电子学报 ›› 2018, Vol. 46 ›› Issue (8): 1898-1905.DOI: 10.3969/j.issn.0372-2112.2018.08.014

• 学术论文 • 上一篇    下一篇

脉冲神经网络硬件互连系统的动态优先级仲裁策略

刘俊秀1, 黄星月1, 罗玉玲1, 曹弋2   

  1. 1. 广西师范大学电子工程学院, 广西桂林 541004;
    2. 萨里大学萨里商学院 商业转型和企业可持续性发展系, 英国萨里 GU27XH
  • 收稿日期:2017-07-27 修回日期:2017-11-02 出版日期:2018-08-25
    • 通讯作者:
    • 罗玉玲
    • 作者简介:
    • 刘俊秀 男,1984年生于黑龙江绥化.博士、广西师范大学电子工程学院讲师.研究方向为生物神经网络、类脑系统;黄星月 男,1992年生于河南开封.广西师范大学电子工程学院硕士研究生.研究方向为脉冲神经网络大规模互连架构.
    • 基金资助:
    • 国家自然科学基金 (No.61603104,No.61661008); 广西自然科学基金 (No.2016GXNSFCA380017,No.2015GXNSFBA139256); 广西高校引进海外高层次人才百人计划专项经费; 广西高校科研项目 (No.KY2016YB059); 广西多源信息挖掘与安全重点实验室开发基金 (No.MIMS15-07)

Dynamic Priority Arbitration Strategy for Interconnections of Hardware Spiking Neural Networks

LIU Jun-xiu1, HUANG Xing-yue1, LUO Yu-ling1, CAO Yi2   

  1. 1. Faculty of Electronic Engineering, Guangxi Normal University, Guilin, Guangxi 541004, China;
    2. Department of Business Transformation and Sustainable Enterprise, Surrey Business School, University of Surrey, Surrey GU27 XH, UK
  • Received:2017-07-27 Revised:2017-11-02 Online:2018-08-25 Published:2018-08-25
    • Corresponding author:
    • LUO Yu-ling
    • Supported by:
    • National Natural Science Foundation of China (No.61603104, No.61661008); Natural Science Foundation of Guangxi Zhuang Autonomous Region,  China (No.2016GXNSFCA380017, No.2015GXNSFBA139256); Overseas High-level Talents Introduction Project of Colleges and Universities of Guangxi Zhuang Autonomous Region  ("100-Talents Program") Program"); Scientific Research Project of Universities in Guangxi Zhuang Autonomous Reigion (No.KY2016YB059); Guangxi Key Laboratory Foundation for Multi-source Information Mining and Security (No.MIMS15-07)

摘要: 本文基于EMBRACE脉冲神经网络硬件实现方案提出了一种片上网络路由器的动态优先级仲裁策略,来解决脉冲神经网络脉冲传输的交通负载非均衡问题.该方案使用二维网格片上网络系统实现神经元之间的互连通讯,其基于脉冲发送频率的动态优先级仲裁策略能够降低高频路径的平均延迟及系统丢包风险,提高系统工作稳定性.使用Noxim片上网络模拟器搭建实验平台,测试结果表明采用提出的动态优先级仲裁策略较轮询及固定优先级仲裁器,高频路径延迟平均降低32.33%和34.69%,降低幅度最大达到84.86%和86.20%.在90nm CMOS工艺下,提出的路由器架构硬件面积为213,471μm2,具有较好的扩展性.

关键词: 脉冲神经网络, 片上网络, 路由器, 仲裁器, 动态优先级

Abstract: Based on the EMBRACE,i.e.a hardware architecture for the SNN,a dynamic priority arbitration strategy for a Networks-on-Chip (NoC) router is proposed to solve the non-balance traffic load problem of the SNN system.A2DmeshNoC system is used to provide the interconnected communication for the neurons.The dynamic priority arbitration strategy is based on the spiking transmission frequencies and it can reduce the risk of the packet loss and the average delay of the paths with heavy traffics,and improve the reliability of system.The Noxim simulator is used to build the simulation platform.The experimental results show that compared to the round-robin and fixed priority arbitration schemes,the proposed dynamic priority arbitration strategy has an average of 32.33% and 34.69%(maximum 84.86% and 86.20%) decrease for the delays of paths with heavy traffics.Based on the 90nm CMOS technology,the hardware area is 213,471 μ m2,which demonstrates its scalability.

Key words: spiking neural networks, networks-on-chip, router, arbiter, dynamic priority

中图分类号: