Deep Reinforcement Learning-based Decoupling Capacitor Optimization Method for Multi-Power Domain considering Transfer Noise in 3D-ICs

Abstract

In this paper, we propose a deep reinforcement learning (DRL)-based multi-power distribution network (PDN) decoupling capacitor design optimization method considering transfer noise in 3D-ICs. The transfer noise from multi-PDN with vertical structures could cause system failure, the entire simultaneous switching noise (SSN) with the combined transfer noise should be considered. To address the multi-PDN problem, we use reinforcement learning suitable for solving complex optimization problems. The input dataset and Markov decision process (MDP) were designed to optimize various multi-PDN cases. The 5x4 size of two PDNs with a vertically stacked structure was used for verification. The proposed method successfully optimizes the decoupling capacitors of multi-PDN. In addition, the proposed method was compared to genetic algorithm (GA), the proposed method perfomed better optimization and reduced the time by about 99% compared to GA to 0.08 seconds.