Title: Asymptotic Probability Extraction for Non-Normal Distributions of Circuit Performance By: Sedigheh Hashemi 201C-Spring2009
1Asymptotic Probability Extraction for Non-Normal
Distributions of Circuit PerformanceBy
Sedigheh Hashemi201C-Spring2009
2Asymptotic Probability Extraction for Non-Normal
Distributions of Circuit Performance
- X. Li, P. Gopalakrishnan and L. Pileggi, CMU
- J. Le, Extreme DA
3Overview
- Introduction
- Asymptotic Probability EXtraction (APEX)
- Implementation of APEX
- Numerical examples
- Conclusion
4IC Technology Scaling
Feature Size Scale Down
Year Leff (nm) W L Tox Vth H ?
1997 250 25.0 32.0 8.0 10.0 25.0 22.2
1999 180 26.2 33.3 8.0 10.0 30.0 24.0
2002 130 28.0 34.6 9.8 10.0 30.0 27.3
2005 100 30.0 40.0 12.0 11.4 33.8 31.7
2006 70 33.3 47.1 16.0 13.3 35.7 33.3
Process Variations (3s / Nominal) Nassif 01
Process variation is becoming relatively larger!
5Statistical Problems in IC
- Statistical methods have been proposed to address
various statistical problems - We focus on analysis problem in this work
6Modeling Process Variations
- Assumption
- Process variations ?xi satisfy Normal
distributions N(0,si) - Principle component analysis (PCA)
- ?xi can be decomposed into independent ?yi
N(0,1)
7Response Surface Model
8Response Surface Model
- A low noise amplifier example designed in IBM
0.25 µm process
Normal Distribution ?yi
Performance Linear Quadratic
F0 1.76 0.14
S11 6.40 1.32
S12 3.44 0.61
S21 2.94 0.34
S22 5.56 3.47
NF 2.38 0.23
IIP3 4.49 0.91
Power 3.79 0.70
Nonlinear Transform
Non-Normal Distribution p
Regression Modeling Error for LNA
9Moment Matching
- Key idea
- Conceptually consider PDF as the impulse response
of an LTI system
Nonlinear Transform
Unknown PDF
Normal Distribution
Match Moments
LTI System
Impulse Excitation
10Moment Matching
- Match the first 2M moments
Impulse Excitation
Impulse Response
11Connection to Probability Theory
- F(?) is called characteristic function in
probability theory - We actually match the first 2M terms of Taylor
expansion at ? 0
System Theory
Probability Theory
12Connection to Probability Theory
- Proposition 1
- Proposition 2
- Typical characteristic functions are "low-pass
filters" - A low-pass system is determined by its behavior
in low-freq band (? 0) - Taylor expansion is accurate around expansion
point (? 0) - Moment matching is efficient in approximating
low-pass systems Celik 02
Characteristic Function for Typical Random
Distributions
Celik 02 IC Interconnect Analysis, Kluwer
Academic Publishers, 2002
13The Classical Moment Problem
T. Stieltjes 1894
RSM
Moment Matching
Probability Extraction
pdf(p)
pdf(p)
14APEX Asymptotic Probability Extraction
- Classical moment problem
- Existence uniqueness of the solution
- Find complete bases to expand PDF function space
15Direct Moment Evaluation
- If ?y1, ?y2,... are independent standard Normal
distribution N(0,1) - Require computing symbolic expression for pk(Y)
16Binomial Moment Evaluation
- Key idea
- Recursively compute high order moments
- Derived from eigenvalue decomposition
statistical independence theory
17Step 1 Model Diagonalization
?zi are independent N(0,1) since eigenvectors U
are orthogonal !
18Step 2 Moment Evaluation
- NOT compute symbolic expression for pk(Y)
- Achieve more than 106x speedup compared with
direct evaluation
19Overview
- Introduction
- Asymptotic Probability EXtraction (APEX)
- Implementation of APEX
- PDF/CDF shifting
- Reverse PDF/CDF evaluation
- Numerical examples
- Conclusion
20PDF/CDF Shifting
- PDF/CDF shifting is required in two cases
- Over-shifting results in large approximation
error - The challenging problem is to accurately
determine ?
?
?
pdf(p)
pdf(p)
p
0
0
p
Mean µ
Mean µ
Case 1 Not Causal
Case 2 Large Delay
21PDF/CDF Shifting
- Exact ? doesn't exist since pdf(p) is unbounded
- Define a bound ? such that the probability P(p
µ-?) is sufficiently small - Propose a generalized Chebyshev inequality to
estimate ? using central moments
22Reverse PDF/CDF Evaluation
- Final value theorem of Laplace transform
- Moment matching is accurate for estimating upper
bound - Use flipped pdf(-p) for estimating lower bound
23Overview
- Introduction
- Asymptotic Probability EXtraction (APEX)
- Implementation of APEX
- Numerical examples
- Conclusion
24ISCAS'89 S27
Longest Path in ISCAS'89 S27
- ST 0.13 µm process
- 6 principal random factors
- MOSFET variations
- No intra-die variation
- No interconnect variation
- Linear delay modeling error
- 4.48
- Quadratic delay modeling error
- 1.10 (4x smaller)
25ISCAS'89 S27
- Binomial moment evaluation achieves more than
106x speedup
Moment Order Direct Direct Binomial
Moment Order of Terms Time (Sec.) Time (Sec.)
