Fast IR drop analysis using U-Net convolutional network

Abstract

As the technology node keeps shrinking, IR drop has become a critical issue of VLSI system design. IR drop impairs circuit performance and may lead to timing issues such as setup/hold violations. Conventional IR drop analyzers solve a large linear equation to attain static/dynamic IR drop, but it takes substantial runtime when the design has millions of nodes. In this paper, we present a machine-learning model that uses image-to-image translation to predict static/dynamic IR drop rapidly. In contrast to most of previous approaches that applied machine learning to calculate IR drop, our model can predict IR drop of designs with different PDN structures. In addition, we predict IR drop region-by-region rather than cell-by-cell to reduce inference time. Compared to commercial IR drop analysis tool (ANSYS Redhawk), our model reduced runtime by 4-5x while achieving an error rate below 18%.