大规模脉冲神经网络动态加载仿真方法  

Dynamic Loading Simulation Method for Large-scale Spiking Neural Network

作　　者：沈嘉玮 才大业 杨国青[1] 吕攀 李红[1] Shen Jiawei;Cai Daye;Yang Guoqing;LüPan;Li Hong(College of Computer Science and Technology,Zhejiang University,Hangzhou 310013,China)

机构地区：[1]浙江大学计算机科学与技术学院,浙江杭州310013

出　　处：《系统仿真学报》2025年第2期541-550,共10页Journal of System Simulation

基　　金：异构融合类脑计算研究平台(2021ZD0200300)。

摘　　要：针对大规模脉冲神经网络仿真时存在GPU内存需求高的问题,提出一种针对大规模脉冲神经网络的动态加载仿真方法。通过子网络粒度的数据移动,利用主机内存作为更大的内存池,减少GPU显存对于模型仿真规模的限制,实现在单GPU的计算机进行大规模脉冲神经网络仿真,并使用流水线加速技术减少数据移动对仿真速度的影响。最终实现了在单机GPU的实验环境下仿真百万级别神经元规模的仿真,解决了在脉冲神经网络仿真过程中内存不足的问题。To address the problem of high GPU memory requirements in large-scale spiking neural network simulation,a dynamic loading simulation method for large-scale spiking neural networks is proposed.This method uses data movement at the sub-network granularity and utilizes the host memory as a larger memory pool to reduce the limitation of GPU memory on the model simulation scale,enabling large-scale spiking neural network simulation on a single GPU computer:The pipeline acceleration technique is adopted to reduce the impact of data movement on simulation speed.The simulation of a million-scale neural network is achieved in a single GPU experimental environment,which solves the problem of insufficient memory during spiking neural network simulation.

关 键 词：类脑计算 脉冲神经网络 神经元 突触 仿真 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]