Bayesian Exploration Imitation Learning-Based Contextual via Design Optimization Method of PAM-4-Based High-Speed Serial Link

Abstract

This article presents a novel contextual via optimization method, Bayesian exploration imitation learning (BE-IL), designed to efficiently enhance the signal integrity (SI) of PAM-4-based high-speed serial links in a peripheral component interconnect express (PCIe) 6.0. The proposed method optimizes via design parameters by considering the strong correlation between the channel and via parameters. It designs a solver capable of rapidly and efficiently optimizing differential via design parameters, even when channel parameters change. We utilize a deep neural network (DNN) to represent a policy (decision maker) and train it using BE-IL, an innovative learning framework that synthesizes two promising optimization methods: Bayesian optimization (BO) and imitation learning (IL). BE-IL collects high-quality guiding solutions (via parameters) for various contexts (channel parameters) using BO and subsequently trains the DNN policy to imitate these guiding solutions through IL. The context-aware DNN can then efficiently find near-optimal via parameters for any channel (context) without further searches or training processes. We verify the effectiveness of proposed method by comparing with deep reinforcement learning and BO for optimizing via parameters in previously unseen PAM-4-based differential channels within PCIe 6.0 of SSD boards.