Large-Scale Analog/RE Performance Modeling by Statistical Regression

In this paper, we introduce several large-scale modeling techniques to analyze the high-dimensional, strongly-nonlinear performance variability observed in nanoscale manufacturing technologies. Our goal is to solve a large number of (e.g., 10(4)similar to 10(6)) model coefficients from a small set of (e.g., 10(2)similar to 10(3)) sampling points without over-fitting. This is facilitated by exploiting the underlying sparsity of model coefficients. Our circuit example designed in a commercial 65nm process demonstrates that the proposed techniques achieve 25x speedup compared with the traditional response surface modeling(1).