the three LHS-based training set sizes (64, 128, 256) we obtain 3.07,
3.00, 2.90 average Kth prediction error in testing.
143
4.2.4 Conclusion
A machine learning (ML)-assisted design-technology co-optimization (DTCO) frame-

work [24] is applied to study sub-3nm node cell and block-level DTCO. We perform cell- and
design (block)-level routability assessments for >400 unique standard-cell architectures by
varying technology options along axes of CPP, MP, RT, and use/non-use of BPR. The regime of
CH <120nm with four available tracks shows increasing block-level area due to routing difficulty,
indicating diminishing return on efforts to push ground rules. BPR improves block-level scaling
with CH <120nm, but scaling benefits slow down with CH <100nm. A 1:2 gear ratio (MP:CPP)
improves block-level area compared to a 2:3 ratio. The proposed DTCO method will extend to
evaluate power and performance along with the cell- and block-level area density assessment
presented in this study.
4.3 PROBE3.0: A Systematic Framework for Design-

Technology Pathfinding with Improved Design Enable-
ment
With the slowdown of Moore’s-Law dimensional scaling, the concept of DTCO has
emerged as a crucial approach at advanced technology nodes. DTCO has become the primary
methodology for determining which technology and design configurations are suitable for
mass production. As noted in Section 4.1, scaling boosters such as backside power delivery
networks (BSPDN) and buried power rails (BPR) have gained importance in optimizing power,
performance, area and cost (PPAC) in future technologies. Recognizing the growing significance
of scaling boosters, Section 4.1 introduces a systematic framework for routability assessments,
providing an exploration methodology for area and cost (AC) considerations. However, there is
a substantial increase in the demand for comprehensive PPAC explorations, encompassing all
aspects of PPAC.
Contributions of Our Work. Compared to the previous works PROBE1.0 [64] and PROBE2.0
144
[24], our new framework provides three main technical achievements.
(1) We establish the first comprehensive end-to-end design and technology pathfinding
framework. [24] [64] focus on area and cost without considering power and performance.
Thus, there is a significant discrepancy between [24] [64] and the actual DTCO process in the
industry. In this work, we propose a more complete and systematic PROBE3.0 framework,
which incorporates power and performance aspects for design-technology pathfinding at an early
stage of technology development. PROBE3.0 enables fast and accurate PPAC evaluations by
generating configurable PDKs, including standard-cell libraries.
(2) We improve our designs for PPAC explorations. Design is a critical factor for PPAC
explorations, and artificially generated designs enable us to explore a wider solution space.
We leverage [78] to generate artificial designs. To create more realistic artificial designs, we
develop a machine learning (ML)-based parameter tuning flow built on [78] to find the best input
parameters for generating such designs. Section 4.3.4 details our artificial design generation flow.
Further, cell width-regularization is employed in [24] [64] to prevent illegal placements when
swapping neighboring cells to assess the routability metric, Kth . We propose a clustering-based
cell width-regularization to achieve more realistic utilization (and faster routability assessment)
as described in Section 4.3.5.
(3) We demonstrate the PPAC exploration of scaling boosters. We incorporate scaling
boosters (BSPDN and BPR) to support P&R and IR drop analysis flows within the framework,
as detailed in Section 4.3.3. Our results show that incorporating BSPDN and BPR leads to a
reduction in power consumption by up to 8% and area by up to 24% based on our predictive 3nm
technology. The area reduction results are consistent with those reported in previous industry
works [55] [104] [138] [141], which have demonstrated area reductions of 25% to 30% through
the use of BSPDN and BPR techniques.
Due to limited access to advanced technology for academic research, we build our
predictive 3nm technology, named the PROBE3.0 technology. To calibrate the technology, we
refer to the International Roadmap for Devices and Structures (IRDS) [156], open-sourced PDKs,
145
Figure 4.23. Scope of PROBE. PROBE1.0 [64] and PROBE2.0 [24] (Section 4.1 of this thesis)
address AC given BEOL and FEOL/BEOL, respectively. PROBE-3nm [27] (Section 4.2 of this
thesis) studies routability with sub-3nm configurations. PROBE3.0 provides true full-stack PPAC
pathfinding.
and other publications [8] [29] [106] [150]. We open-source our work, including process design
kits (PDKs), standard-cell libraries, and scripts for P&R and IR drop analysis; this is available in
our GitHub repository [177]. In Section 4.3.2, we provide details on the automated PDKs and
library generation flows, while in Section 4.3.6, we present three experiments to demonstrate the
effectiveness of the PROBE3.0 framework for PPAC pathfinding.
4.3.1 Related Work
In this section, we divide the relevant previous works into the three categories of (i)
advanced-technology research PDKs, (ii) design-technology co-optimization and (iii) scaling
boosters, along with (iv) “PROBE” frameworks.
Advanced Technology Research PDKs. PDKs of advanced node technologies are highly
confidential. Academic research can be blocked by limited access to relevant information. To
unblock academic research, predictive advanced-node PDKs have been published. ASAP7 [29]
is a predictive PDK for 7nm FinFET technology that includes standard cells which support
146
commercial logic synthesis and P&R. FreePDK3 [106] [150] and FreePDK15 [8] are open-
source PDKs for 3nm and 15nm technology. [77] proposes a 3nm predictive technology called
NS3K with nanosheet FETs (NSFET). The authors of [77] also create 5nm FinFET and 3nm
NSFET libraries to compare power, performance and area.
Design-Technology Co-Optimization. Previous DTCO works evaluate block-level PPAC and
optimize design and technology simultaneously. [113] proposes UTOPIA to evaluate block-level
PPAC with thermally limited performance, and to optimize device and technology parameters.
[84] proposes a fast pathfinding DTCO flow for FinFET and complementary FET (CFET). [14]
also proposes a fast and agile technology pathfinding platform with compact device models
to accelerate the DTCO process. [63] describes power delivery network pathfinding for 3-D
IC technology to study tradeoffs between IR drop and routability. [22] uses ML to predict
sensitivities to changes for DTCO.
Scaling Boosters. Scaling boosters are used in advanced nodes to maximize benefit of new
technology. BSPDN and BPR are among the most promising scaling boosters in sub-5nm
nodes. [102] carries out a CPU implementation with BSPDN and BPR in their 3nm technology,
demonstrating a reduction of up to 7X in worst IR drop. Similarly, [104] investigates BSPDN
and BPR at sub-3nm nodes and finds that they can lead to a 30% reduction in area based on
IR drop mitigation. [18] also explores the impact of BSPDN and BPR on design, concluding
that their use can lead to a 43% reduction in area with 4X less IR drop. [55] studies BSPDN
configurations with µTSVs, and observes 25% to 30% reduction in area using BSPDN and BPR.
Additionally, [112] investigates BSPDN with nTSVs and µTSVs and finds that the average IR
drop with BSPDN improves by 69% compared to traditional frontside PDN (FSPDN). Finally,
[116] conducts holistic evaluations for BSPDN and BPR, demonstrating that FSPDN with BPR
achieves a 25% lower on-chip IR drop, while BSPDN with BPR achieves an 85% lower on-chip
IR drop with iso-performance and iso-area. In contrast to these previous DTCO works, here we
propose a highly configurable framework that enables more efficient investigation of scaling
boosters in advanced nodes.
147
“PROBE” Frameworks. Prior “PROBE” [24] [64] works propose systematic frameworks for
assessing routability with different FEOL and BEOL configurations. Specifically, [64] begins
with an easily-routable placement and increases the routing difficulty by random neighbor-swaps
until the routing fails with greater than a threshold number of design rule violations (DRCs). The
normalized number of swaps at which routing failure occurs, denoted by Kth , is a metric used to
measure the inherent routability of the given parameters. On the other hand, [24] introduces an
automatic standard-cell layout generation using satisfiability modulo theory (SMT) to support
explorations of both FEOL and BEOL configurations. The authors of [24] also employ machine
learning (ML)-based Kth prediction to expedite the DTCO pathfinding process. The work of [24]
corresponds to Section 4.1 of this thesis. Additionally, [27] employs PROBE2.0 in a routability
study with sub-3nm technology configurations; this work corresponds to Section 4.2 of this
thesis.
4.3.2 Standard-Cell Library and PDK Generation
Expediting the DTCO process requires automation of the standard-cell library and PDK
generation flows. Therefore, the PROBE2.0 framework [24] introduces standard-cell layout
and PDK generation flows and utilizes them for routability assessments. In this work, we
extend the PROBE2.0 framework to include proper electrical models of standard-cell libraries
and interconnect layers for design-technology pathfinding. Additionally, we enhance the PDK
generation flow to support advanced nodes. While the PROBE2.0 framework solely focuses
on the physical layout of standard cells, the PROBE3.0 framework enables true full-stack
PPAC pathfinding through automated, configurable standard-cell and PDK generation flows for
advanced nodes. To demonstrate use of PROBE3.0 for advanced-node PPAC pathfinding, we
use a technology that incorporates cutting-edge (3nm FinFET) technology predictions based on
the works of [29] [156].
148
Figure 4.24. Automatic standard-cell library and PDK generation (Design Enablement) in
the PROBE3.0 framework. In addition to technology and design parameters in the PROBE3.0
framework, other technology-related inputs are required.
Overall Flow
Figure 4.24 describes our overall flow of standard-cell and PDK generation. Technology
and design parameters are defined as input parameters for the flow. Beyond these input parame-
ters, there are additional inputs required to generate standard-cell libraries and PDKs, as follows:
(i) SPICE model cards, (ii) Liberty template and PVT conditions, (iii) Interconnect technology
files (ICT/ITF), (iv) LVS rule deck, and (v) SPICE netlists. Given the inputs, our SMT-based
standard-cell layout generation and GDS/LEF generation are executed sequentially. Generation
of timing and power models (Liberty) requires additional steps including LVS, parasitic extrac-
tion and library characterization flow. Aside from the standard-cell library generation, we also
generate interconnect models from ICT/ITF, and P&R routing technology files from technology
and design parameters. The PDK elements that we generate feed seamlessly into commercial
logic synthesis and P&R tools. Further, to the best of our knowledge, ours is the first-ever work
that is able to disseminate all associated EDA tool scripts for research purposes.
149
Table 4.18. Layer definition in the PROBE3.0 technology.
Layer Name Description

WELL N-Well
FIN Fin.
GATE Poly (gate).
GCUT Gate cut.
ACTIVE Active area for fin definition.
NSELECT N-implant.
FEOL
PSELECT P-implant.
CA Contact (via) between LIG/LISD and M0.
LIG Gate interconnect layer.
LISD Source-drain interconnect layer.
SDT Source-drain trench (ACTIVE to LIG/LISD).
BOUNDARY Boundary layer for P&R.
M0-M13 Metal layers.
BEOL
V0-V12 Via layers.
PROBE3.0 Technology
We build our own predictive 3nm technology node, called the PROBE3.0 technology.
Based on [29], we define our FEOL and BEOL layers as described in Table 4.18. We assume that
all BEOL layers are unidirectional routing layers. Hence, we first change M1 to a unidirectional
routing layer with vertical preferred direction, since the work of [29] has a bidirectional M1
routing layer. We add an M0 layer with horizontal preferred direction below the modified M1
layer, and we add contact layers V0 and CA which respectively connect between M1 and M0,
and between gate/source-drain and M0.
Also, electrical features of technologies are critical to explore “PP” aspects. Therefore,
parasitic extractions of standard cells and BEOL metal stacks are important steps. To extract
parasitic elements, interconnect technology files are required to use commercial RC extraction,
P&R and IR drop analysis tools. In this work, we use commercial tools [147] [168] [173]
for extractions, and each tool has its own technology file format.23 Interconnect technology
23The file formats for each tool are unique. The MIPT file format is for Siemens Calibre [168] for extraction, and
is converted to an RC rule file for standard-cell layout extractions. On the other hand, the ICT and ITF file formats
are for Cadence and Synopsys extraction tools, respectively. We convert ICT to QRC techfile, and ITF to TLUPlus
file, to enable P&R tools and IR drop analysis.
150
Table 4.19. Key features of the PROBE3.0 technology.
Layer Feature Value

Fin pitch 24nm
Fin width 6nm
Gate pitch (CPP) 45nm
FEOL
Gate width 16nm
Standard-cell height 100 / 120 / 144nm
Dielectric constant 3.9
Aspect ratio (width/thickness) 1.5
Power/ground pin width (M0) 36nm
M0/M2/M3 pitch 24nm
M1 pitch 30nm
M4-M11 pitch 64nm
BEOL
M12-M13 / BM1-BM2 pitch 720nm
V0-V3 via resistance 50ohm/via
V4-V11 via resistance 5ohm/via
V12 via resistance 0.06294ohm/via
Dielectric constant 2.5-3
files include layer structures of technology and electrical parameters, such as thickness, width,
resistivity, dielectric constant and via resistance. We refer to the values of physical features
in the 3nm FinFET technology of [156], such as fin pitch, fin width, gate pitch, gate width,
metal pitch and aspect ratio. We also refer to [156] for the values for electrical parameters such
as via resistance and dielectric constant. Table 4.19 describes key features of the PROBE3.0
technology.
Improved Standard-Cell Library Generation
We generate standard-cell libraries via several steps illustrated in Figure 4.24: (i) SMT-
based standard-cell layout generation, (ii) generation of GDS and LEF files, (iii) LVS and PEX
flow, and (iv) library characterization flow.
SMT-Based Standard-Cell Layout Generation. In recent technology nodes, standard-cell
architectures use a variety of pitch values for different layers in order to optimize power, perfor-
mance, area and cost (PPAC). To accommodate this, PROBE3.0 improves the SMT-based layout
151
Table 4.20. List of 41 standard cells per generated library.
Cell list Size