1 28 1.00 ? 10-2 0.01
3 924 3.02 ? 100 0.01
5 8008 2.33 ? 102 0.01
6 18564 1.57 ? 103 0.01
7 38760 8.43 ? 103 0.02
8 74613 3.73 ? 104 0.02
15 0.04
20 0.07
Moment Evaluation
Computation Time for Moment Evaluation
26ISCAS'89 S27
Delay
- Numerical oscillation for low order approximation
- Increasing approx. order provides better accuracy
- Typical approx. order is 7 10
Cumulative Distribution Function for Delay
27ISCAS'89 S27
- APEX is the most accurate approach
- APEX achieves more than 200x speedup compared
with MC 104 runs - APEX 0.18 seconds
- MC 104 runs 43.44 seconds
Linear Legendre APEX
1 Point 1.43 0.87 0.04
10 Point 4.63 0.02 0.01
25 Point 5.76 0.12 0.03
50 Point 6.24 0.05 0.02
75 Point 5.77 0.03 0.02
90 Point 4.53 0.16 0.03
99 Point 0.18 0.78 0.09
Comparison on Estimation Error
28Low Noise Amplifier
- IBM 0.25 µm process
- 8 principal random factors
- MOSFET RCL variations
- No mismatches
Performance Linear Quadratic
F0 1.76 0.14
S11 6.40 1.32
S12 3.44 0.61
S21 2.94 0.34
S22 5.56 3.47
NF 2.38 0.23
IIP3 4.49 0.91
Power 3.79 0.70
Regression Modeling Error for LNA
Circuit Schematic for LNA
29Low Noise Amplifier
- APEX is the most accurate approach
- APEX achieves more than 200x speedup compared
with MC 104 runs - APEX 1.29 seconds
- MC 104 runs 334.37 seconds
Performance Corner Corner Linear Linear Legendre Legendre APEX APEX
Performance 1 99 1 99 1 99 1 99
F0 15.8 20.1 1.11 1.10 0.20 0.55 0.06 0.05
S11 45.4 51.5 5.78 1.40 2.94 3.28 0.09 0.08
S12 38.9 44.6 3.88 1.16 0.39 0.27 0.14 0.28
S21 60.3 51.6 2.91 4.69 0.37 0.01 0.17 0.19
S22 23.1 36.0 1.01 5.61 1.11 0.84 0.07 0.19
NF 51.9 72.8 3.70 3.52 0.34 0.37 0.06 0.12
IIP3 54.6 59.7 5.02 5.93 0.29 0.43 0.33 0.26
Power 16.6 42.5 0.01 1.24 0.92 0.93 0.09 0.02
Comparison on Estimation Error
30Operational Amplifier
- IBM 0.25 µm process
- 49 principal random factors
- MOSFET variations from design kit
- Include mismatches
Performance Linear Quadratic
Gain 3.92 1.57
Offset 21.80 7.49
UGF 1.14 0.45
GM 0.96 0.52
PM 1.11 0.41
SR (P) 0.82 0.66
SR (N) 1.27 0.44
SW (P) 0.38 0.16
SW (N) 0.36 0.12
Power 1.00 0.64
Circuit Schematic for OpAmp
Regression Modeling Error for OpAmp
31Operational Amplifier
- APEX achieve more than 100x speedup compared with
MC 104 runs
Performance Linear Linear Legendre Legendre APEX APEX
Performance 1 99 1 99 1 99
Gain 22.7 10.4 22.0 81.7 1.45 0.32
Offset 11.5 74.7 222 159 0.58 3.20
UGF 3.78 4.30 0.39 0.33 0.03 0.18
GM 2.72 2.46 0.37 0.20 0.08 0.04
PM 4.41 3.79 0.40 0.52 0.13 0.02
SR (P) 0.81 0.97 0.35 0.34 0.11 0.07
SR (N) 3.83 4.31 0.24 0.27 0.13 0.24
SW (P) 0.13 0.03 0.37 0.37 0.16 0.06
SW (N) 0.06 0.03 0.34 0.43 0.09 0.01
Power 0.69 0.65 0.35 0.41 0.11 0.00
Comparison on Estimation Error
32Application of APEX
- APEX can be incorporated into statistical
analysis/synthesis tools - E.g. robust analog design Li 04
Optimization Engine
Unsized Topology
Optimized Circuit Size
Simulation Engine
APEX
Li 04 Robust analog/RF circuit design with
projection-based posynomial modeling, IEEE ICCAD,
2004
33Conclusion
- APEX applies moment matching for PDF/CDF
extraction - Propose a binomial moment evaluation for
computing high order moments - Moments are efficiently matched to a pole/residue
formulation - Solve several implementation issues of APEX
- PDF/CDF shifting using generalized Chebyshev
inequality - Reverse PDF/CDF Evaluation
- APEX can be incorporated into statistical
analysis/synthesis tools - Statistical timing analysis
- Yield optimization