Parameter-importance based Monte-Carlo technique for variation-aware analog yield optimization

Abstract

The Monte-Carlo method is the method of choice for accurate yield estimation. Standard Monte-Carlo methods suffer from a huge computational burden even though they are very accurate. Recently a Monte-Carlo method was proposed for the parametric yield estimation of digital integrated circuits [13] that achieves significant computational savings at no loss of accuracy by focusing on those statistical variables that have a significant impact on yield. We adapt this technique to the context of analog circuit yield estimation. The inputs to the proposed method are the designable parameters, the uncontrollable statistical variations, and the operating conditions of interest. The technique of [13] operates on a linear model of circuit variations. In our work we first convexify the nonlinear design constraints to obtain a convex feasible region. We then consider an accurate polytope-approximation of the convex feasible region by taking tangent hyperplanes at various points on the surface of the convex region. The hyperplanes give rise to a matrix of design variable sensitivities, which is then used to glean information about the importance of design variables for yield estimation. Finally the knowledge of which design variables are very important for yield estimation is used to allow the Monte-Carlo technique achieve a lower error compared to standard Monte-Carlo in the same amount of simulation time.