Inverter (INV), Buffer (BUF) X1, X2, X4, X8
4-input NAND/NOR (NAND4/NOR4) X1, X2
2-1 AND-OR-Inverter (AOI21), 2-2 AND-OR-Inverter (AOI22) X1, X2
2-1 OR-AND-Inverter (OAI21), 2-2 OR-AND-Inverter (OAI22) X1, X2
D flip-flop (DFFHQN), D flip-flop with reset (DFFRNQ) X1
2-input MUX/XOR (MUX2/XOR2), Latch (LHQ) X1
generation used in PROBE2.0 to support non-unit gear ratios for M1 pitch (M1P) and contacted
poly pitch (CPP).
Our standard-cell layouts are generated using SPICE netlists, technology and design
parameters from [24]. However, in PROBE3.0 we change two key parameters: metal pitch (MP)
and power delivery network (PDN). Instead of using MP, we define parameters for pitch values
of each layer. Since M0, M1 and M2 layers are used for standard-cell layouts, we define M0P,
M1P and M2P as pitches of M0, M1 and M2 layers, respectively. Table 4.25 shows four layouts
of AND2 X1 cells with four parameter settings. The four standard-cell libraries (Lib1, Lib2,
Lib3 and Lib4) along with their corresponding parameter sets are used for our experiments in
Section 4.3.6. For our PPAC exploration, we generate 41 standard cells for each standard-cell
library as shown in Table 4.20.
GDS/LEF Generation and LVS/PEX Flow. While [24] only supports LEF generation for
P&R, PROBE3.0 generates standard-cell layouts in both GDS and LEF formats. The GDS files
are used to extract parasitics from standard-cell layouts and check LVS between layouts and
schematics. We use Calibre [168] to check LVS and generate extracted netlists for standard
cells with intra-cell RC parasitics. Scripts for GDS/LEF generation and LVS/PEX flows are
open-sourced in [177].
Library Characterization Flow. We perform library characterization to generate standard-cell
152
Figure 4.25. Example standard cells (AND2 X1). The cells are generated by our standard-cell
layout generation with the following parameters (Fin, RT, PGpin, CH): (a) Lib1 (2Fin, 4RT, BPR,
5T), (b) Lib2 (2Fin, 4RT, M0, 6T), (c) Lib3 (3Fin, 5RT, BPR, 6T) and (d) Lib4 (3Fin, 5RT, M0,
7T).
libraries in the Liberty format. The inputs to the flow are model cards for FinFET devices,
Liberty template including PVT conditions, and interconnect technology files. We use model
cards from [150]. For the Liberty template, we define the PVT conditions, and the capacitance
and transition time indices of (7×7) tables for electrical models (delay, output transition time,
and power). We use 5, 10, 20, 40, 80, 160, and 320ps as the transition time indices. For the
input capacitance, we obtain the input pin capacitance Cinv of an X1 inverter, then multiply this
value by predefined multipliers, 2, 4, 8, 16, 24, 32, and 64. For characterization, we use the PVT
corner (T T , 0.7V, 25◦C).
153
4.3.3 Power Delivery Network
We study PDN scaling boosters to showcase the DTCO and pathfinding capability of
PROBE3.0. There are two key challenges of traditional PDNs at advanced technologies:
• High resistance of BEOL [89]: Elevated resistance in BEOL layers exacerbates IR drop
issues, necessitating denser PDN topologies.
• Routing overheads (routability) [114]: PDN occupies routing resources that are shared
with signal and clock distribution. The routability and area density impact of PDN becomes
more severe with denser PDN at advanced nodes.
To overcome these challenges, multiple foundries have begun implementing backside

power delivery networks (BSPDN) and buried power rails (BPR) as scaling boosters in their
sub-5nm technologies. We use these scaling boosters, BSPDN and BPR, to demonstrate use
of PROBE3.0. We establish four options for PDN parameter in the PROBE3.0 framework: (i)
Frontside PDN without BPR (PFS ); (ii) Frontside PDN with BPR (PFB ); (iii) Backside PDN
without BPR (PBS ); and (iv) Backside PDN with BPR (PBB ). Figure 4.26 illustrates the four PDN
configurations in the PROBE3.0 framework.
Frontside and Backside Power Delivery Network
We have defined realistic structures for both frontside power delivery networks (FSPDN)
and backside power delivery networks (BSPDN), and enabled IR drop analysis within our
framework. Table 4.21 shows the configurations for FSPDN and BSPDN. Since BEOL layers
with smaller pitches (e.g., 24nm-pitch layer) have high resistance, we add power stripes for every
layer. While the work of [24] has multiple options for FSPDN, the PROBE3.0 framework has
only one PDN structure for FSPDN. Instead, we add other options such as PFB , PBS and PBB .
Furthermore, while the Backside option in [24] assumes no PDN at the frontside for the BSPDN
option, we add power stripes at the backside for BSPDN in the PROBE3.0 framework to enable
IR drop analysis for BSPDN.
154
Figure 4.26. Cross-section view of four PDN options in the PROBE3.0 framework: (a) Frontside
PDN (PFS ), (b) Frontside PDN with BPR (PFB ), (c) Backside PDN (PBS ) and (d) Backside PDN
with BPR (PBB ).
Figures 4.26(a) and (c) respectively show cross-section views of PFS and PBS options.
The PFS option has M0 power and ground pins for standard cells, which connect to power stripes
at the frontside of the die. The PBS option uses the same M0 power and ground pins for standard
cells but connects to power stripes at the backside of the die. For the PBS option, we employ two
backside metal layers (BM1 and BM2) and one via layer (BV1) between the backside metal
layers. The layer characteristics (width, pitch and spacing) are identical to the top two layers
(M12 and M13) of FSPDN. Additionally, the M0 pins of standard cells and BSPDN are connected
using Through-Silicon Vias (TSVs). We assume nano-TSVs with 90nm [112] width for the PBS
option, and 1:10 width-to-height aspect ratio. For the PBS option, TSV insertions necessitate
155
Table 4.21. PDN configurations for FSPDN and BSPDN. A pair of power (VDD) and ground
(VSS) stripes are placed every pitch, while maintaining the spacing between VDD and VSS.
Density denotes the percentage of routing tracks occupied by PDN.
Pitch Width Spacing Density

PDN Layer
(µm) (µm) (µm) (%)
M3 1.08 0.012 0.508 4
M4 1.152 0.032 0.544 11
FSPDN
M5-M11 5.0 1.0 1.5 20
M12-M13 4.32 1.8 0.36 100
BSPDN BM1-BM2 4.32 1.8 0.36 100
reserved spaces in front-end-of-line (FEOL) layers, including keepout margins surrounding the
TSVs. To accommodate this, we insert power tap cells prior to standard-cell placement.
Frontside and Backside PDN with Buried Power Rail
In advanced nodes, power rails on BEOL metal layers can be “buried” into FEOL levels
with shallow-trench isolation (STI). Using deep trench and creating space between devices lowers
the resistance of power rails. In addition to the resistance benefits, standard-cell height (area)
can be further reduced with deep and narrow widths of power and ground pins. Figures 4.26(b)
and (d) respectively show cross-section views of FSPDN with BPR (PFB ) and BSPDN with
BPR (PBB ) options. In the case of PFB , connections between FSPDN and BPR are made through
nano-TSVs with the same 90nm width as in the PBS option (but, with 1:7 aspect ratio). These
nano-TSVs also necessitate insertion of reserved spaces.
Power Tap Cell Insertion
Although use of BSPDN and BPR can reduce area and mitigate IR drop problems,
connecting frontside layers to BSPDN and/or BPR remains a critical challenge. To establish
“tap” connections from frontside metals to BPR, or from backside to frontside metals, space must
be reserved on device layers – e.g., [102] proposes power tap cells for the connection between
BPR to MINT (M0) layers. More frequent “taps” will mitigate IR drop problems, but occupy
156
Figure 4.27. Power tap cells for (a) PFB and (b) PBS .
more placement area. In PROBE3.0, we define two types of power tap cells for the PFB and PBS
options. Tap cells for PFB connect BPR to M1, and tap cells for PBS connect BM1 to M0. By
contrast, PFS and PBB do not require power tap cells.
Power Tap Cell Structure. Figure 4.27(a) shows a structure of power tap cells for PFB .
Double-height power tap cells for PFB have 2CPP cell width. The connection between BPR and
M0 is through a 1×2 via array, and the two M1 metals are aligned with M1 vertical routing
tracks. There are also two types of power tap cells for PFB according to starting power and
ground pins: power/ground pins on the double-height power tap cells are ordered as Power-
Ground-Power (VDD-VSS-VDD) or Ground-Power-Ground (VSS-VDD-VSS). On the other
hand, Figure 4.27(b) shows a structure of power tap cells for PBS . While power tap cells for
PFB have 2CPP width, double-height power tap cells for PBS have 6CPP width due to the
∼90nm width of nano-TSVs [112]. We also assume a 50nm keepout spacing around nano-TSVs.
Similar to power tap cells for PFB , there are two types of double-height power tap cells for PFB ,
Power-Ground-Power and Ground-Power-Ground.
Power Tap Cell Insertion Scheme. Power tap cell insertion affects routability and IR drop, and
hence affect PPAC of designs. In this work, we define five tap cell insertion pitches and two
power tap insertion schemes, as follows.
157
Figure 4.28. Four power tap cell insertion results: (a) Power tap cells for PFB (2CPP width) with
Column; (b) power tap cells for PFB with Staggered; (c) power tap cells for PBS (6CPP width)
with Column; and (d) power tap cells for PBS with Staggered.
• I pitch : 24, 32, 48, 96 and 128CPP
• Ischeme : Column and Staggered
I pitch and Ischeme denote tap cell insertion pitch and tap cell insertion scheme, respectively.
Tap cell insertion scheme Column places double-height power tap cells on every two placement
rows with the given tap cell pitch. Conversely, tap cell insertion scheme Staggered places double-
height power tap cells on every four placement rows with the given tap cell pitch. Figure 4.28
shows four power tap cell insertion results for PFB and PBS with Column and Staggered insertion
schemes.
IR Drop Analysis Flow
We develop two IR drop analysis flows for FSPDN and BSPDN. Figure 4.29(a) presents
our IR drop analysis flow for FSPDN. After P&R, we generate DEF and SPEF files for routed
designs using a commercial P&R tool to perform standalone vectorless dynamic IR drop analysis.
Additionally, an interconnect technology file (QRC techfile) is needed for RC extraction as input
for the IR drop analysis flow. In contrast, Figure 4.29(b) depicts our IR drop analysis flow for
BSPDN. After P&R, we only create a SPEF file from routed designs. We then remove all routed
158
Figure 4.29. IR drop analysis flow for (a) FSPDN and (b) BSPDN. For the IR drop flow for
BSPDN, we delete all the signal and clock routing after P&R and build power stripes for BSPDN.
signals and clocks from the P&R database and construct new power stripes for BSPDN. Since the
standalone IR drop analysis tool obtains power stripe information from a DEF file, we generate
a DEF file after creating power stripes on the backside. There are two backside metal layers,
BM1 and BM2. When creating PDN on backside metal layers, we consider M1 as BM1 and
M2 as BM2, respectively. For RC extraction with BSPDN, the QRC techfile must be scaled for
backside metals since we assume BM1 and BM2 have the same pitches as M12 and M13. Full
details are visible in open-source scripts at [177].
4.3.4 Enhanced Artificial Designs for PPAC Exploration
The use of specific real designs in DTCO and PPAC exploration can bring risk of biases
and incorrect decisions regarding technology configurations (e.g., cell architecture or BEOL
stack). To avoid such biases, the PROBE1.0 [64] bases its routability assessment on a mesh-like
netlist topology, and PROBE2.0 [24] similarly uses a knight’s tour-based topology. However,
these artificial topologies have two main limitations as we bring “PP” aspects of PPAC into
the picture. First, they are highly regular and cannot capture a wide range of circuit types.
159
Table 4.22. Definition of topological parameters in ANG [78] [140].
Parameter Definition
Ninst (T1 ) Number of instances.
N prim (T2 ) Number of primary inputs/outputs.
Average net degree. The net degree of a net is the number of terminals of the
Davg (T3 )
net.
Average size of net bounding box. The√ placed (or routed) layout is divided into
Bavg (T4 )
a bin grid where each bin contains Ninst instances.
Average depth of timing paths. The depth of a given timing endpoint is the
Tavg (T5 ) maximum number of stages in any fanin combinational path of that endpoint.
Tavg is the average of all endpoint depths.
Ratio of the number of sequential cells to the total number of cells. Sratio equals
Sratio (T6 )
to number of sequential cells over total number of instances.
Second, they do not mimic timing and power properties of real netlists, as they target routability
assessment without regard to timing path structure.
PROBE3.0 overcomes these limitations by generating artificial but realistic netlists with
the Artificial Netlist Generator (ANG) of [78] [140], for use in PPAC studies. We use the six
topological parameters of ANG (see Table 4.22) to generate and explore circuits with various
sizes, interconnect complexity, routed wirelengths and timing. Moreover, we apply machine
learning (AutoML) to improve the match of generated artificial netlists to targeted (real) netlists.
Comparison of ANG and Real Designs
In this section, we study four real designs from OpenCores [161] and the corresponding
artificial netlists generated by ANG [78]. Each design is taken through commercial logic
synthesis and P&R tools [171] [172] in the PROBE3.0 technology, to obtain a final-routed layout.
For AES, JPEG, LDPC and VGA, we respectively use target clock periods of 0.2ns, 0.2ns, 0.6ns
and 0.2ns, and utilizations of 0.7, 0.7, 0.2 and 0.7. We then extract the six topological parameters
from the routed designs and use these parameters to generate artificial netlists with ANG.
We introduce a Score metric to quantify similarity between artificial and real netlists, as
defined in Equation (4.3).
160
Table 4.23. Topological parameters for real netlists from OpenCores [161] and artificial netlists
generated by [78]. Design names followed by ∗ indicate ANG-generated artificial netlists.
Parameters
Design Score
Ninst N prim Davg Bavg Tavg Sratio
AES 12318 394 3.28 0.55 7.98 0.04 -
JPEG 70031 47 3.09 0.21 10.36 0.07 -
LDPC 77379 4102 2.85 1.00 12.94 0.03 -
VGA 60921 185 3.71 0.42 8.25 0.28 -
AES* 10371 394 3.28 0.79 5.19 0.13 8.53
JPEG* 63185 47 3.16 0.70 6.97 0.15 12.03
LDPC* 58699 4106 3.10 0.78 6.96 0.13 14.8
VGA* 64412 188 3.32 0.26 6.39 0.25 2.8
N Titarget Tiout
Score = Πi=1 max( out , target ) (4.3)
Ti Ti
where: Titarget = Ti in target parameter set

