GridVAE: Fast Power Grid EM-Aware IR Drop Prediction and Fixing Accelerated by Variational AutoEncoder

Yibo Liu and Sheldon Tan
University of Califronia, Riverside


Electromigration (EM) remains the primary failure mechanism for copper-based interconnects in today's and future nanometer chip technologies. To ensure the longevity of on-chip power grids, effective EM-aware IR drop analysis is crucial. However, the existing power grid optimization approaches suffer high computational costs from the full-chip EM-aware IR drop analysis and sensitivity computation. This paper proposes a novel and efficient framework for full-chip power grid EM-aware IR drop prediction and fixing framework. We developed a conditional VAE-based framework, named GridVAE, for fast and accurate EM-aware IR drop prediction and full-chip power grid fixing. Compared to the state-of-the-art generative adversarial network (GAN)-based methods, our GridVAE model offers a remarkable 40% reduction in prediction RMSE on synthesized power grid benchmarks from ARM Cortex-M0 processor design. Building on the accurate EM-aware IR drop predictions and fast acquired sensitivities, we apply the sequence of linear programming-based optimizations to efficiently size the wires. Our proposed GridVAE method achieves up to an 140X speedup (at least one order of magnitude) compared to conventional SLP-based methods for power grid EM-aware IR drop fixing.