A Quantile-Based SUM/MAX Operating Method for Statistical Timing Analysis

Instead of expressing the delay distribution shape with complicated forms, we propose a quantile-based statistical SUM/MAX operating method by converting the distribution to several quantiles. Based on the numerical convolution, the SUM/MAX operations are converted to DNN with two groups of quantiles. The configuration of DNN models is discussed in this work, such as the hidden layer number and the transfer function. With 28 nm technology, the average errors of the DNN models for estimating the μ+3σ arrival time are less than 3% compared to the Monte Carlo simulations. The proposed method is more efficient than the NN-SSTA which needs to fit the shape of the whole distribution.