Tiout = Ti of output parameter set
N = Number of parameters (N = 6)
In Equation (4.3), target and output parameters are elements Titarget and Tiout of the target and
output parameter sets. For each parameter, we calculate the discrepancy (ratio) between target
and output values. The Score value is the product of these ratios. Ideally, if output parameters
are exactly the same as target parameters, Score is 1. Larger values of Score indicate greater
discrepancy between ANG-generated netlists and the target netlists.
Table 4.23 shows the input parameters, extracted parameters and Score metric in our
comparison of real and artificial designs. The causes of discrepancy are complex, e.g., [78] has
steps that heuristically adjust average depths of timing paths Tavg and the ratio of sequential
cells Sratio . Also, performing P&R will change the number of instances Ninst , the average net
degree Davg , and the routing which determines Bavg . Hence, it is difficult to identify the input
161
parameterization of ANG that will yield artificial netlists whose post-route properties match
those of (target) real netlists. We use machine learning to address this challenge.
Machine Learning-Based ANG Parameter Tuning
We improve the realism of generated artificial netlists with ML-based parameter tuning
for ANG. Figure 4.30(a) shows the training flow in the parameter tuning. First, to generate
training data, we sweep the six ANG input parameters to generate 21,600 combinations of
input parameters, as described in Table 4.24. Second, we use ANG with these input parameter
combinations to generate artificial gate-level netlists. Third, we perform P&R with the (21,600)
artificial netlists and extract the output parameters. The extracted output parameters are used as
output labels for the ML model training. We use the open-source H2O AutoML package [154]
(version 3.30.0.6) to predict the output parameters; the StackedEnsemble AllModels model
consistently returns the best model. The model training is a one-time overhead which took 4
hours using an Intel Xeon Gold 6148 2.40GHz server. Executing commercial P&R required just
over 7 days in our academic lab setting, and is again a one-time overhead.24
Figure 4.30(b) shows our inference flow. First, we define ranges around the target
parameter and sweep the parameters to generate multiple combinations of input parameters as
candidates, which are shown in Table 4.24. Second, we use our trained model to predict the
output parameters from each input parameter combination. Note that although there are 12.3M
combinations as specified in the rightmost two columns of Table 4.24, this step requires less than
10 minutes on an Intel Xeon Gold 6148 2.40GHz server.25 Third, we calculate a predicted Score
per each input parameter combination, and then choose the parameter combination with lowest
predicted Score. Finally, we use ANG and the chosen parameter combination to generate an
24The average P&R runtime on our 21,600 ANG netlists is 0.4 hours on an Intel Xeon Gold 6148 2.40GHz server.
The data generation used 50 concurrently-running licenses of the P&R tool, with each job running single-threaded.
(21,600 × 0.4 / 50 / 24 ∼= 7.2 days. With multi-threaded runs, we estimate that data generation would have taken
3 to 4 days.)
2511 × 11 × 21 × 21 × 11 × 21 = 12,326,391. We apply simple filtering based on lower and upper bounds, to
avoid parameter values for which ANG does not work properly. Specifically, parameter values are restricted to be
within: 0 < Bavg ≤ 1.0; 0 < Sratio ≤ 1.0; 1 < Davg < 2.6; and 3 < Tavg . For example, the AES testcase then has
∼3M input parameter combinations, and predicting output parameters for all of these takes 441 seconds of runtime.
162
Table 4.24. Parameter sets for training and testing. We train our ML model with ANG input
parameters and post-P&R output parameters. The total number of datapoints is 4 × 6 × 6 × 5 ×
5 × 6 = 21600. Testing is performed in the ranges around given target parameters, according to
the step sizes.
Testing value
Parameter Training value
Range Step
target
Ninst (T1in ) 10000, 20000, 40000, 80000 T1 ± 500 100
N prim (T2in ) 100, 200, 500, 1000, 2000, 4000 T2target ± 5 1
target
Davg (T3in ) 1.8, 2.0, 2.2, 2.4, 2.6 T3 ± 0.2 0.02
target
Bavg (T4in ) 0.70, 0.75, 0.80, 0.85, 0.90, 0.95 T4 ± 0.2 0.02
target
Tavg (T5in ) 6, 8, 10, 12, 14, 16 T5 ± 10 2
target
Sratio (T6in ) 0.2, 0.4, 0.6, 0.8, 1.0 T6 ± 0.2 0.02
artificial netlist for P&R and PPAC explorations.

Table 4.25 shows the benefit from ML-based ANG parameter tuning. Columns 2-5 show
parameters from real netlists, which we use as target parameters. The trained ML model and the
inference flow produce the tuned parameters for ANG shown in Columns 6-9 of the table, and
corresponding results are shown in Columns 10-13. The average Score decreases from to 4.89
from the original value of 8.87 for ANG without ML-based parameter tuning (Table 4.23).
The ML-enabled improvement of realism in ANG netlists can be seen using t-SNE
visualization [91] from P&R results. We perform P&R for the four real designs by sweeping
initial utilization from 0.6 to 0.8 with a 0.01 step size, and target clock period from 0.15 to
0.25ns with a 0.01ns step size; this results in 21 × 11 = 231 P&R runs. (For LDPC, we sweep
utilization from 0.1 to 0.3 with a 0.01 step size, and clock period from 0.55 to 0.65ns with a
0.01ns step size.) We then perform P&R for artificial netlists with and without our parameter
tuning flow, with 0.7 utilization (0.2 for LDPC) and 0.2ns (0.6ns LDPC) target clock period.
Figure 4.31 shows t-SNE visualization26 of the real and artificial designs. The 231 real datapoints
per design form well-defined clusters. In Figure 4.31(a), the datapoints of the artificial AES
26 For t-SNE visualization, we collect ten features from P&R results: Number of instances, number of nets,
number of primary input/output pins, average fanout, number of sequential cells, wirelength, area, number of design
rule violations, worst negative slack, total negative slack, and number of failing endpoints.
163
Table 4.25. Topological parameters for target, input and output netlists. The design names followed by ∗∗ indicate ANG-generated
artificial netlists with ML-based ANG parameter tuning.
Parameters of target netlists ANG input parameters Parameters from artificial netlists
Parameter
AES JPEG LDPC VGA AES JPEG LDPC VGA AES** JPEG** LDPC** VGA**
Ninst 12318 70031 77379 60921 12718 69531 76979 60421 10200 64296 64796 65113
164
N prim 394 47 4102 185 390 42 4106 199 394 46 4110 202
Davg 3.28 3.09 2.85 3.71 3.40 3.10 3.03 3.53 3.26 3.13 3.18 3.30
Bavg 0.55 0.21 1.00 0.42 0.49 0.31 1.98 0.28 0.72 0.21 0.73 0.36
Tavg 7.98 10.36 12.94 8.25 13.98 18.36 20.94 12.25 8.01 9.29 11.64 8.54
Sratio 0.04 0.07 0.03 0.28 0.01 0.27 0.01 0.16 0.11 0.20 0.13 0.16
Score - - - - - - - - 4.39 3.59 2.77 8.81
Figure 4.30. ML-based parameter tuning for ANG.
and JPEG designs are located in the corresponding designs’ clusters. However, the artificial
LDPC and VGA designs are not close to the corresponding clusters of real designs. By contrast,
Figure 4.31(b) shows that with our ML-based ANG parameter tuning, datapoints of all four
artificial designs are located within the corresponding clusters of real designs. This suggests that
the ML-based ANG parameter tuning helps create artificial netlists that better match targeted
design parameters – including parameters that are relevant to PPAC exploration.
4.3.5 Cell Width-Regularized Placements for More Realistic Routability

Assessment
Recall that in the PROBE approach, routability (“AC”) is evaluated using the K-threshold
(Kth ) metric [64]. That is, given a placed netlist, routing difficulty is gradually increased by
iteratively swapping random pairs of neighboring instances. The cell-swaps progressively “tangle”
the placement until it becomes unroutable (> 500 DRCs post-detailed routing). The number of
swaps K – expressed as a multiple of the instance count – at which routing fails is the Kth metric.
Larger Kth implies greater routing capacity or intrinsic routability.
165
Figure 4.31. Comparison between real and artificial designs by t-SNE [91]. (a) t-SNE visualiza-
tion for real and artificial (ANG, design names followed by ∗) designs without our parameter
tuning flow. (b) Real and artificial (ANG) designs with our ML-based parameter tuning flow.
Both PROBE1.0 [64] and PROBE2.0 [24] enable the study of real netlists through the
concept of a cell width-regularized placement. In this approach, combinational cells are inflated
(by LEF modification) to match the maximum width among all the combinational cells in the
cell library. This process, called cell width-regularization, prevents illegal placements (i.e.,
cell overlaps due to varying widths) from arising due to neighbor-swaps during Kth evaluation.
Unfortunately, while cell width-regularization permits real designs to be placed and then tangled
by random neighbor-swaps, it also forces low utilizations that harm the realism of the study.
(Moreover, high whitespace leads to high Kth values that require more P&R runs to determine.)
We now describe a clustering-based cell width-regularization methodology that generates
placements with realistic utilizations, based on real designs. Our experiments in Section 4.3.6
show that clustering-based cell width-regularization obtains the same Kth rank-ordering of design
enablement, with less P&R expense, than the previous cell width-regularization approach.
166
Figure 4.32. Two example clustered cells NAND X1 AND X1 and INV X1 OR X1, in
clustering-based cell width-regularization. (a) Schematic view, and (b) physical layout view
assuming Lib2. Here, the maximum clustered cell width wmax is 12CPP.
Clustering-Based Cell Width-Regularization
We propose clustering-based cell width-regularization using bottom-up hypergraph

clustering, as detailed in Algorithm 7. In the following, we refer to standard cells of the original
netlist as cellsorig , and (clustered) cells of the clustered netlist as cellsclustered .
Clustered Hypergraph Creation. For a given design, we first obtain a netlist hypergraph using
OpenDB [162]. We perform cell width-regularized clustering, where cells cellsorig (vertices) in
the original netlist hypergraph are clustered such that clustered cell width27 does not exceed wmax ,
the maximum cell width in the library. The inputs to cell width-regularized clustering are (i) a
hypergraph H(V, E,W ) with vertices V , hyperedges E and cell widths W , (ii) the maximum cell
width, wmax , and (iii) a limit on number of clustering iterations, Niter .28 The output is a clustered
hypergraph (Hout ). We use First-Choice (FC) clustering [74] and refer to our clustering method
as cell width-regularized clustering with FC, or CWR-FC.
27 Given vertex set V with cell widths W , clustering vertices vi , v j ∈ V yields a clustered cell with width W [vi ] +
W [v j ].
28 In our experiments, we set N
iter = 20. However, the cell width-regularized clustering is strongly constrained by
wmax , and we observe on our testcases that clustering stops after ∼ 3 iterations.
167
Algorithm 7. Cell width-regularization by clustering.
Inputs: Hypergraph H(V, E,W ), Maximum cell width wmax , Number of iterations Niter
Outputs:, Clustered hypergraph Hout (Vout , Eout ,Wout )
1: Ncluster ← |V |
2: Hypergraph at iteration 0, H0 (V0 , E0 ,W0 ) ← H(V, E,W )
3: for k ← 0; k < Niter ; k + + do
4: Vordered ← Sorted Vk in increasing order of Wk
5: visited[v] ← f alse ∀v ∈ Vk
6: Cluster assignments, cmap[v] ← v ∀v ∈ Vk
7: Clustered cell widths, Wk+1 ← Wk
8: for vi ∈ Vordered do
9: if visited[vi ] == true or vi is a sequential cell then
10: continue
11: end if
12: Vneighbor ← Find adjacent vertices of vi
13: Best cluster score, φbest ← 0; Best cluster candidate, vbest ← −1
14: for v j in Vneighbor do
15: if Wk [vi ] +Wk+1 [cmap[v j ]] ≤ wmax then
weighte
∑v ∈e,v ∈e
i j |e|−1
16: φ (vi , v j ) ← W [vi ]+W // Cluster Score
k k+1 [cmap[v j ]]
17: if φ (vi , v j ) > φbest then vbest ← v j
18: end if
19: end if
20: end for
21: if vbest == −1 then
22: Wk+1 [vi ] ← Wk [vi ]
23: visited[vi ] ← true
24: else
25: cmap[vi ] ← cmap[vbest ]
26: Wk+1 [vbest ] ← Wk [vi ] +Wk+1 [cmap[vbest ]]
27: visited[vi ] ← true; visited[vbest ] ← true
28: Ncluster ← Ncluster − 1
29: end if
30: end for
31: if Ncluster == |Vk−1 | then
32: break
33: else
34: Hk+1 (Vk+1 , Ek+1 ,Wk+1 ) ← Build clustered hypergraph using cmap
35: end if
36: end for
37: Hc ← Clustered hypergraph generated at last iteration
38: Hout ← Best-fit bin packing on Hc
39: Return Hout
168
CWR-FC first sorts vertices in increasing order of cell widths (Line 4) and initializes
cluster assignments (Line 6). The cluster assignment cmap is the mapping of vertices to clusters
(Vk to Vk+1 ). Clustered cell widths Wk+1 are initialized in Line 7. Next, vertices are traversed
in order to perform pairwise clustering; note that only combinational cells are considered for
clustering (Line 8). For each vertex vi that is traversed, we find its neighbors v j in the hypergraph
(Line 11). Each v j is considered only if it does not violate the wmax limit (Line 14); a cluster
score φ (vi , v j ) is calculated in Line 15. In the cluster score, weighte is the weight of hyperedge
e and Wk [vi ] is the width of vertex vi . The numerator aims to cluster vertices that are strongly
connected (i.e., share many hyperedges) while the denominator promotes clusters of similar
widths. If all neighboring vertices v j violate the threshold width constraint, then no new clusters
are formed (Lines 17-19). Otherwise, the vertex with the highest cluster score is selected, and a
new cluster is created (Lines 21-24). After all vertices are visited, we construct the clustered
hypergraph and proceed with subsequent iterations (Line 28). If no further clustering is feasible,
the process terminates (Line 26).
Note that CWR-FC clusters vertices that are adjacent to each other in the hypergraph.
However, if all pairings of vertices selected for clustering violate the wmax width constraint, the
algorithm can stall (Line 25). To address this issue and improve the uniformity of cluster contents,
we perform best-fit bin-packing [60] with bins having capacity wmax (Line 30).29 Finally, the
output is the clustered hypergraph Hout .
Clustered Netlist Creation. We convert the clustered hypergraph Hout into Verilog using
OpenDB. Then, to run P&R we require a new LEF file that captures the cluster assignments from
cell width-regularized clustering. I.e., we require a new netlist over the clusters, cellsclustered .
Figure 4.32(a) provides a schematic view of two clustered cells, NAND X1 AND X1 and
INV X1 OR X1. These correspond to two clusters of original cells: NAND X1 and AND X1,
and INV X1 and OR X1. Figure 4.32(b) shows how the clustered cells are composed from
29The choice of best-fit is motivated by its simplicity and intuitiveness. Best-fit also enjoys a better approximation
ratio compared to first-fit or next-fit alternatives [60].
169
Figure 4.33. Cell width distributions pre-clustering (i.e., original netlist) and post-clustering
(i.e., by CWR-FC) for (a) AES, (b) JPEG, (c) LDPC and (d) VGA.
original standard-cell layouts. In this case, a non-integer gear ratio between M1P (30nm)
and CPP (45nm) forces cells in cellsclustered to be positioned at even CPP sites, to avoid M1
pin misalignment. In the first cluster, NAND X1 width (3CPP) is an odd number of CPPs,
necessitating addition of 1CPP padding between the two cells. In the second cluster, the total
cell width is less than wmax , so whitespace is included along with the clustered original cells.
We distribute whitespace uniformly, (i) at the sides of cellsclustered and (ii) between consecutive
cells in each cluster, as illustrated in Figure 4.32(b). During this whitespace allocation, we first
allocate whitespace at junctions (between consecutive original cells) where no extra padding has
been previously allocated.
170
Performance of Clustered Cell Width-Regularization
We now document advantages of our proposed clustered cell width-regularization, i.e.,

more realistic utilization in P&R blocks, and realistic topological and wirelength characteristics
of P&R outcomes.
Comparison to Previous Cell Width-Regularization. Figure 4.33 compares cell width distri-
butions for instances in the clustered netlist and instances in the original netlist. The blue lines
show the distribution of cell widths in the original netlist, where smaller cell widths predominate.
The red lines indicate that CWR-FC increases the prevalence of cells with larger widths through
creation of the merged cellsclustered . The larger amount of actual cell widths in cellsclustered leads
to smaller amounts of added whitespace needed to regularize cell widths.
As anticipated, clustered cell width-regularization significantly reduces whitespace in the
placed designs. With Lib2 and FSPDN for P&R, placing cell width-regularized instances used in
PROBE2.0 at 90% density achieves actual utilizations of 0.21, 0.21 and 0.40 for AES, JPEG and
VGA, respectively. For LDPC, placing cell width-regularized instances at 30% density achieves
actual utilization of 0.08. By contrast, with clustered cell width-regularization, we achieve actual
utilizations of 0.71, 0.74 and 0.71 for AES, JPEG and VGA, respectively. For LDPC, we achieve
actual utilization of 0.23. In this way, our new methodology enables Kth evaluation by iterated
neighbor-swapping while preserving realistic placement utilizations.
Topological and Wirelength Comparisons to Real Designs. We have confirmed additional
similarities between between clustered cell width-regularized netlists and the original real designs.
Table 4.26 compares characteristics of our clustering-based cell width-regularized netlists and
placements ([C]), versus analogous characteristics of real netlists and placements ([A]). We
also implement another plausible clustering methodology, which is to induce clusters from a
placement of the original design ([B]). In [B], clusters from the placement are induced by (i)
traversing combinational cells left-to-right in each standard cell row, and (ii) clustering maximal
contiguous sets of cells without exceeding Wmax .
171
Table 4.26. Comparison of the width-regularized clustered netlist produced by CWR-FC ([C])
with the original flat netlist ([A]) and a width-regularized clustered netlist induced from a
placement of the flat netlist ([B]).
Stage Design #Insts Area (µm2 ) Util WL (µm) Avg. FO

