Graph convolutional network-based reinforcement learning for large-scale pin assignment optimization of micro-bumps in high bandwidth memory considering signal integrity

Abstract

Recently, to meet the increasing demand of high bandwidth systems, the number of I/Os and interconnections for 2.5D/3D ICs is also growing. Accordingly, the pin count of the ball grid array (BGA) is getting larger along with signal integrity issues. In this paper, we propose BGA-opt, a novel deep reinforcement learning (DRL)-based pin assignment method that represents ball grid array (BGA) packages on graphs for minimizing signal integrity degradation. The proposed method represents the pin arrangement of BGAs in graphs to formulate the pin assignment task to a variant of the maximum independent set (MIS). Then, a state-of-the-art DRL-based MIS solver was introduced to solve our task. Unlike previous methods of BGA optimization, the proposed graph representation of pins makes it possible to assign pins of any shape. Moreover, the significant scaling performance enables us to handle BGA with high pin count. We verify that the proposed DRL-based method with graph representation is effective by comparing it with conventional meta-heuristic methods including genetic algorithms.