Title: YieldAware Analog Integrated Circuit Optimization using Geostatistics Motivated Performance Modeling
1Yield-Aware Analog Integrated Circuit
Optimization using Geostatistics Motivated
Performance Modeling
Guo Yu and Peng Li
Department of ECE Texas AM University yuguo,
pli_at_neo.tamu.edu
2Motivation
- Analog circuit optimization
- Find best performance point(s) with simulation or
performance model - Simulation-based optimization ? accurate, time
consuming - Model-based optimization ? fast, maybe inaccurate
- Pareto front
- Trade-off of competing performances Stehr et al,
DAC 03 - Esp. useful in hierarchical analog optimization
Zou et al, DAC 06
Pareto front
Corresponding performance area
Map
Design space
Performance space
3Motivation
- Pareo front in large analog circuit optimization
- Hierachical Phase-locked loop (PLL) optimization
Zou et al, DAC 06
Pareto fronts
Pareto fronts for VCO
System level optimization
Map back to transistor size
Phase locked loop
4Motivation
- Robust analog circuit optimization
- Conventional optimization may push performance
into corners - Not robust to process variations ? Low yield
- Consider yield when doing optimization Tiway et
al, DAC 06 - Yield-aware Pareto fronts
- Guarantee to achieve the performance with certain
yield - Relax performance for higher yield ? trade off
P2
80 yield
50 yield
nominal
? 20 yield
5Motivation
- Yield-aware pareto front generation
- Performance distribution with Monte-Carlo
simulation - Select performance according to the required
yield - Computation cost may increase 100X
- An efficient and accurate modeling technique is
needed - Pareto front generation needs more evaluation
points - Yield analysis needs Monte-Carlo simulation
- Propose to use Kriging performance mode
- Self-evaluation of confidence level ? increase
accuracy - Suitable for fast Monte-Carlo simulation
6Geostatiscs Motivated Kriging Modeling
- Modeling stochastics process in geography with
Kriging model Matheron, Economic Geology 63 - Use measured data to fit geology model
- Can be applied for deterministic performance
modeling - Complex geology modeling ? good accuracy
- Stochastics ? error/uncertainty prediction
Modeling error too large?
Include measurement results of these points into
the Kriging model
7Performance Modeling with Kriging Model
- Key benefits of Kriging model
- Good accuracy with limited training data
- Uncertainty level prediction for iterative search
Resample in design space
Kriging model
If predicted err gt errmax Include this point as
training data
Design parameter space
8Outline
- Motivations
- Background
- Nominal analog circuit optimization
- Construct Kriging models with design parameters
- Build pareto fronts in nominal case
- Yield-aware analog circuit optimization
- Include process variations into Kriging models
- Partial Kriging model for efficient Monte-Carlo
simulation - Generate yield-aware pareto fronts with updated
Kriging models - Experimental results
- Conclusions
9Kriging Model Fundamental
- Deterministic function approximation
Model detailed nonlinearities
Global trend
Compensation
Input
Output
Model training data
New point
Correlation of Z
Correlation parameter E.g. P2 ? distance based
Correlation parameter Fine correlation tuning
10Kriging Model Fundamental (cont.)
- Performance prediction in new point
- Uncertainty prediction in new point
Distance of new point and training data
New point
Model parameter Uncertainty level prediction
11Kriging Model Generation
- Model construction with maximum likelihood
estimation - Deterministic problem ? optimization problem
MLE of model para.
Prediction of
Optimization problem
Maximize the MLE
MSE
Model para.
Perf. prediction
12Nominal Pareto Front Generation
Uniformly sample in design space
Find corresponding design parameters
Pareto front after 1st run
Search for best performances
13Nominal Pareto Front Generation (cont.)
- Iterative search to push to the boundaries
Sample more points near init. pareto fronts
Pareto front after 1st run
Regenerate pareto front
Not converged? Search again
Converged
Final pareto front in nominal case
14Kriging Models for Yield-aware Optimization
- Regenerate Kriging model with process variation
information - Add process variables as inputs in Kriging model
- Kriging model generation issues
- Model dimension increase with process variations
- E.g. 7 design variables, 23 process variables for
ring VCO - Computation cost raises for yield analysis ?
partial Kriging model
15Yield Evaluation with Kriging Model
- Replace nominal perf. with the perf. at certain
yield - Guarantee to achieve the performance at that
yield level - Evaluate performance distribution with Kriging
model
Monte-Carlo sim. for design point
50
20
80
Use performance at specified yield level for
Pareto front generation
Design space
16Partial Kriging Model
- Save computation cost by reusing pre-calculated
terms
Training data, evaluate once
Recalculate with new input
Same set of process dis. for all the design points
Process variables Can be reused!!
Design parameters Need to be modified in the
optimization
17Yield-aware Pareto Front Generation
Search in the design space
Generate yield pareto fronts
Not converged? Search again
Initial pareto fronts
Converged
Yield-aware pareto fronts
18Experimental Results
- Five-stage ring oscillator
- Design parameter transistor sizes
- Process variations transistor threshold voltages
- Performance power, VCO gain, max. frequency
Kriging model accuracy verification
Ring oscillator schematic
19Experimental Results
- Yield-aware pareto front for ring oscillator
20Experimental Results
- Verification of pareto front
- Sample in the design space for verification
Verification samples
50 yield pareto front
21Experimental Results
- LC oscillator
- Design parameter transistor sizes, inductance,
biasing current - Process variations transistor threshold
voltages, inductor mismatch, parasitic capacitors
and resistors - Performance power and VCO gain
Kriging model accuracy verification
LC oscillator schematic
22Experimental Results
- Yield-aware pareto fronts for LC oscillator
80 yield pareto front
50 yield pareto front
Nominal pareto front
23Experimental Results
- Two-stage operational amplifier
- Design parameter transistor sizes
- Process variation transistor threshold voltages
- Performance DC gain, 3-dB bandwidth
Two-stage Opamp schematic
24Experimental Results
- Iterative pareto front generation
Initial pareto front
Sort for pareto front
Converged pareto front
25Experimental Results
- Yield-aware pareto fronts for operational
amplifier
80 yield pareto front
50 yield pareto front
20 yield pareto front
26Experimental Results
- Run-time cost for optimization
- Kriging model sampling with SPICE optimization
for model generation - Pareto front iterative search with Kriging
models
Run-time for pareto front generation
27Conclusion
- Apply Kriging model for performance modeling
- Connect performance with design parameters and
process variables - Update model with uncertainty prediction
- Nominal pareto front generation
- Sample in the design space with Kriging model
- Iterative search to build pareto fronts
- Yield-aware pareto front generation
- Use nominal pareto fronts as staring points
- Model both design parameters and process
variables - Partial MC simulation for efficient yield
calculation
28Thank You!
29Kriging Performance Model Generation
N
Y