Scalable Transformer Network-based Reinforcement Learning Method for PSIJ Optimization in HBM

Abstract

In this paper, we first propose a scalable transformer network-based reinforcement learning (RL) method for power supply induced jitter (PSIJ) optimization in high bandwidth memory (HBM). The proposed method can provide an optimal power distribution network (PDN) decoupling capacitor (decap) design to satisfy the target PSIJ with the minimum number of NMOS decaps. For the given number of decaps, the network is trained to maximize the impedance reduction from 10 MHz to 20 GHz compared to the initial PDN. Also, the network has scalability on the number of decap assignments. Therefore, for given any number of decaps, the scalable network can provide minimized PDN impedance profiles by one inference without re-training. Then, by increasing the decap assignments, the network can find out the minimum number to meet the given target PSIJ. For verification, the proposed network is applied to the HBM2 I/O interface. The network successfully provides the optimized decap designs to satisfy the given target PSIJ values.