异构多芯粒神经网络加速器  

Heterogeneous Multi-Chiplets Neural Network Accelerator

作　　者：朱郭益 马胜[1] 张春元[1] 王波[1] Zhu Guoyi;Ma Sheng;Zhang Chunyuan;Wang Bo(College of Computer,National University of Defense Technology,Changsha 410073)

机构地区：[1]国防科技大学计算机学院,长沙410073

出　　处：《计算机辅助设计与图形学学报》2023年第5期811-818,共8页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(62172430);湖南省自然科学基金(2021JJ10052);国家重点研发计划(2021YFB0300300)。

摘　　要：随着神经网络技术的快速发展,出于安全性等方面考虑,大量边缘计算设备被应用于智能计算领域.首先,设计了可应用于边缘计算的异构多芯粒神经网络加速器其基本结构及部件组成.其次,通过预计算异构芯粒上的计算负载,将计算任务在神经网络通道上进行划分,不断加入新的任务,逐芯粒测试并进行迭代,选取异构芯粒组合以构建神经网络加速器.最后,分别在抽样构造的测试神经网络、MobileNet以及ShuffleNet上使用这种粗粒度优化的方法构建了异构多芯粒神经网络加速器,并测试了其能耗与性能表现.实验结果表明,这种异构的设计方法可以在控制能耗同时,分别取得7.43,2.30和5.60的加速比.With the rapid development of neural network technology,a large number of edge computing devices are used in the field of intelligent computing for security reasons.Firstly,this paper designs the basic structure and components of a heterogeneous multi-chiplets neural network accelerator that can be applied to edge computing.Secondly,we precompute the computational load on the heterogeneous cores,divide the computational tasks on the neural network channels,continuously add new tasks,test and iterate chipet by chiplet,and select the combination of heterogeneous chiplets to build the neural network accelerator.Finally,the heterogeneous multi-chiplets neural network accelerator is constructed on the test neural network,Mobile-Net,and ShuffleNet with this coarse-grained optimization method,and its energy consumption and performance are tested respectively.The experimental results show that this heterogeneous design approach can achieve acceleration ratios of 7.43,2.30,and 5.60,respectively,while controlling the energy consumption.

关 键 词：芯粒 异构 神经网络 加速器 

分 类 号：TP391.72[自动化与计算机技术—计算机应用技术]