AES 12318 426.254 0.83 30849 2.32
JPEG 70031 2781.981 0.73 112605 2.15
[A]
LDPC 77379 6250.563 0.43 567630 1.85
VGA 60921 4238.205 0.76 208845 2.71
AES 4275 426.254 0.83 32632 1.96
JPEG 23281 2781.981 0.73 111241 1.86
[B]
LDPC 42383 6250.563 0.43 585923 1.43
VGA 40084 4238.205 0.76 189612 2.14
AES 4661 426.254 0.83 32679 2.08
JPEG 25961 2781.981 0.73 143693 1.76
[C]
LDPC 30636 6250.563 0.43 637417 1.29
VGA 32768 4238.205 0.76 220915 2.02
We run P&R using Lib2 and PFS for PDN, maintaining the same core area and utilization.
Clustering decreases the number of instances and average fanouts for [B] and [C], relative to
[A]. However, wirelengths exhibit no significant changes. The similarities between [A], [B]
and [C] suggest that our CWR-FC methodology can preserve netlist properties relevant to P&R
outcomes, with more realistic placement utilizations.
We extensively study the design-technology pathfinding capability of the PROBE3.0

framework using the PROBE3.0 technology. In this section, we report three main experiments.
Expts. 1 and 2 show PROBE3.0’s capability to assess PPAC trends and tradeoffs, using real
and artificial designs respectively. Expt. 3 performs assessments of routability and achievable
utilization.
In Expts. 1 and 2, we analyze four tradeoffs. (i) We present Performance-Power plots
that quantify tradeoffs between performance (maximum frequency) and power. (ii) We present
Performance-Area plots to quantify the tradeoffs between performance and area. (iii) To address
172
PP aspects, we use the Energy-Delay Product (EDP) [84] as a single metric for power and
performance. EDP-Area plots depict tradeoffs between performance/power and area. (iv)
We present IR drop-Area plots to demonstrate tradeoffs between IR drop and area. We also
compare results obtained using artificial designs with those obtained using real designs. Expt. 3
assesses routability and achievable utilization using our clustering-based cell width-regularized
placements.
Experimental Setup
Based on the definition of technology and design parameters in [24], we define ten
technology parameters and eight design parameters as the input parameters for the PROBE3.0
framework. Table 4.27 describes the definitions of these parameters and the options used in our
experiments. Also, we use commercial tools for PDK generation, logic synthesis, P&R, and
IR drop analysis. We use open-source tools for GDT-to-GDS translation [151] and SMT solver
[178]. Table 4.28 summarizes the tools and versions that we use in our experiments.
Criteria for Valid Result. In our experiments, for given Design, PDN and technology parameters,
we perform logic synthesis, P&R and IR drop analysis with multiple sets of parameters including
I pitch , Ischeme , Util and Clkp. We use 24, 32, 48, 96 and 128CPP for I pitch , and Column and
Staggered for Ischeme . For Util, we use values ranging from 0.70 to 0.94 with a step size of 0.02,
and for Clkp, we use values ranging from 0.12 to 0.24ns with a step size of 0.02ns. Importantly,
after the implementation and the analysis steps, we filter out results that are deemed invalid – in
that they are likely to fail signoff criteria even with additional human engineering efforts.
To be precise, a “valid” result must satisfy three conditions: (i) the worst negative slack
is larger than -50ps; (ii) the number of post-route DRCs is less than 500; and (iii) the 99.7
percentile of the effective instance voltage is greater than 80% of the operating voltage (Vop ). To
assess (iii), we use a commercial IR drop analysis tool [148] to measure vectorless dynamic IR
drop, and calculate the effective instance voltage as Vop −Vdrop per each instance, where Vop is
an operating voltage (0.7V) and Vdrop is the worst voltage drop per instance. We take the 99.7
173
Table 4.27. Technology and design parameters in our experiments.
Type Parameter Description Option

Fin The number of fins for devices of standard cells. 2, 3
CPP Contacted poly pitch for standard cells in nm. 45
M0P M0 (horizontal) layer pitch in nm. 24
M1P M1 (vertical) layer pitch in nm. 30
M2P M2 (horizontal) layer pitch in nm. 24
RT The number of available M0 routing tracks in standard cells. 4, 5
PGpin Power/ground pin layer for standard cells. BPR, M0
Technology Cell height of standard cells, expressed as a multiple of M0P.
When the cell height is 120nm and M0P is 24nm, the cell
CH 5, 6, 7
height (CH) is 5. The cell height value is calculated as RT +
2 for M0 PGpin and RT + 1 for BPR PGpin.
MPO The number of minimum pin openings (access points). 2
Design rules. We define the same grid-based design rules,
minimum area rule (DR-MAR), end-of-line spacing rule (DR-
174
DR EOL) and via spacing rule (DR-VR) as [24]. We use the EUV-Tight
EUV-tight (ET) design rule set, which includes DR-MAR =
1, DR-EOL = 2 and DR-VR = 1.
Metal stack options. The 14M metal option which contains
BEOL 14 metal layers (M0 to M13). We define 1.2X, 2.6X, 3.2X 14M
and 30X layer pitches based on 24nm as the 1X pitch.
PDN Power delivery network options. PFS , PFB , PBS , PBB
I pitch Power tap cell pitch in CPP. 24, 32, 48, 96, 128
Design Ischeme Power tap cell insertion scheme. Column, Staggered
Tool Commercial P&R tools. Synopsys IC Compiler II
Util Initial placement utilization. 0.70 to 0.94 with 0.02 step size
Designs studied in our experiments: four open-source de-
Design signs from OpenCores [161] and artificial netlists generated AES, JPEG, LDPC, VGA
by ANG with our ML-based parameter tuning.
Target clock period for logic synthesis and P&R. Clkp values
Clkp 0.12 to 0.24ns with 0.02ns step size
reflect maximum achievable frequencies of the designs.
Table 4.28. Tools and versions in our experiments.
Purpose Tool Version Ref.

Format conversion GDT-to-GDS translator 4.0.4 [151]
IR drop analysis Cadence Voltus 19.1 [148]
Library characterization Cadence Liberate 16.1 [145]
Cadence Genus 21.1 [143]
Logic synthesis
Synopsys Design Compiler R-2020.09 [171]
LVS Siemens Calibre 2017.4 19 [168]
P&R Synopsys IC Compiler-II R-2020.09 [172]
Cadence QRC Extraction 19.1 [147]
PEX Synopsys StarRC O-2018.06 [173]
Siemens Calibre 2017.4 19 [168]
SMT solver Z3 4.8.5 [34] [178]
percentile of effective instance voltage as representative of IR drop for the post-P&R result, as it
is within three standard deviations from the mean per the empirical rule [137].
Expt. 1: PPAC Exploration with Real Designs
Performance versus Power. We first present PPAC explorations that show tradeoffs between
performance and power. (We assume that area is proportional to cost, since chip area is closely
related to cost.) In this study, we show results for JPEG with four standard-cell libraries (Lib1-4).
Also, we use four PDN structures, PFS , PFB , PBS and PBB , and measure improvements due to
scaling boosters relative to the traditional frontside PDN (PFS ).
Figure 4.34(a) gives Performance-Power plots that show tradeoffs between performance
and power for JPEG, and improvements from the traditional FSPDN. We calculate the maximum
achievable frequency ( fmax ) as 1/(Clkp −W NS) where Clkp is the target clock period and W NS
is the worst negative slack. Also, we add up leakage and dynamic power to obtain the total
power. To measure the improvement from PFS , we compare the second-largest value (on the
x-axis) attained with each PDN configuration. From the result, we make two main observations.
(i) Power consumption with PBS and PBB decreases by 7 to 8%, compared to PFS with the same
performance. (ii) Power consumption with PFB is similar to PFS , with the same performance. We
175
observe power reductions from use of scaling boosters, BSPDN and BPR. However, use of BPR
without BSPDN does not reduce power consumption.
Performance versus Area. Figure 4.34(b) shows Performance-Area tradeoffs for JPEG. We
make two main observations. (i) Area with PFB , PBS and PBB decreases by up to 8%, 5% and
24%, respectively, as compared to PFS , while maintaining the same level of performance. (ii) We
find that use of scaling boosters results in area reductions across all four standard-cell libraries.
The area reduction results obtained using the PROBE3.0 framework are consistent with previous
industry works [55] [104] [138] [141], which show that use of BSPDN and BPR techniques can
result in area reductions of 25% to 30%.
Energy-Delay Product (EDP) versus Area. Given the tradeoffs among PPAC criteria, a simpler
metric is useful to comprehend multiple aspects simultaneously. The Energy-Delay Product
(EDP) is adopted by, e.g., [84] as a single-value metric that captures both power efficiency and
maximum achievable frequency (performance). EDP is calculated as P × fmax 2 , where P denotes
power consumption and fmax denotes maximum achievable frequency. Lower EDP means more
energy-efficient operations for the chip. Since we address power, performance and area (cost),
we draw EDP-Area plots to show PPAC tradeoffs of various PDN structures. We again use four
standard-cell libraries (Lib1-4).
From Figure 4.34(c), we derive four key observations. (i) For 4RT (Lib1 and Lib2), EDP
with PFB , PBS and PBB decreases by 0.2, 0.2 and 0.4 mW · ns2 , respectively, compared to PFS with
the same area. (ii) For 5RT (Lib3 and Lib4), EDP with PBB decreases by 0.3 mW · ns2 , compared
to PFS with the same area. (iii) For 5RT, EDP with PFB shows no improvements, and EDP with
PBS increases by 0.1 mW · ns2 , as compared to PFS with the same area. (iv) Use of PBB better
optimizes area than other PDN structures with the same EDP.
Supply Voltage (IR) Drop versus Area. With recent advanced technologies and designs, denser
PDN structures are required due to large resistance seen in tight-pitch BEOL metal layers. The
denser PDN structures bring added routability challenges which critically impact area density.
In light of this, we measure IR drop and area from valid runs, and plot IR drop-Area tradeoffs
176
177
Figure 4.34. PPAC tradeoffs for JPEG with four standard-cell libraries (Lib1-4) and four PDN structures. We measure improvements
relative to traditional PDN (PFS ), to show the benefits of BSPDN and BPR: (a) Performance-Power, (b) Performance-Area and (c)
Energy-Delay Product-Area.
Figure 4.35. IR drop-Area plots for JPEG with four standard-cell libraries (Lib1, Lib2, Lib3 and
Lib4). (a) JPEG with 4RT (Lib1/2). (b) JPEG with 5RT (Lib3/4).
in Figure 4.35. In the plots, we compare the points with the minimum area for each PDN
configuration in terms of area and 99.7 percentile (three-sigma) of effective instance voltage
(EIV). Note that larger effective instance voltage means better IR drop mitigation. Figures 4.35(a)
and (b) show IR drop-Area tradeoffs for JPEG with 4RT (Lib1 and Lib2) and 5RT (Lib3 and Lib4),
respectively. From the results, we make four main observations. (i) Area with PFB decreases by 2
to 6% compared to PFS , while the effective instance voltage (EIV) increases by 3 to 4%. (ii) Area
with PBS increases by 1 to 4% compared to PFS , while EIV decreases by 4 to 12%. (iii) Area with
PBB decreases by 15 to 18% compared to PFS , while EIV decreases by 17%. (iv) We observe that
there is IR drop mitigation from use of backside PDN, while use of BPR (PFB ) worsens IR drop.
This implies that more power tap cells will need to be inserted to mitigate IR drop. However, the
area overhead of power tap cells will degrade the IR drop quality achieved by use of BPR.
Expt. 2: PPAC Exploration with Artificial Design
Our second main experiment uses the artificial JPEG design generated by ANG using
our ML-based parameter tuning. We conduct the same studies as in Expt. 1 and analyze the
results.
Performance versus Power. Figure 4.36(a) shows the tradeoffs between performance and
power with the artificial JPEG design. From the result, we make three main observations. (i)
178
Power consumption with PBS and PBB decreases by 6 to 14%, compared to PFS with the same
performance. (ii) Power consumption with PFB is similar to PFS with the same performance. (iii)
Results with the artificial JPEG show up to 7% differences, but with similar trends, compared to
the results obtained with the real JPEG design.
Performance versus Area. Figure 4.36(b) shows tradeoffs between performance and area
with the artificial JPEG design. We make three main observations. (i) Area with PFB and PBB
decreases up to 14% and 21% compared to PFS with the same performance. (ii) Area with PBS
increases by 0% to 3% compared to PFS with the same performance. This area penalty is caused
by power tap cell insertion for PBS . (iii) We observe that the results with the artificial JPEG show
up to 9% differences, but with similar trends, compared to the results obtained with the real
JPEG design. However, area for PBS shows opposite trends to what we observe with the real
design, although the discrepancy is not too large.
Energy-Delay Product (EDP) versus Area. From Figure 4.36(c), we make three main observa-
tions. (i) For 4RT (Lib1 and Lib2), EDP with PBB decreases by 0.5 mW · ns2 , compared to PFS
with the same area. However, EDP with PFB and PBS shows no improvements. (ii) For 5RT (Lib3
and Lib4), EDP with PFB and PBB decreases by 0.6 and 0.9 mW · ns2 , compared to PFS with the
same area. However, EDP with PBS shows no improvements. (iii) We observe that results with
the artificial JPEG show similar trends as results obtained with the real JPEG design.
Supply Voltage (IR) Drop versus Area. Figures 4.37(a) and (b) show tradeoffs between IR
drop and area for the artificial JPEG design with 4RT (Lib1 and Lib2) and 5RT (Lib3 and
Lib4), respectively. We make four main observations. (i) Area with PFB decreases by 9 to 14%,
compared to PFS , while the effective instance voltage (EIV) increases by 1 to 6%. (ii) Area with
PBS increases by 2%, compared to PFS , while EIV decreases by 2 to 3%. (iii) Area with PBB
decreases by 14 to 18%, compared to PFS , while EIV decreases by 6 to 11%. (iv) We observe
that results with the artificial JPEG show similar trends as results obtained with the real JPEG
design, and that discrepancies are reasonably small.
179
180
Figure 4.36. PPAC tradeoffs for “artificial” JPEG with four standard-cell libraries (Lib1, Lib2, Lib3 and Lib4). (a) Performance-Power.
(b) Performance-Area. (c) Energy-Delay Product-Area.
Figure 4.37. IR drop-Area plots for “artificial” JPEG with four standard-cell libraries (Lib1,
Lib2, Lib3 and Lib4). (a) JPEG with 4RT (Lib1/2). (b) JPEG with 5RT (Lib3/4).
Expt. 3: Routability Assessment and Achievable Utilization
Our third main experiment measures Kth using our clustering-based cell width-regularized
placements, and explores the relationship between Kth and achievable utilization. We note that
the previous work of [24] introduced Achievable Utilization as the maximum utilization for
which the number of DRCs is less than a predefined threshold of 500 DRCs. Here, we include
all three criteria for a valid result, and define Achievable Utilization as the maximum utilization
among all valid runs seen.
Figure 4.38 shows experimental results for Kth and achievable utilization. We conduct
our experiment with artificial JPEG and four cell width-regularized libraries (Lib1-4). From
the plots, we make two observations. (i) We compare the results with 2Fin/4RT standard-cell
libraries (Lib1/2) to those with 3Fin/5RT standard-cell libraries (Lib3/4). The data show that
a larger number of M0 routing tracks brings better routability. (ii) Compared to PFS , PFB and
PBB , the plots for PBS are skewed to the right for each design, showing better routability than
the other PDN configurations. We observe that the routability improvement of PBS comes from
regularly-placed power tap cells: the power tap cell placement eases routing congestion caused
by high cell and/or pin density.
Finally, we compare the Kth results obtained with the previous cell width-regularized
181
Figure 4.38. Kth and achievable utilization for (a) AES and (b) JPEG, with various libraries and
power delivery methodologies.
placements used in the PROBE2.0 work ([A]) and clustering-based cell width-regularized
placements obtained using the CWR-FC algorithm ([C]). We perform routability assessments as
summarized in Table 4.29. We rank-order Kth across the eight combinations of four PDN and
two RT with the JPEG design. The main observation from this comparison is that the ordering
of enablement based on Kth is the same for both placements, even as the area utilization of the
clustering-based cell width-regularized placements is closer to the initial utilization (0.6). We
conclude that our clustering-based cell width-regularized placement methodology successfully
provides more realistic placements without disrupting the Kth -based rank-ordering of enablement.
Moreover, the generally smaller Kth values seen in the rightmost two columns of Table 4.29
imply fewer P&R trials needed to evaluate the Kth metric.
4.3.7 Conclusion
We have presented PROBE3.0, a systematic and configurable framework for “full-stack”

