Abstract

As nanoscale IC technology advances, process variation induces a new set of challenges for designing SRAM. This project aims to extend the existing importance sampling method to develop a unified statistical analysis engine of SRAM yield, reliability and testability. In particular, the team adapts the recent advance of Gibbs sampling from the statistics community to find the optimal distribution function for importance sampling with minimum computational cost (i.e., minimum number of simulation runs). This goal is achieved by adaptively searching the failure region using a novel conditional probability formulation. Preliminary experimental results demonstrate that the proposed Gibbs sampling method achieves 3~10× runtime speed-up over other state-of-the-art methods without surrendering any accuracy. In future research, the researchers will further apply the proposed Gibbs sampling technique to two other applications: (1) study the SRAM reliability due to aging effects, and (2) estimate yield loss and escape rate for SRAM testing.