专用处理器比较分析  

Comparative study of the domain-specific processors

作　　者：鄢贵海[1] 卢文岩 李晓维[1] 孙凝晖[1] Guihai YAN;Wenyan LU;Xiaowei LI;Ninghui SUN(State Key Laboratory of Computer Architecture,Institute of Computing Technology,Chinese Academy of Sciences,Beijing 100190,China)

机构地区：[1]中国科学院计算技术研究所,计算机体系结构国家重点实验室,北京100190

出　　处：《中国科学：信息科学》2022年第2期358-375,共18页Scientia Sinica(Informationis)

基　　金：国家自然科学基金(批准号:61872336,62002340,62090020);中国科学院2020年度青年创新促进会优秀会员(批准号:Y201923);中国科学院B类战略性先导科技专项(批准号:XDB44030100)资助。

摘　　要：微处理器是现代信息系统的核心基础设施.大数据、人工智能、5G等技术的快速发展催生了数据量的爆发性增长,随之对数据处理能力的需求也急剧增长.专用计算技术被广泛认为是后摩尔时代的计算机体系结构演化的重要方向.专用处理器技术的发展一直伴生着通用处理器的发展,数字信号处理技术甚至早于传统意义上的通用处理器.通用处理器技术的发展,不仅在商业上取得了巨大的成功,很多关键技术也被专用处理器吸收借鉴用于提升专用计算的性能、优化可编程性等.本文主要分析了数字信号处理器(DSP)、图像处理器(GPU)、深度学习处理器(AI芯片)和网络处理器(NPU)的关键技术特征,并进一步对专用计算架构未来发展可能涉及的关键点作出了简要的评述.Microprocessor is the core component in modern information systems.The explosive growth of data volume,fueled by the rapidly evolved technology including big data,artificial intelligence and 5 G,demands an unprecedented high computing capability.It has been believed that domain-specific computing will be one of the major directions of computer architecture in the post-Moore’s Law era.The evolution of domain-specific processors is actually always accompanied by general-purpose processors.Digital signal processor(DSP),one of the typical domain-specific architectures,appeared even earlier than common CPU technology.General-purpose computing,which has been witnessed to achieve huge commercial success,also shares lots of key technologies to boost domain-specific architectures,in terms of performance and programmability.By surveying representative commercial products of DSP,GPU,AI chips and NPU,this paper aims to explore the key architectural features of these domain-specific architectures,and is supposed to shed some light on the future research and development of domain-specific computing technology.

关 键 词：专用处理器 数字信号处理 图像处理 深度学习 网络处理 

分 类 号：TP332[自动化与计算机技术—计算机系统结构]