PPAC exploration and pathfinding in advanced technology nodes. We introduce automated PDK
and standard-cell library generation, along with enablement of scaling boosters in a predictive
3nm technology. Our work is permissively open-sourced in GitHub [177], and includes open-
182
Table 4.29. Kth comparison for the JPEG design with cell width-regularized placements ([A])
and clustering-based cell width-regularized placements ([C]). Util denotes real utilization with
0.6 initial utilization.
[A] [C]
Rank PDN RT Library
Kth Util Kth Util
1 PFB 4 Lib1 6 0.14 3 0.50
2 PFS 4 Lib2 9 0.14 5 0.49
3 PBB 4 Lib1 12 0.14 7 0.50
4 PFS 5 Lib4 15 0.14 8 0.50
5 PFB 5 Lib3 16 0.14 9 0.50
6 PBS 4 Lib2 17 0.14 13 0.49
7 PBB 5 Lib3 18 0.14 16 0.50
8 PBS 5 Lib4 23 0.14 26 0.50
sourceable PDKs and EDA tool scripts that incorporate power and performance considerations
into the framework.
We employ artificial netlist generation with a machine learning-based parameter tuning
to mimic properties of arbitrary real designs. Along with a new CWR-FC clustering-based
width-regularized netlist and placement methodology, this enables PPAC exploration of a much
wider space of technology, design enablement, and design options. From our experiments, we
find that the use of backside power delivery network (BSPDN) and buried power rails (BPR)
can lead to up to 8% reduction in power consumption and up to 24% reduction in area using
our predictive 3nm technology. These results from PROBE3.0 closely match previous works
[55] [104] [138] [141] which estimated 25% to 30% area reduction from use of BSPDN and
BPR.
Ongoing and future directions include the following. (i) Improving the software archi-
tecture of PROBE3.0 will make it more accessible and flexible for users to pursue their own
PPAC explorations. Supporting the addition of user-defined variables can help capture and
study variant technology and design assumptions. (ii) To improve robustness of the framework,
and its usefulness as a “proxy” in real-world advanced technology development and DTCO,
183
improved device models, parasitic extraction models, signoff corner definitions, relevant design
examples, etc. will be beneficial. It will also be necessary to add generation of DRC rule decks
for commercial tools. (iii) While our scripts for commercial tools are shared publicly in our
GitHub repository, using these tools still requires valid licenses. Incorporation of open-source
tools into the PROBE3.0 framework can potentially lead to highly-scaled deployments, shorter
turnaround times, and improved utility to a broader audience.
4.4 Acknowledgments
Chapter 4 contains a reprint of Chung-Kuan Cheng, Andrew B. Kahng, Hayoung Kim,

Minsoo Kim, Daeyeal Lee, Dongwon Park and Mingyu Woo, “PROBE2.0: A Systematic
Framework for Routability Assessment from Technology to Design in Advanced Nodes”, IEEE
Transactions on Computer-Aided Design of Integrated Circuits and System 41(5), 2022; Chidi
Chidambaram, Andrew B. Kahng, Minsoo Kim, Giri Nallapati, S. C. Song and Mingyu Woo,
“A Novel Framework for DTCO: Fast and Automatic Routability Assessment with Machine
Learning for Sub-3nm Technology Options”, Proc. IEEE Symposium on VLSI Technology,
2021; and Suhyeong Choi, Jinwook Jung, Andrew B. Kahng, Minsoo Kim, Chul-Hong Park,
Bodhisatta Pramanik and Dooseok Yoon, “PROBE3.0: A Systematic Framework for Design-
Technology Pathfinding with Improved Design Enablement”, in submission to IEEE Transactions
on Computer-Aided Design of Integrated Circuits and Systems, 2023. The dissertation author is
a main contributor to, and a primary author of, each of these papers.
I would like to thank my coauthors, Professor Chung-Kuan Cheng, Dr. Chidi Chi-
dambaram, Suhyeong Choi, Dr. Jinwook Jung, Professor Andrew B. Kahng, Hayoung Kim, Dr.
Daeyeal Lee, Dr. Giri Nallapati, Dr. Chul-Hong Park, Dr. Dongwon Park, Bodhisatta Pramanik,
Dr. S. C. Song, Mingyu Woo and Dooseok Yoon.
184
Chapter 5
Conclusion
This thesis presents robust physical design and design-technology co-optimization

methodologies at advanced VLSI technology. The proposed methodologies address signifi-
cant challenges posed by (i) aggressive PPAC optimization for various applications in modern IC
designs, (ii) local layout effect and high resistance of advanced nodes, and (iii) limitations of the
current DTCO process.
Chapter 2 presents two general physical design methodologies. First, we have proposed a
new k-active dynamic power minimization problem that arises in clock distribution for memory-
dominant IC designs. We formulate an ILP for the k-active dynamic power minimization problem,
and the ILP minimizes the weighted sum of the total costs (wirelengths) and the maximum power
across all sets of k active sinks. Our experimental results give new insights into the tradeoff for
maximum k-(consecutive)-active dynamic power and wirelengths of subtrees as well as skew
and total wirelengths. Second, we have described CoRe-ECO, a novel concurrent refinement
framework for the automated ECO flow. We propose simultaneous and perturbation-minimized
refinements of placement and routing to remove the remaining DRCs after the ECO stage. By
ripping up and refining a local window of the entire layout design, CoRe-ECO is capable of
achieving a DRC-clean layout solution. We have demonstrated that our framework successfully
resolves an average of 58.6% (range: 33.3% to 100.0%) of remaining post-ECO route DRCs,
across a range of benchmark circuits with various cell architectures, with no adverse effect on
185
total routed wirelength (average of 0.003% reduction).
Chapter 3 presents two technology-aware physical design methodologies. First, we study
the 2nd DB effect, which is one of the critical LLEs for physical design in sub-10nm technologies,
along with the mixed-DB design that can have both SDB and DDB cells. We propose new
heuristics to recover leakage power increments caused by the 2nd DB effect, including the use of
2nd DB-aware relocation, gate sizing, Vt swapping and DB swapping while satisfying placement
constraints. Our work achieves up to 80% leakage recovery on average using the proposed flow.
Second, we have presented novel dynamic programming-based single- and double-row detailed
placement optimizations with metaheuristics to maximize power staples. We perform extensive
studies on the scalability and sensitivity to parameters, impact of the benefit table, as well as
balance of power and ground staples. Compared to the traditional post-placement flow, we
achieve up to 13.2% (10mV) reduction in IR drop, with almost no WNS degradation compared
to a pre-placement flow.
Chapter 4 presents three design-technology co-optimization works. First, we have
proposed a novel framework to evaluate routability impacts of advanced-node scaling options,
spanning across technology, design enablement and design parameters. Crucially, our framework
can reduce the timelines for design enablement and design implementation that limit today’s
DTCO methodologies. Also, our framework can flexibly support additional technology and
design options. We integrate a powerful SMT-based standard-cell layout generation capability.
Optimal layout solutions are obtained via a unified constraint satisfaction formulation that spans
technology and cell architecture parameters. We validate these aspects of our methodology using
a novel cell-based routability assessment with knight’s tour-based topology generation. We also
perform design-based routability assessment using open-source testcases, and show correlations
of the Kth routability metric to achievable utilization in P&R. Furthermore, we propose learning-
based Kth prediction to reduce runtime and required disk space, and to mitigate P&R tool
license overheads. The experimental studies confirm the capability of our framework to produce
routability assessments across a large range of technology and design parameters. Second, the
186
ML-assisted DTCO framework is applied to study sub-3nm node cell and block-level DTCO.
The regime of CH <120nm with four available tracks shows increasing block-level area due to
routing difficulty, indicating diminishing return on efforts to push ground rules. The proposed
DTCO method will extend to evaluate power and performance along with the cell- and block-
level area density assessment presented in this study. Last, we propose a systematic framework
to explore PPAC tradeoffs with advanced technology nodes, by enabling scaling boosters and
improved design enablements including standard-cell libraries. We enable our predictive 3nm
technology and the corresponding PDKs. Also, we propose an ML-based parameter tuning flow
and clustering-based cell width-regularized placements to generate realistic artificial designs.
We experimentally demonstrate 24% area reduction through use of BSPDN and BPR. Our results
are comparable to the previous works from industry which report 25% to 30% area reduction
from use of BSPDN and BPR.
187
Bibliography
[1] A. Agnesina, K. Chang and S. K. Lim, “VLSI Placement Parameter Optimization Using
Deep Reinforcement Learning”, Proc. IEEE/ACM International Conference on Computer-
Aided Design, 2020, pp. 1-9.
[2] T. Ajayi, V. A. Chhabria, M. Fogaça, S. Hashemi, A. Hosny, A. B. Kahng, M. Kim, J. Lee,

U. Mallappa, M. Neseem, G. Pradipta, S. Reda, M. Saligane, S. S. Sapatnekar, C. Sechen,
M. Shalan, W. Swartz, L. Wang, Z. Wang, M. Woo and B. Xu, “Toward an Open-Source
Digital Flow: First Learnings from the OpenROAD Project”, Proc. ACM/IEEE Design
Automation Conference, 2019, pp. 76:1-76:4.
[3] Y. P. Aneja, “An Integer Linear Programming Approach to the Steiner Problem in Graphs”,
Networks 10(2) (1980), pp. 167-178.
[4] C. Auth, A. Aliyarukunju, M. Asoro, D. Bergstrom, V. Bhagwat, J. Birdsall, N. Bisnik,

M. Buehler, V. Chikarmane, G. Ding, Q. Fu, H. Gomez, W. Han, D. Hanken, M. Haran,
M. Hattendorf, R. Heussner, H. Hiramatsu, B. Ho, S. Jaloviar, I. Jin, S. Joshi, S. Kirby, S.
Kosaraju, H. Kothari, G. Leatherman, K. Lee, J. Leib, A. Madhavan, K. Marla, H. Meyer,
T. Mule, C. Parker, S. Parthasarathy, C. Pelto, L. Pipes, I. Post, M. Prince, A. Rahman, S.
Rajamani, A. Saha, J. Dacuna Santos, M. Sharma, V. Sharma, J. Shin, P. Sinha, P. Smith,
M. Sprinkle, A. St. Amour, C. Stuas, R. Suri, D. Towner, A. Tripathi, A. Tura, C. Ward
and A. Yeoh, “A 10nm High Performance and Low-Power CMOS Technology Featuring
3rd Generation FinFET Transistors, Self-Aligned Quad Patterning, Contact Over Active
Gate and Cobalt Local Interconnects”, Proc. IEEE International Electron Devices Meeting,
2017, pp. 29.1.1-29.1.4.
[5] H. B. Bakoglu and J. D. Meindl, “Optimal Interconnect Circuits for VLSI”, IEEE Transac-
tions on Electron Devices ED-32(5) (1985), pp. 903–909.
[6] R. Berthelon, F. Andrieu, E. Josse, R. Bingert, O. Weber, E. Serret, A. Aurand, S.

