Distributionally Robust Circuit Design Optimization under Variation Shifts

TL;DR

This paper introduces a distributionally robust optimization framework for circuit design that accounts for shifts in process variation distributions, enhancing robustness without needing exact distribution knowledge.

Contribution

It formulates variation-aware circuit optimization as a distributionally robust problem and applies Bayesian optimization to handle distribution shifts effectively.

Findings

Optimized circuits demonstrate robustness against variation shifts.

Method validated on photonic and electronic ICs.

Performance remains stable under diverse process variation distributions.

Abstract

Due to the significant process variations, designers have to optimize the statistical performance distribution of nano-scale IC design in most cases. This problem has been investigated for decades under the formulation of stochastic optimization, which minimizes the expected value of a performance metric while assuming that the distribution of process variation is exactly given. This paper rethinks the variation-aware circuit design optimization from a new perspective. First, we discuss the variation shift problem, which means that the actual density function of process variations almost always differs from the given model and is often unknown. Consequently, we propose to formulate the variation-aware circuit design optimization as a distributionally robust optimization problem, which does not require the exact distribution of process variations. By selecting an appropriate uncertainty set for the probability density function of process variations, we solve the shift-aware circuit optimization problem using distributionally robust Bayesian optimization. This method is validated with both a photonic IC and an electronics IC. Our optimized circuits show excellent robustness against variation shifts: the optimized circuit has excellent performance under many possible distributions of process variations that differ from the given statistical model. This work has the potential to enable a new research direction and inspire subsequent research at different levels of the EDA flow under the setting of variation shift.