Deep reinforcement learning(DRL)-based hybrid bonding TSV design optimization method for next generation high bandwidth memory (HBM)

Abstract

In this paper, we propose the deep reinforcement learning-based (DRL) optimization method by defining design parameter optimization of hybrid bonding TSV for next-generation high-bandwidth memory (HBM) as a problem. The proposed method is configured in the form of extracting the action that obtains the best reward for the state using deep reinforcement learning. The agent trained through the proposed method can quickly and accurately provide optimal hybrid bonding TSV design based on the Cu pad dimension, which is an interconnection parameter, considering signal integrity. In the process, a recurrent neural network-based policy network to reflect the coupling between design parameters, a fast and accurate modeling-based reward simulation method for evaluating signal integrity of design parameters, and a clipping policy gradient (CAPG) algorithm for stable learning were proposed. To verify the proposed method, it was applied to the hybrid bonding TSV of high bandwidth memory (HBM) and compared with the conventional optimization method in terms of performance. As a result, the proposed methodology has time efficiency and optimality compared to the conventional optimization method.