Encoder-Decoder Networks for Analyzing Thermal and Power Delivery Networks

Power delivery network (PDN) analysis and thermal are computationally expensive tasks that essential for successful integrated circuit (IC) design. Algorithmically, both these analyses have similar computational structure complexity as they involve the solution to a partial differential equation of same form. This article converts into image-to-image sequence-to-sequence translation tasks, which allows leveraging class machine learning models with an encoder-decoder–based generative (EDGe) architecture address time-intensive nature tasks. For PDN analysis, we propose two networks: (i) IREDGe: full-chip static dynamic IR drop predictor (ii) EMEDGe: electromigration (EM) hotspot classifier based on input power, power grid distribution, pad distribution patterns. ThermEDGe, temperature estimator patterns analysis. These networks transferable across designs synthesized within technology packing solution. The predict on-chip drop, EM locations, in milliseconds negligibly small errors against commercial tools requiring several hours.