TDR Optimization Method of 112G Serdes Interface in PKG Using Deep Reinforment Learing

This paper presents a deep reinforcement learning (DRL)-based optimization methodology for TDR optimization of high-speed serdes interface channels in high performance computing (HPC) PKG. For optimization, a transformer and a DNN-based prediction operator were built and modeled. The prediction operator provided accurate information to the RL-based agent, and the AI agent is designed to provide general and scalable operations based on PPO. The method proposed in this paper is not only applicable to serdes in PKG but also to board or system-level designs, and it has high value as it can be applied to various high-speed interface tasks.