Delmedico, V. Farys, C. Bernicot, E. Bechet, E. Bernard, T. Poiroux, D. Rideau, P. Scheer,
E. Baylac, P. Perreau, M.A. Jaud, J. Lacord, E. Petitprez, A. Pofelski, S. Ortolland, P.
Sardin, D. Dutartre, A. Claverie, M. Vinet, J.C. Marin and M. Haond, “Design / Technol-
ogy Co-optimization of Strain-induced Layout Effects in 14nm UTBB-FDSOI CMOS:
Enablement and Assessment of Continuous-RX Designs”, Proc. IEEE Symposium on VLSI
Technology, 2016, pp. 1-2.
188
[7] P. Besser, “BEOL Interconnect Innovations for Improving Performance”, Proc. NCCAVS
Joint Users Group Technical Symposium, 2017.
[8] K. Bhanushali and W. R. Davis, “FreePDK15: An Open-Source Predictive Process Design

Kit for 15nm FinFET Technology”, Proc. International Symposium on Physical Design,
2015, pp. 165–170.
[9] D. Bhattacharya, V. Boppana, R. Roy and J. Roy, “Method for Automated Design of Inte-
grated Circuits with Targeted Quality Objectives Using Dynamically Generated Building
Blocks”, US Patent, US7225423B2, 2007.
[10] K. Boese and A. B. Kahng, “Zero-Skew Clock Routing Trees With Minimum Wirelength”,
Proc. IEEE International ASIC Conference and Exhibit, 1992, pp. 1.1.1 - 1.1.5.
[11] J. L. Burns and J. A. Feldman, “C5M-A Control-Logic Layout Synthesis System for
High-Performance Microprocessors”, IEEE Transactions on Computer-Aided Design of
Integrated Circuits and Systems 17(1) (1998), pp. 14–23.
[12] A. Ceyhan, J. Quijas, S. Jain, H.-Y. Liu, W. E. Gifford and S. Chakravarty, “Machine
Learning-Enhanced Multi-Dimensional Co-Optimization of Sub-10nm Technology Node
Options”, Proc. IEEE International Electron Devices Meeting, 2019, pp. 36.6.1-36.6.4.
[13] W.-T. J. Chan, Y. Du, A. B. Kahng, S. Nath and K. Samadi, “BEOL Stack-Aware Routability
Prediction from Placement Using Data Mining Techniques”, Proc. IEEE International
Conference on Computer Design, 2016, pp. 41-48.
[14] D. Chanemougame, J. Smith, P. Gutwin, B. Byrns and L. Liebmann, “Agile Pathfind-

ing Technology Prototyping: the Hunt for Directional Correctness”, Proc. International
Conference on Simulation of Semiconductor Processes and Devices, 2020, pp. 301-305.
[15] T.-H. Chao, Y.-C. Hsu, J.-M. Ho, K. D. Boese and A. B. Kahng, “Zero Skew Clock Routing
With Minimum Wirelength”, IEEE Transactions on Computer-Aided Design of Integrated
Circuits and Systems 39(11) (1992), pp. 799-814.
[16] W.-C. Chao and W.-K. Mak, “Low-Power Gated and Buffered Clock Network Construc-
tion”, ACM Transactions on Design Automation of Electronic Systems 13(1) (2008), pp.
1-20.
[17] B. Chava, J. Ryckaert, L. Mattii, S. M. Y. Sherazi, P. Debacker, A. Spessot and D. Verkest,

“DTCO Exploration for Efficient Standard Cell Power Rails”, Proc. SPIE 10588, Design-
Process-Technology Co-optimization for Manufacturability XII, 105880B, 2018, pp. 1-6.
[18] B. Chava, K. A. Shaik, A. Jourdain, S. Guissi, P. Weckx, J. Ryckaert, G. Van Der Plaas,
A. Spessot, E. Beyne and A. Mocuta, “Backside Power Delivery as a Scaling Knob for
Future Systems”, Proc. SPIE 10962, Design-Process-Technology Co-optimization for
Manufacturability XIII, 1096205, 2019, pp. 1-6.
189
[19] G. Chen and E. F. Y. Young, “Dim Sum: Light Clock Tree by Small Diameter Sum”, Proc.
Design, Automation & Test in Europe Conference & Exhibition, 2019, pp. 174-179.
[20] T. Chen and C. Guestrin, “XGBoost: A Scalable Tree Boosting System”, Proc. ACM
SIGKDD International Conference on Knowledge Discovery and Data Mining, 2016, pp.
785–794.
[21] L. Chen, C. Huang, Y. Chang and H. Chen, “A Learning-Based Methodology for Routability
Prediction in Placement”, Proc. International Symposium on VLSI Design, Automation and
Test, 2018, pp. 1-4.
[22] C.-K. Cheng, C.-T. Ho, C. Holtz and B. Lin, “Design and System Technology Co-
Optimization Sensitivity Prediction for VLSI Technology Development Using Machine
Learning”, Proc. ACM/IEEE International Workshop on System-Level Interconnect Predic-
tion. 2021, pp. 1-8.
[23] C.-K. Cheng, C.-T. Ho, D. Lee and D. Park, “A Routability-Driven Complimentary-FET
(CFET) Standard Cell Synthesis Framework Using SMT”, Proc. IEEE/ACM International
Conference on Computer-Aided Design, 2020, pp. 1-8.
[24] C.-K. Cheng, A. B. Kahng, H. Kim, M. Kim, D. Lee, D. Park and M. Woo, “PROBE2.0: A
Systematic Framework for Routability Assessment from Technology to Design in Advanced
Nodes”, IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
41(5) (2022), pp. 1495-1508.
[25] C.-K. Cheng, A. B. Kahng, I. Kang and L. Wang, “RePlAce: Advancing Solution Quality
and Routability Validation in Global Placement”, IEEE Transactions on Computer-Aided
Design of Integrated Circuits and Systems 38(9) (2019), pp. 1717-1730.
[26] C.-K. Cheng, D. Lee and D. Park, “Standard-Cell Scaling Framework with Guaranteed
Pin-Accessibility”, Proc. IEEE International Symposium on Circuits and Systems, 2020,
pp. 1-5.
[27] C. Chidambaram, A. B. Kahng, M. Kim, G. Nallapati, S. C. Song and M. Woo, “A Novel

Framework for DTCO: Fast and Automatic Routability Assessment with Machine Learning
for Sub-3nm Technology Options”, Proc. IEEE Symposium on VLSI Technology, 2021, pp.
1-2.
[28] C. Chu and Y.-C. Wong, “FLUTE: Fast Lookup Table Based Rectilinear Steiner Minimal
Tree Algorithm for VLSI Design”, IEEE Transactions on Computer-Aided Design of
Integrated Circuits and Systems 27(1) (2008), pp. 70-83.
[29] L. T. Clark, V. Vashishtha, L. Shifren, A. Gujja, S. Sinha, B. Cline, C. Ramamurthy and G.

Yeric, “ASAP7: A 7-nm FinFET Predictive Process Design Kit”, Microelectronics Journal
53 (2016), pp. 105–115.
190
[30] J. Cong, A. B. Kahng, C. K. Koh and C.-W. A. Tsao, “Bounded-Skew Clock and Steiner
Routing”, ACM Transactions on Design Automation of Electronic Systems 3(3) (1998), pp.
341-388.
[31] J. M. Cohn, J. Venuto, I. L. Wemple and P. S. Zuchowski, “Method for Supply Voltage
Drop Analysis During Placement Phase of Chip Design”, US Patent, US6523154B2, 2000.
[32] P. Cremer, S. Hougardy, J. Schneider and J. Silvanus, “Automatic Cell Layout in the 7nm
Era”, Proc. International Symposium on Physical Design, 2017, pp. 99-106.
[33] W. M. Dai, T. Asano and E. S. Kuh, “Routing Region Definition and Ordering Scheme
for Building-Block Layout”, IEEE Transactions on Computer-Aided Design of Integrated
[34] L. De Moura and N. Bjørner, “Z3: An Efficient SMT Solver”, Proc. International Con-
ference on Tools and Algorithms for the Construction and Analysis of Systems, 2008, pp.
337–340.
[35] Y. Ding, C. Chu and W.-K. Mak, “Pin Accessibility-Driven Detailed Placement Refine-
ment”, Proc. International Symposium on Physical Design, 2017, pp. 133–140.
[36] M. Donno, E. Macii and L. Mazzoni, “Power-Aware Clock Tree Planning”, Proc. Interna-
tional Symposium on Physical Design, 2004, pp. 138-147.
[37] Y. Du and M. D. F. Wong, “Optimization of Standard Cell Based Detailed Placement

for 16nm FinFET Process”, Proc. Design, Automation & Test in Europe Conference &
Exhibition, 2014, pp. 1-6.
[38] M. Edahiro, “A Clustering-Based Optimization Algorithm in Zero-Skew Routings”, Proc.

ACM/IEEE Design Automation Conference, 1993, pp. 612-616.
[39] M. Fernández-Delgado, M. S. Sirsat, E. Cernadas, S. Alawadi, S. Barro and M. Febrero-

Bande, “An Extensive Experimental Survey of Regression Methods”, Neural Networks 111
(2019), pp. 11-34.
[40] G. Flach, T. Reimann, G. Posser, M. Johann and R. Reis, “Effective Method for Simultane-
ous Gate Sizing and Vth Assignment Using Lagrangian Relaxation”, IEEE Transactions
on Computer-Aided Design of Integrated Circuits and Systems 33(4) (2014), pp. 546-557.
[41] J. H. Friedman, “Greedy Function Approximation: A Gradient Boosting Machine”, Annals

of Statistics 29(5) (2001), pp. 1189-1232.
[42] P. Geurts, D. Ernst and L. Wehenkel, “Extremely Randomized Trees”, Machine Learning
63(1) (2006), pp. 3-42.
[43] B. Giraud, O. Thomas, A. Amara, A. Vladimirescu and M. Belleville, SRAM Circuit Design,
Dordrecht, Springer, 2009.
191
[44] P. Gupta, A. B. Kahng, P. Sharma and D. Sylvester, “Gate-Length Biasing for Runtime
Leakage Control”, IEEE Transactions on Computer-Aided Design of Integrated Circuits
and Systems 25(8) (2006), pp. 1475-1485.
[45] M. Guruswamy, R. L. Maziasz, D. Dulitz, S. Raman, V. Chiluvuri, A. Fernandez and L. G.

Jones, “Cellerity: A Fully Automatic Layout Synthesis System for Standard Cell Libraries”,
Proc. Design Automation Conference, 1997, pp. 327-332.
[46] N. Ha, Y.-D. Kim, B.-H. Lee, H.-O. Kim, K. Jeong and J. Kim, “System and Method
of Designing Integrated Circuit by Considering Local Layout Effect”, US Patent,
US20180032658A1, 2018.
[47] L. Hagen, A. B. Kahng, F. Kurdahi and C. Ramachandran, “On the Intrinsic Rent Parameter
and New Spectra-Based Methods for Wireability Estimation”, IEEE Transactions on
Computer-Aided Design of Integrated Circuits and Systems 13(1) (1994), pp. 27-37.
[48] C. Han, K. Han, A. B. Kahng, H. Lee, L. Wang and B. Xu, “Optimal Multi-Row Detailed
Placement for Yield and Model-Hardware Correlation Improvements in Sub-10nm VLSI”,
Proc. IEEE/ACM International Conference on Computer-Aided Design, 2017, pp. 667-674.
[49] K. Han, A. B. Kahng and H. Lee, “Evaluation of BEOL Design Rule Impacts Using
an Optimal ILP-Based Detailed Router”, Proc. ACM/EDAC/IEEE Design Automation
Conference, 2015, pp. 1-6.
[50] K. Han, A. B. Kahng and H. Lee, “Scalable Detailed Placement Legalization for Complex
Sub-14nm Constraints”, Proc. IEEE/ACM International Conference on Computer-Aided
Design, 2015, pp. 867-873.
[51] K. Han, A. B. Kahng, C. Moyes and A. Zelikovsky, “A Study of Optimal Cost-Skew Trade-
off and Remaining Suboptimality in Interconnect Tree Constructions”, Proc. ACM/IEEE
International Workshop on System-Level Interconnect Prediction, 2018, pp. 2:1-2:8.
[52] C. Han, A. B. Kahng, L. Wang and B. Xu, “Enhanced Optimal Multi-Row Detailed Place-
ment for Neighbor Diffusion Effect Mitigation in Sub-10nm VLSI”, IEEE Transactions on
Computer-Aided Design of Integrated Circuits and Systems 38(9) (2019), pp. 1703-1716.
[53] M. Hanan, “On Steiner’s Problem with Rectilinear Distance”, SIAM Journal on Applied
Mathematics 14(2), pp. 255–265.
[54] X. He, Y. Wang, Y. Guo and E. F. Y. Young, “Ripple 2.0: Improved Movement of Cells in
Routability-Driven Placement”, ACM Transactions on Design and Automation of Electronic
Systems 22(1) (2016), pp. 10:1-10:26.
[55] M. O. Hossen, B. Chava, G. Van der Plas, E. Beyne and M. S. Bakir, “Power Delivery
Network (PDN) Modeling for Backside-PDN Configurations With Buried Power Rails and
µ TSVs”, IEEE Transactions on Electron Devices 67(1) (2020), pp. 11-17.
192
[56] M.-K. Hsu, N. Katta, H. Y.-H. Lin, K. T.-H. Lin, K. H. Tam and K. C.-H. Wang, “Design
and Manufacturing Process Co-Optimization in Nano-Technology”, Proc. IEEE/ACM
International Conference on Computer-Aided Design, 2014, pp. 574-581.
[57] J. Hu, A. B. Kahng, S. Kang, M.-C. Kim and I. L. Markov, “Sensitivity-Guided Metaheuris-
tics for Accurate Discrete Gate Sizing”, Proc. IEEE/ACM International Conference on
Computer-Aided Design, 2012, pp. 233-239.
[58] J.-H. Huang, A. B. Kahng and C.-W. A. Tsao, “On the Bounded-Skew Clock and Steiner
Routing Problems”, Proc. Design Automation Conference, 1995, pp. 508-513.
[59] K. Jo, S. Ahn, J. Do, T. Song, T. Kim and K. Choi, “Design Rule Evaluation Framework
Using Automatic Cell Layout Generator for Design Technology Co-Optimization”, IEEE
Transactions on Very Large Scale Integration Systems 27(8) (2019), pp. 1933-1946.
[60] D. Johnson, “Near-Optimal Bin Packing Algorithms”, Ph.D. Thesis, Massachusetts Institute
of Technology, 1973.
[61] A. B. Kahng, J. Dodrill and P. Gupta, “Designing in Advanced Technologies: A Quick

