Predictable Coupling Effect Model for Global Placement Using Generative Adversarial Networks With an Ordinary Differential Equation Solver

Abstract

One of the most important issues in physical design is coupling capacitance. However, the issue is typically addressed during the routing stage, which necessitates the execution of a time-consuming algorithm. Based on the generative adversarial networks (GAN) model, we propose a coupling-free global placement (CFGP) model with different orders of ordinary differential equations (ODE) solver. Experiments on the ISPD’11/DAC’12 contest benchmark revealed that using the ODE-GAN architecture, our coupling effect estimator (CEE) model can achieve 0.91X similarity to the ground-truth image and a 50X speedup over traditional global routers such as NCTUgr. Compared to the original framework without the CEE model, the CFGP implemented using DREAMPlace results in a 41% reduction in coupling effect.