HFGCN: High-speed and Fully-optimized GCN Accelerator

MinSeok Han1, Jiwan Kim1, Donggeon Kim1, Hyunuk Jeong1, Gilho Jung1, Myeongwon Oh1, Hyundong Lee2, Yunjeong Go2, HyunWoo Kim1, Jongbeom Kim1, Taigon Song1
1Kyungpook National University (KNU), 2Kyungpook National University


Abstract

graph convolutional network (GCN) is a type of neural network that inference the new nodes based on the connectivity of the graphs. GCN requires high-calculation volume for processing, similar to other neural networks requiring significant calculation. In this paper, we propose a new hardware architecture for GCN that tackles the problem of wasted cycles during processing. We propose a new scheduler module that reduces memory access through aggregation and an optimized systolic array with improved delay. We compare our study with the state-of-the-art GCN accelerator and show outperforming results.