Computer Engineering and Applications ›› 2020, Vol. 56 ›› Issue (21): 65-71.DOI: 10.3778/j.issn.1002-8331.1912-0339

Previous Articles     Next Articles

Research and Implementation of Accelerating Neuromorphic Computing Based on ZYNQ Cluster

ZHANG Xinwei, LI Kang, YU Gongjian, LIU Jiahang, LI Peiqi, CHAI Zhilei   

  1. 1.School of Internet of Things Engineering, Jiangnan University, Wuxi, Jiangsu 214122, China
    2.State Key Laboratory of Mathematical Engineering and Advanced Computing, Wuxi, Jiangsu 214125, China
  • Online:2020-11-01 Published:2020-11-03

基于ZYNQ集群的神经形态计算加速研究与实现

张新伟,李康,郁龚健,刘家航,李佩琦,柴志雷   

  1. 1.江南大学 物联网工程学院 物联网技术应用教育部工程研究中心,江苏 无锡 214122
    2.数学工程与先进计算国家重点实验室,江苏 无锡 214125

Abstract:

Spiking Neural Network(SNN)-based neuromorphic computing is considered to be a better way to solve artificial intelligence problems because of its working mechanism closer to that of the biological brain. However, how to meet the high performance, low power consumption, and adapt to the scaling needs is a challenging problem for neuromorphic computing systems. Based on the FPGA heterogeneous computing platform ZYNQ cluster, on the NEST-like brain simulator, the problems of Spike-Timing Dependent Plasticity(STDP) synaptic computation with high complexity, low parallelism, and large hardware resource occupation are mainly solved. The experimental results show that the designed method has an performance of 14.7 times that of the Xeon E5-2620 CPU on the 8-node ZYNQ 7030 cluster. In terms of energy efficiency, it is 51.6 times that of the Xeon E5-2620 CPU and 20.6 times that of the 8-node ARM Cortex-A9.

Key words: neuromorphic computing, Spiking Neuron Networks(SNN), Spike?Timing Dependent Plasticity(STDP), FPGA cluster, NEST simulator

摘要:

基于脉冲神经网络(SNN)的神经形态计算由于工作机理更接近于生物大脑,被认为有望克服深度学习的不足而成为解决人工智能问题的更佳途径。但是如何满足高性能、低功耗和适应规模伸缩需求是神经形态计算系统需要解决的挑战性问题。基于FPGA异构计算平台ZYNQ集群,在NEST类脑仿真器上,重点解决了具有脉冲时间依赖可塑性(STDP)突触计算复杂度高、并行度低、硬件资源占用大的问题。实验结果表明,设计的方法在8节点ZYNQ 7030集群上,性能是Xeon E5-2620 CPU的14.7倍。能效比方面,是Xeon E5-2620 CPU的51.6倍,是8节点ARM Cortex-A9的20.6倍。

关键词: 神经形态计算, 脉冲神经网络(SNN), 脉冲时间依赖可塑性(STDP), FPGA集群, NEST仿真器