Effcient approaches to overcome non-convexity issues in analog design automation

Abstract

We propose two approaches to overcome limitation of convex programming technique for analog design automation. Analog design performance constraints are cast in posynomial inequality format for suitability into convex optimization based application. But in most cases original equations are not in posynomial so, either they are deliberately modeled or approximated. This leads to inaccuracy. Our first approach is based on exploiting apparent benefit of convex programming based global optimizer but still making use of highly accurate non-posynomial or signomial model. To achieve that, we combine both global and local optimizer. Global optimizer handles less accurate posynomial equation ensuring global optimality. And the initial guess obtained therefore is used by local optimizer to handle accurate signomial model. Second approach demonstrates that the targeted region to be modeled for circuit sizing can be reduced which invariably leads to better model accuracy. We develop low dropout regulator (LDO) performance metrics in posynomial and signomial format for use in the proposed methodology.