Review of Approaches”, Tutorial 4, ACM/ESDA/IEEE Design Automation Conference,
2018.
[62] A. B. Kahng and H. Lee, “Minimum Implant Area-Aware Gate Sizing and Placement”,
Proc. Great Lakes Symposium on Very Large Scale Integration, 2014, pp. 57-62.
[63] A. B. Kahng, S. Kang, S. Kim and B. Xu, “Enhanced Power Delivery Pathfinding for
Emerging 3D Integration Technology”, IEEE Transactions on Very Large Scale Integration
Systems 29(4) (2021), pp. 591-604.
[64] A. Kahng, A. B. Kahng, H. Lee and J. Li, “PROBE: Placement, Routing, Back-End-of-
Line Measurement Utility”, IEEE Transactions on Computer-Aided Design of Integrated
[65] A. B. Kahng, S. Kang, H. Lee, I. L. Markov and P. Thapar, “High-Performance Gate Sizing
with a Signoff Timer”, Proc. IEEE/ACM International Conference on Computer-Aided
Design, 2013, pp. 450-457.
[66] A. B. Kahng, J. Kuang, W.-H. Liu and B. Xu, “In-Route Pin Access-Driven Placement
Refinement for Improved Detailed Routing Convergence”, IEEE Transactions on Computer-
Aided Design of Integrated Circuits and Systems 41(3) (2022), pp. 784-788.
[67] A. B. Kahng, B. Liu and Q. Wang, “Supply Voltage Degradation Aware Analytical Place-
ment”, Proc. International Conference on Computer Design, 2005, pp. 437-443.
[68] A. B. Kahng, I. L. Markov and S. Reda, “On Legalization of Row-Based Placements”,

Proc. ACM Great Lakes Symposium on Very Large Scale Integration, 2004, pp. 214-219.
193
[69] A. B. Kahng and C.-W. A. Tsao, “Planar-DME: A Single-Layer Zero-Skew Clock Tree
Router”, IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
15(1) (1996), pp. 8-19.
[70] A. B. Kahng and C.-W. A. Tsao, “Practical Bounded-Skew Clock Routing”, Journal of
VLSI Signal Processing 16 (1997), pp. 199-215.
[71] A. B. Kahng and Q. Wang, “A Faster Implementation of APlace”, Proc. International

Symposium on Physical Design, 2006, pp. 218-220.
[72] A. B. Kahng, L. Wang and B. Xu, “The Tao of PAO: Anatomy of a Pin Access Oracle for
Detailed Routing”, Proc. ACM/IEEE Design Automation Conference, 2020, pp. 1–6.
[73] I. Kang, D. Park, C. Han and C.-K. Cheng, “Fast and Precise Routability Analysis with
Conditional Design Rules”, Proc. ACM/IEEE International Workshop on System-Level
Interconnect Prediction, 2018, pp. 1-4.
[74] G. Karypis, R. Aggarwal, V. Kumar and S. Shekhar, “Multilevel Hypergraph Partitioning:

Applications in VLSI Domain”, IEEE Transactions on Very Large Scale Integration Systems
7(1) (1999), pp. 69-79.
[75] M.-C. Kim, J. Hu, D. J. Lee and I. L. Markov, “A SimPLR Method for Routability-Driven
Placement”, Proc. IEEE/ACM International Conference on Computer-Aided Design, 2011,
pp. 67-73.
[76] S. Kim, K. Jo and T. Kim, “Boosting Pin Accessibility through Cell Layout Topology
Diversification”, Proc. Asia and South Pacific Design Automation Conference, 2021, pp.
1–6.
[77] T. Kim, J. Jeong, S. Woo, J. Yang, H. Kim, A. Nam, C. Lee, J. Seo, M. Kim, S. Ryu, Y. Oh
and T. Song, “NS3K: A 3-nm Nanosheet FET Standard Cell Library Development and Its
Impact”, IEEE Transactions on Very Large Scale Integration Systems (2022), pp. 1-14.
[78] D. Kim, S.-Y. Lee, K. Min and S. Kang, “Construction of Realistic Place-and-Route
Benchmarks for Machine Learning Applications”, IEEE Transactions on Computer-Aided
Design of Integrated Circuits and Systems (2022), doi:10.1109/TCAD.2022.3209530.
[79] G. Ke, Q. Meng, T. Finley, T. Wang, W. Chen, W. Ma, Q. Ye and T.-Y. Liu, “LightGBM: A
Highly Efficient Gradient Boosting Decision Tree”, Proc. Advances in Neural Information
Processing Systems, 2017, pp. 3146-3154.
[80] E. D. Kurniawan, H. Yang, C.-C. Lin and Y.-C. Wu, “Effect of Fin Shape of Tapered
FinFETs on the Device Performance in 5-nm Node CMOS Technology”, Microelectronics
Reliability 83 (2018), pp. 254-259.
[81] B. Landman and R. Russo, “On a Pin Versus Block Relationship for Partitions of Logic
Graphs”, IEEE Transactions on Computers C-20(12) (1971), pp. 1469-1479.
194
[82] D. Lee, D. Park, C.-T. Ho, I. Kang, H. Kim, S. Gao, B. Lin and C.-K. Cheng, “SP&R:
SMT-based Simultaneous Place-&-Route for Standard Cell Synthesis of Advanced Nodes”,
IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems 40(10)
(2021), pp. 2142-2155.
[83] Y.-L. Li, S.-T. Lin, S. Nishizawa, H.-Y. Su, M.-J. Fong, O. Chen and H. Onodera, “NCTU-
cell: A DDA-Aware Cell Library Generator for FinFET Structure with Implicitly Adjustable
Grid Map”, Proc. ACM/IEEE Design Automation Conference, 2019, pp. 1-6.
[84] L. Liebmann, D. Chanemougame, P. Churchill, J. Cobb, C.-T. Ho, V. Moroz and J. Smith,
“DTCO Acceleration to Fight Scaling Stagnation”, Proc. SPIE 11328, Design-Process-
Technology Co-optimization for Manufacturability XIV, 2020, 113280C, pp. 1-15.
[85] L. Liebmann, V. Gerousis, P. Gutwin, X. Zhu and J. Petykiewicz, “Exploiting Regular-

ity: Breakthroughs in Sub-7nm Place-and-Route”, Proc. SPIE 10148, Design-Process-
Technology Co-optimization for Manufacturability XI, 2017, pp. 1-10.
[86] T. Lin and C. Chu, “POLAR 2.0: An Effective Routability-Driven Placer”, Proc.
ACM/EDAC/IEEE Design Automation Conference, 2014, pp. 1-6.
[87] Y. Lin, B. Yu, X. Xu, J.-R. Gao, N. Viswanathan, W.-H. Liu, Z. Li, C. J. Alpert and D. Z.
Pan, “MrDP: Multiple-row Detailed Placement of Heterogeneous-sized Cells for Advanced
Nodes”, Proc. IEEE/ACM International Conference on Computer-Aided Design, 2016, pp.
7:1-7:8.
[88] W.-H. Liu, C.-K. Koh and Y.-L. Li, “Optimization of Placement Solutions for Routability”,
Proc. ACM/EDAC/IEEE Design Automation Conference, 2013, pp. 1-9.
[89] L.-C. Lu, “Physical Design Challenges and Innovations to Meet Power, Speed, and Area
Scaling Trend”, Proc. ACM International Symposium on Physical Design, 2017, p. 63.
[90] T. Luo, D. Newmark and D. Z. Pan, “Total Power Optimization Combining Placement,
Sizing and Multi-Vt Through Slack Distribution Management”, Proc. Asia and South
Pacific Design Automation Conference, 2008, pp. 352-357.
[91] L. Maaten and G. Hinton, “Visualizing Data Using t-SNE”, Journal of Machine Learning
Research 9 (2008), pp. 2579-2605.
[92] P. Majumder, B. Kumthekar, N. R. Shah, J. Mowchenko, P. A. Chavda, Y. Kojima, H.

Yoshida and V. Boppana, “Method of IC Design Optimization via Creation of Design-
Specific Cells from Post-Layout Patterns”, US Patent, US7941776B2, 2011.
[93] W. K. Mak, W. S. Kuo, S. H. Zhang, S. I. Lei and C. Chu, “Minimum Implant Area-Aware
Placement and Threshold Voltage Refinement”, IEEE Transactions on Computer-Aided
Design of Integrated Circuits and Systems 36(7) (2017), pp. 1103-1112.
195
[94] L. Mattii, D. M. Milojevic, P. Debacker, Y. Sherazi, M. Berekovic and P. Raghavan, “IR-
Drop Aware Design & Technology Co-Optimization for N5 Node with Different Device
and Cell Height Options”, Proc. IEEE/ACM International Conference on Computer-Aided
Design, 2017, pp. 89-94.
[95] M. D. McKay, W. J. Conover and R. J. Beckman, “A Comparison of Three Methods for

Selection Values of Input Variables in the Analysis of Output from a Computer Code”,
Technometrics 21(2) (1979), pp. 239–245.
[96] C. W. Moon, P. Gupta, P. J. Donehue and A. B. Kahng, “Method of Designing a Digital

Circuit by Correlating Different Static Timing Analyzers”, US Patent, US7823098B1,
2010.
[97] J. A. Nelder and R. W. M. Wedderburn, “Generalized Linear Models”, Journal of the Royal
Statistical Society Series A (Statistics in Society) 135(3) (1972), pp. 370-384.
[98] J. Oh, I. Pyo and M. Pedram, “Constructing Lower and Upper Bounded Delay Routing Trees
Using Linear Programming”, Proc. Design Automation Conference, 1996, pp. 401-404.
[99] D. Park, I. Kang, Y. Kim, S. Gao, B. Lin and C.-K. Cheng, “ROAD: Routability Analysis
and Diagnosis Framework Based on SAT Techniques”, Proc. International Symposium on
Physical Design, 2019, pp. 65-72.
[100] D. Park, D. Lee, I. Kang, C. Holtz, S. Gao, B. Lin and C.-K. Cheng, “Grid-Based
Framework for Routability Analysis and Diagnosis with Conditional Design Rules”, IEEE
Transactions on Computer-Aided Design of Integrated Circuits and Systems 39(12) (2020),
pp. 5097-5110.
[101] G. Posser, V. Mishra, P. Jain, R. Reis and S. S. Sapatnekar, “A Systematic Approach

for Analyzing and Optimizing Cell-Internal Signal Electromigration”, Proc. IEEE/ACM
International Conference on Computer-Aided Design, 2014, pp. 486-491.
[102] D. Prasad, S. S. T. Nibhanapudi, S. Das, O. Zografos, B. Chehab, S. Sarkar, R. Baert,

A. Robinson, A. Gupta, A. Spessot, P. Debacker, D. Verkest, J. Kulkarni, B. Cline and
S. Sinha, “Buried Power Rails and Back-side Power Grids: Arm® CPU Power Delivery
Network Design Beyond 5nm”, Proc. IEEE International Electron Devices Meeting, 2019,
pp. 19.1.1-19.1.4.
[103] X. Qiu and M. Marek-Sadowska, “Can Pin Access Limit the Footprint Scaling?”, Proc.
Design Automation Conference, 2012, pp. 1100–1106.
[104] J. Ryckaert, A. Gupta, A. Jourdain, B. Chava, G. Van der Plas, D. Verkest and E. Beyne,
“Extending the Roadmap Beyond 3nm through System Scaling Boosters: A Case Study on
Buried Power Rail and Backside Power Delivery”, Proc. IEEE Electron Devices Technology
and Manufacturing Conference, 2019, pp. 50-52.
196
[105] N. Ryzhenko, S. Burns, A. Sorokin and M. Talalay, “Pin Access-Driven Design Rule
Clean and DFM Optimized Routing of Standard Cells under Boolean Constraints”, Proc.
International Symposium on Physical Design, 2019, pp. 41–47.
[106] S. Sadangi, “FreePDK3: A Novel PDK for Physical Verification at the 3nm Node”, M.S.
Thesis, Computer Engineering, North Carolina State University, 2021.
[107] R. Sengupta, J. G. Hong and M. Rodder, “Semiconductor Device Having Buried Power
Rail”, US Patent, US9570395, 2017.
[108] J. Seo, J. Jung, S. Kim and Y. Shin, “Pin Accessibility-Driven Cell Layout Redesign and
Placement Optimization”, Proc. ACM/EDAC/IEEE Design Automation Conference, 2017,
pp. 54:1-54:6.
[109] S. M. Y. Sherazi, J. K. Chae, P. Debacker, L. Mattii, P. Raghavan, V. Gerousis, D.

Verkest, A. Mocuta, R. H. Kim, A. Spessot and J. Ryckaert, “Track Height Reduction
for Standard-Cell in Below 5nm Node: How Low Can You Go?”, Proc. SPIE 10588,
Design-Process-Technology Co-optimization for Manufacturability XII, 1058809, 2018, pp.
1058809-1:1058809-13.
[110] S. M. Y. Sherazi, M. Cupak, P. Weckx, O. Zografos, D. Jang, P. Debacker, D. Verkest,

A. Mocuta, R. H. Kim, A. Spessot and J. Ryckaert, “Standard-Cell Design Architecture
Options Below 5nm Node: The Ultimate Scaling of FinFET and Nanosheet”, Proc. SPIE
10962, Design-Process-Technology Co-optimization for Manufacturability XIII, 1096202,
2019, pp. 1096202:1-1096202:15.
[111] S. M. Y. Sherazi, C. Jha, D. Rodopoulos, P. Debacker, B. Chava, L. Mattii, M. G. Bardon,

P. Schuddinck, P. Raghavan, V. Gerousis, A. Spessot, D. Verkest, A. Mocuta, R. H. Kim
and J. Ryckaert, “Low Track Height Standard Cell Design in iN7 Using Scaling Boosters”,
Proc. SPIE 10148, Design-Process-Technology Co-optimization for Manufacturability XI,
101480Y, 2017, pp. 101480Y:1-101480Y:8.
[112] G. Sisto, B. Chehab, B. Genneret, R. Baert, R. Chen, P. Weckx, J. Ryckaert, R. Chou, G.

