IR-GCN:二值图卷积神经网络推理加速器  

IR-GCN:binary graph convolutional neural network inference accelerator

作　　者：于启航 文渊博 杜子东 YU Qihang;WEN Yuanbo;DU Zidong(State Key Laboratory of Processors,Institute of Computing Technology,Chinese Academy of Sciences,Beijing 100190;University of Chinese Academy of Sciences,Beijing 100049)

机构地区：[1]中国科学院计算技术研究所处理器芯片国家重点实验室,北京100190 [2]中国科学院大学,北京100049

出　　处：《高技术通讯》2024年第10期1024-1035,共12页Chinese High Technology Letters

基　　金：国家重点研发计划(2022YFB4501601);国家自然科学基金(62222214,61925208,U19B2019,U22A2028);中国科学院稳定支持基础研究领域青年国家团队计划(YSBR-029)资助项目。

摘　　要：针对图卷积神经网络(GCN)中数据规模庞大、不适合边缘端低功耗处理器高效推理计算的问题,本文提出一种将新型二值数据量化算法(IR-Net)应用于GCN模型推理计算的方法,并设计了对应的硬件加速器IR-GCN。同时,针对计算过程中工作负载分布不均衡的问题,实现了一种负载均衡模块设计,显著提高了计算效率。实验结果表明,在较小精度损失范围内,IR-GCN加速器可以同时降低计算延迟以及访存开销。与现有性能最优的研究相比,IR-GCN加速器平均具有2.4倍的计算加速比、7.9倍的功耗降低、13.7倍的芯片面积减少以及21.0倍的访存量降低。This article proposes a method of applying a novel binary quantization algorithm IR-Net to the inference calculation of graph convolutional neural networks(GCN)and designing the corresponding hardware accelerator IR-GCN to address the problem of the large-scale data in GCN,which is not suitable for efficient inference calculation on low-power edging processors.Additionally,to address the issue of uneven workload distribution during the computation process,a load balancing module is implemented,which significantly improves the calculation efficiency.The experimental results show that the IR-GCN algorithm can reduce both calculation delay and memory access overhead simultaneously within a small range of precision loss.Compared with the existing state-of-art research,the IR-GCN accelerator has an average of 2.4 times calculation acceleration,7.9 times power reduction,13.7 times chip area reduction,and 21.0 times memory access reduction.

关 键 词：图卷积神经网络(GCN) 二值神经网络(BNN) 硬件加速器 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]