Generative Multi-Physics Models for System Power and Thermal Analysis Using Conditional Generative Adversarial Networks

As system performance increases, chip density and power consumption also increase. Power integrity and thermal management have become critical to the design flow and are codependent on each other. Advanced simulation tools perform co-simulation of electrical and thermal analysis on package-board designs. This paper describes a novel way of using a class of deep learning algorithms called conditional GANs (cGANs) to efficiently model the power/thermal co-simulation task. As the name suggests, cGANs are generative models that can predict unseen simulation conditions. Using the cGAN, the root-mean-squared error on unseen test cases is 0.015 in a [-1,1] range, translating to an error under 0.3 C degrees. Furthermore, a trained network exhibits fast inference speeds, allowing for near real-time generation of analysis results. This is a common goal of digital twins for dynamic system performance tuning.