Van der Plas, E. Beyne and D. Milojevic, “IR-Drop Analysis of Hybrid Bonded 3D-ICs
with Backside Power Delivery and µ- & n-TSVs”, Proc. IEEE International Interconnect
Technology Conference, 2021, pp. 1-3.
[113] S. C. Song, J. Xu, D. Yang, K. Rim, P. Feng, J. Bao, J. Zhu, J. Wang, G. Nallapati, M.
Badaroglu, P. Narayanasetti, B. Bucki, J. Fischer and G. Yeap, “Unified Technology Opti-
mization Platform using Integrated Analysis (UTOPIA) for Holistic Technology, Design
and System Co-Optimization at <= 7nm Nodes”, Proc. IEEE Symposium on VLSI Circuits,
2016, pp. 1-2.
[114] H. Su, J. Hu, S. S. Sapatnekar and S. R. Nassif, “Congestion-Driven Codesign of Power

and Signal Networks”, Proc. Design Automation Conference, 2002, pp. 64-69.
197
[115] S. Sutanthavibul, E. Shragowitz and J. B. Rosen, “An Analytical Approach to Floorplan
Design and Optimization”, IEEE Transactions on Computer-Aided Design of Integrated
Circuits and Systems 10(6) (1991), pp. 761–769.
[116] S. S. Teja Nibhanupudi, D. Prasad, S. Das, O. Zografos, A. Robinson, A. Gupta, A.
Spessot, P. Debacker, D. Verkest, J. Ryckaert, G. Hellings, J. Myers, B. Cline and J. P.
Kulkarni, “A Holistic Evaluation of Buried Power Rails and Back-Side Power for Sub-5nm
Technology Nodes”, IEEE Transactions on Electron Devices 69(8), 2022, pp. 4453-4459.
[117] I. Tseng, Z. C. Lee, V. Tripathi, C. M. T. Yip, Z. Chen and J. Ong, “A System for Standard
Cell Routability Checking and Placement Routability Improvements”, Proc. IEEE Asia
Pacific Conference on Circuits and Systems, 2019, pp. 125-128.
[118] P. Van Cleeff, S. Hougardy, J. Silvanus and T. Werner, “BonnCell: Automatic Cell Layout
in the 7nm Era”, IEEE Transactions on Computer-Aided Design of Integrated Circuits and
Systems 39(10) (2020), pp. 2872-2885.
[119] V. Vashishtha, M. Vangala and L. T. Clark, “ASAP7 Predictive Design Kit Development
and Cell Design Technology Co-optimization”, Proc. IEEE/ACM International Conference
on Computer-Aided Design, 2017, pp. 992-998.
[120] A. Vittal and M. Marek-Sadowska, “Low-Power Buffered Clock Tree Design”, IEEE
Transactions on Computer-Aided Design of Integrated Circuits and Systems 16(9) (1997),
pp. 965-975.
[121] D. M. Warme, P. Winter and M. Zachariasen, “Exact Algorithms for Plane Steiner Tree
Problems: A Computational Study”, Advances in Steiner Trees (D. Z. Du, J. M. Smith and
J. H. Rubinstein, eds.), Kluwer Academic Publishers, 2000, pp. 81-116.
[122] L. Wei, K. Roy and C.-K. Koh, “Power Minimization by Simultaneous Dual-Vth As-
signment and Gate-Sizing”, Proc. IEEE Custom Integrated Circuits Conference, 2000, pp.
413-416.
[123] H.-S. P. Wong, “Semiconductor Technology: A System Perspective”, Opening
Keynote, ACM/IEEE Design Automation Conference, 2020. http://www2.dac.com/events/
eventdetails.aspx?id=295-151
[124] Z. Xie, Y.-H. Huang, G.-Q. Fang, H. Ren, S.-Y. Fang, Y. Chen and J. Hu, “RouteNet:
Routability Prediction for Mixed-Size Designs Using Convolutional Neural Network”,
Proc. IEEE/ACM International Conference on Computer-Aided Design, 2018, pp. 1-8.
[125] R. Xie, K.-Y. Lim, M. G. Sung and R. R.-H. Kim, “Methods of Forming Single and
Double Diffusion Breaks on Integrated Circuit Products Comprised of FinFET Devices and
the Resulting Products”, US Patent, US9412616, 2016.
[126] X. Xu, B. Cline, G. Yeric, B. Yu and D. Z. Pan, “Self-Aligned Double Patterning Aware
Pin Access and Standard Cell Layout Co-Optimization”, IEEE Transactions on Computer-
Aided Design of Integrated Circuits and Systems 34(5) (2015), pp. 699–712.
198
[127] X. Xu, Y. Lin, V. Livramento and D. Z. Pan, “Concurrent Pin Access Optimization for
Unidirectional Routing”, Proc. ACM/EDAC/IEEE Design Automation Conference, 2017,
pp. 1-6.
[128] S. Yang, Y. Liu, M. Cai, J. Bao, P. Feng, X. Chen, L. Ge, J. Yuan, J. Choi, P. Liu, Y. Suh,
H. Wang, J. Deng, Y. Gao, J. Yang, X.-Y. Wang, D. Yang, J. Zhu, P. Penzes, S.C. Song, C.
Park, S. Kim, J. Kim, S. Kang, E. Terzioglu, K. Rim and PR. C. Chidambaram, “10nm
High Performance Mobile SoC Design and Technology Co-Developed for Performance,
Power, and Area Scaling”, Proc. Symposium on VLSI Technology, 2017, pp. T70-T71.
[129] G. Yeric, “Circuit Application Requirements”, Proc. IEEE International Electron Devices
Meeting, Short Course, 2014.
[130] T.-C. Yu, S.-Y. Fang, H.-S. Chiu, K.-S. Hu, P. H.-Y. Tai, C. C.-F. Shen and H. Sheng,
“Pin Accessibility Prediction and Optimization with Deep Learning-Based Pin Pattern
Recognition”, Proc. ACM/IEEE Design Automation Conference, 2019, pp. 1–6.
[131] C.-Y. Yu, Y.-T. Hou, C.-M. Fu, W.-H. Chen and W.-Y. Lo, “Apparatus and Method for Miti-
gating Dynamic IR Voltage Drop and Electromigration Affects”, US Patent, US9768119B2,
2017.
[132] A. Z. Zelikovsky and I. I. Mandoiu, “Practical Approximation Algorithms for Zero- and
Bounded-Skew Trees”, SIAM Journal Discrete Mathematics 15(1) (2002), pp. 97-111.
[133] Z. Zhang, R. Wang, C. Chen, Q. Huang, Y. Wang, C. Hu, D. Wu, J. Wang and R. Huang,
“New-Generation Design-Technology Co-Optimization (DTCO): Machine-Learning As-
sisted Modeling Framework”, Proc. Silicon Nanoelectronics Workshop, 2019, pp. 1-2.
[134] P. Zhao, S. M. Pandey, E. Banghart, X. He, R. Asra, V. Mahajan, H. Zhang, B. Zhu, K.

Yamada, L. Cao, P. Balasubramaniam, M. Joshi, M. Eller, F. Benistant and S. Samavedam,
“Influence of Stress Induced CT Local Layout Effect (LLE) on 14nm FinFET”, Proc.
Symposium on VLSI Technology, 2017, pp. T228-T229.
[135] Q. Zhou, X. Wang, Z. Qi, Z. Chen, Q. Zhou and Y. Cai, “An Accurate Detailed Routing
Routability Prediction Model in Placement”, Proc. Asia Symposium on Quality Electronic
Design, 2015, pp. 119-122.
[136] A. M. Ziesemer and R. A. da Luz Reis, “Simultaneous Two-Dimensional Cell Layout

Compaction Using MILP with ASTRAN”, Proc. IEEE Computer Society Annual Sympo-
sium on VLSI, 2014, pp. 350-355.
[137] 68-95-99.7 Rule. https://en.wikipedia.org/wiki/68-95-99.7 rule
[138] Applied Materials Logic Master Class. https://ir.appliedmaterials.com/static-files/

acba6be3-4778-41eb-9183-5c8e52884dea
[139] An Open-Source Knight’s Tour Solver by Warnsdorff-Rule Algorithm. https://github.com/

NiloofarShahbaz/knight-tour-warnsdorff
199
[140] Artificial Netlist Generator. https://github.com/daeyeon22/artificial netlist generator
[141] D. O’Laughlin, ”Backside Power Delivery and Bold Bets at Intel”, June 2022. https:
//www.fabricatedknowledge.com/p/backside-power-delivery-and-bold
[142] Cadence Design Systems, Inc., Sr Software Engineering Group Director, Personal Com-
munication, 2019.
[143] Cadence Genus User Guide. http://www.cadence.com
[144] Cadence Innovus User Guide. http://www.cadence.com
[145] Cadence Liberate User Guide. http://www.cadence.com
[146] Cadence PVS User Guide. http://www.cadence.com
[147] Cadence QRC Extraction User Guide. http://www.cadence.com
[148] Cadence Voltus IC Power Integrity Solution User Guide. http://www.cadence.com
[149] Design Technology Team, Samsung Electronics Co. Ltd., Personal Communication, May
2018.
[150] FreePDK3 Predictive Process Design Kit. https://github.com/ncsu-eda/FreePDK3
[151] GDT to GDS Format Translator. https://sourceforge.net/projects/gds2
[152] O. Tange, GNU Parallel 2018. https://doi.org/10.5281/zenodo.1146014
[153] Gurobi Optimizer Reference Manual. http://www.gurobi.com
[154] H2O AutoML. https://www.h2o.ai
[155] IBM ILOG CPLEX. http://www.ilog.com/products/cplex
[156] IEEE International Roadmap for Devices and Systems (IRDS) 2020 Edition. https://irds.
ieee.org/editions/2020
[157] i.MX RT600 Crossover MCU with Arm Cortex-M33 and DSP Cores Data Sheet. https:
//www.nxp.com/docs/en/data-sheet/DS-RT600.pdf
[158] LEF/DEF Reference 5.8. http://www.si2.org/openeda.si2.org/projects/lefdefnew
[159] Liberty Library Format. http://www.opensourceliberty.org
[160] lowRISC. https://www.lowrisc.org
[161] OpenCores: Open Source IP-Cores. http://www.opencores.org
[162] OpenDB. https://github.com/The-OpenROAD-Project/OpenDB
200
[163] OpenMP Architecture Review Board, “OpenMP Application Program Interface, Version
4.0”.
[164] OpenSTA: Static Timing Analyzer. https://github.com/abk-openroad/OpenSTA
[165] K. Jeong, A. B. Kahng and H. Yao, RentCon: Rent Parameter Evaluation Using Different
Methods. https://vlsicad.ucsd.edu/WLD/index.html
[166] Scaling The Lowly SRAM. https://semiengineering.com/scaling-the-lowly-sram
[167] Si2 OpenAccess. http://www.si2.org/?page=69
[168] Siemens EDA Calibre User Guide. https://eda.sw.siemens.com/en-US/ic/calibre-design
[169] C.-K. Cheng, D. Lee and D. Park, SMT-Based Standard-Cell Layout Generator for
PROBE2.0.
https://github.com/ckchengucsd/SMT-based-STDCELL-Layout-Generator-for-PROBE2.
0
[170] SweRV RISC-V Core 1.1 from Western Digital.

https://github.com/westerndigitalcorporation/swerv eh1
[171] Synopsys Design Compiler User Guide. http://www.synopsys.com
[172] Synopsys IC Compiler II User Guide. http://www.synopsys.com
[173] Synopsys StarRC User Guide. http://www.synopsys.com
[174] Synopsys DTCO Flow: Technology Development. https://www.synopsys.com/silicon/

resources/articles/dtco-flow.html
[175] The iPhone XS & XS Max Review: Unveiling the Silicon Secrets. https://www.anandtech.
com/show/13392/the-iphone-xs-xs-max-review-unveiling-the-silicon-secrets/2
[176] The OpenROAD Project GitHub Repository. https://github.com/The-OpenROAD-Project
[177] The PROBE3.0 Framework. https://github.com/ABKGroup/PROBE3.0
[178] Z3 SMT Solver. https://github.com/Z3Prover/z3
201

qt5wb14729 Nosplash

Uploaded by

Copyright:

Available Formats

qt5wb14729 Nosplash

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

qt5wb14729 Nosplash

Uploaded by

Copyright:

Available Formats

UNIVERSITY OF CALIFORNIA SAN DIEGO

Robust Physical Design and Design Technology Co-Optimization Methodologies at

A dissertation submitted in partial satisfaction of the

Electrical Engineering (Computer Engineering)

Professor Andrew B. Kahng, Chair

University of California San Diego

Dissertation Approval Page . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii

List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii

List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiv

Abstract of the Dissertation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxiv

Chapter 2 General Physical Design Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

Chapter 3 Technology-Aware Physical Design Methodology . . . . . . . . . . . . . . . . . . . . . 59

Chapter 4 Design-Technology Co-Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

Chapter 5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185

Figure 1.2. Power and performance requirements for applications [89]. . . . . . . . . . . . . . 3

Figure 1.5. Overview of this thesis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

Figure 2.10. Framework overview. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

Figure 2.11. Grid-based place-and-route architecture. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

Figure 2.16. Dynamic cell allocation (DCA). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

Figure 2.17. Weighted extendable pin candidates. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

Figure 2.19. Overall ECO flow using CoRe-ECO framework. . . . . . . . . . . . . . . . . . . . . . . 51

Figure 2.21. DRC reductions by CoRe-ECO rounds for AES (4T/ET/MPO2). . . . . . . . . 56

Figure 3.8. Illustration of power staple insertions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

Figure 3.9. The case represented by D[i][1][4][2][2]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86

Figure 3.15. Heatmaps of vectorless dynamic IR drop for AES. . . . . . . . . . . . . . . . . . . . . 95

Figure 4.4. A grid-based standard-cell architecture and technology parameters for

Figure 4.5. DR-MAR, DR-EOL and DR-VR design rules. . . . . . . . . . . . . . . . . . . . . . . . . 109

Figure 4.9. Example of knight’s tour-based connections in a placement of 2-input cells.

Figure 4.12. Design-level assessments with various technology parameters. . . . . . . . . . . 127

Figure 4.17. An example of neighbor-swap operation. The blue-colored cell is swapped

Figure 4.30. ML-based parameter tuning for ANG. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165

Table 2.1. Notations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

Table 2.2. Average ILP runtime when α = 1, β = 0.5, L = M − B. M denotes the

Table 2.4. Notations for the proposed CoRe-ECO framework. . . . . . . . . . . . . . . . . . . . . 38

Table 2.5. Standard-cell architectures. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

Table 2.6. Experimental statistics. RT = the number of routing tracks, DRSet =

Table 3.2. Notations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

Table 3.3. Experimental results for all the design blocks. . . . . . . . . . . . . . . . . . . . . . . . . 71

Table 3.6. Notations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80

Table 3.7. Design information. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88

Table 3.9. IR drop vs. staple length. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91

Table 3.10. Experimental results using double-row optimization. . . . . . . . . . . . . . . . . . . . 93

Table 4.1. Standard-cell architectures in our experiments. . . . . . . . . . . . . . . . . . . . . . . . . 109

Table 4.4. Technology and design parameters in our experiments. . . . . . . . . . . . . . . . . . 112

Table 4.7. Input features of the Kth prediction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121

Table 4.8. Default parameters in our experiments. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122

Table 4.9. List of 39 standard cells per generated library. . . . . . . . . . . . . . . . . . . . . . . . . . 123

Table 4.18. Layer definition in the PROBE3.0 technology. . . . . . . . . . . . . . . . . . . . . . . . . 150

Table 4.19. Key features of the PROBE3.0 technology. . . . . . . . . . . . . . . . . . . . . . . . . . . . 151

Table 4.20. List of 41 standard cells per generated library. . . . . . . . . . . . . . . . . . . . . . . . . . 152

Table 4.22. Definition of topological parameters in ANG [78] [140]. . . . . . . . . . . . . . . . . 160

Table 4.26. Comparison of the width-regularized clustered netlist produced by CWR-FC

Table 4.28. Tools and versions in our experiments. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175

2011 B. S., Electrical Engineering,

Robust Physical Design and Design Technology Co-Optimization Methodologies at

Doctor of Philosophy in Electrical Engineering (Computer Engineering)

University of California San Diego, 2023

Professor Andrew B. Kahng, Chair

The semiconductor industry has achieved remarkable progress by adhering to Moore’s