Adaptive Low-Rank Tensor Approximation for SRAM Yield Analysis using Bootstrap Resampling

With increasing design complexity and reliability requirement, performance failure has become a main threat for a robust circuit design. The major challenge to develop a high-sigma yield analysis method is “Curse of Dimensionality”. In this paper, we propose a novel Adaptive Low-Rank Tensor Approximation (Adaptive LRTA) method. It starts with a Latin Hypercube Sampling set to build the initial LRTA meta-model. Our adaptive framework has multiple iterations of meta-model refinement. In each iteration, we evaluate the confidence interval by constructing a set of bootstrap resampled LRTA meta-models, and develop a learning function which determines our experimental design enrichment criteria. The experiment results on high-dimensional SRAM column validate that our method has better efficiency and higher accuracy. Our adaptive LRTA uses 2-6X fewer samples than state-of-the-art methods without compromising accuracy.