高技术通讯2024，Vol.34Issue(7)：714-725,12.DOI:10.3772/j.issn.1002-0470.2024.07.005
针对深度学习中不规则内存访问的高吞吐内存管理单元
HTMMU:a memory management unit for irregular memory access in deep learning
摘要
Abstract
The diversification and complexity of artificial intelligence applications lead to irregular memory access pat-tern.The irregular memory access pattern can be defined as bursty and sparse memory access requests,which brings great challenges to the deployment of intelligent applications on mobile devices with strictly limited memory resources.This irregular memory access pattern has caused the memory management unit(MMU)in existing archi-tectures to face the problems of low throughput and long latency,making it a bottleneck of the system.To solve this problem,this paper proposes a novel MMU architecture called high-throughput MMU(HTMMU).HTMMU uses multi-stream parallelism,enhances filtering of redundant requests and allocates limited on-chip memory more rea-sonably to improve system memory access efficiency.Experimental results show that when dealing with the irregular memory accesses in artificial intelligence algorithms,compared with the current MMU design,HTMMU achieves 2.43 times speedup averagely,and reduces the average latency by 65.9％with less than 3.0％area overhead.关键词
内存管理单元(MMU)/地址转换/不规则访存/深度学习/高吞吐Key words
memory management unit(MMU)/address translation/irregular memory access/deep learning/high-throughput引用本文复制引用
丁峰,李曦..针对深度学习中不规则内存访问的高吞吐内存管理单元[J].高技术通讯,2024,34(7):714-725,12.基金项目
国家自然科学基金(U20A20227,U22A2028),中国科学院稳定支持基础研究领域青年团队计划(YSBR-029)和中国科学院青年创新促进会资助项目. （U20A20227,U22A2028）
