低功耗人工智能计算系统研究进展综述  

A review of progresses in low-power artificial intelligence computing systems

作　　者：陈华[1,2] 曲益明 吴文豪 赵毅 CHEN Hua;QU Yiming;WU Wenhao;ZHAO Yi(School of Integrated Circuits,East China Normal University,Shanghai 200241,China;China Nanhu Academy of Electronics and Information Technology,Jiaxing 314001,China;College of Information Science and Electronic Engineering,Zhejiang University,Hangzhou 310027,China)

机构地区：[1]华东师范大学集成电路科学与工程学院,上海200241 [2]中国电子科技南湖研究院,嘉兴314001 [3]浙江大学信息与电子工程学院,杭州310027

出　　处：《功能材料与器件学报》2024年第6期300-309,共10页Journal of Functional Materials and Devices

基　　金：科技创新2030-“新一代人工智能”重大项目(No.2020AAA0109001);国家自然科学基金资助项目(No.U23B2040)。

摘　　要：最近,随着大数据和硬件能力的快速增长,人工智能取得了显著发展,人工神经网络(Artificial Neural Network,ANN)已被成功应用于解决学术界和工业界的许多问题。然而,在边缘设备上部署人工神经网络仍具有挑战性。这些场景一般对功率、体积有严格的限制,同时对系统延迟和实时性有较高要求,因此构建低功耗人工智能计算系统需要在性能、功率、体积之间进行权衡。本文综述了目前低功耗人工智能计算系统的研究现状,介绍和分析了低功耗人工智能计算硬件和软件工具,阐述了存在的技术挑战,讨论了系统的评估方法和指标,并展望了未来发展趋势。Recently,with the rapid growth of big data and hardware capabilities,artificial intelligence(AI)has achieved significant development.Artificial neural networks(ANNs)have been successfully applied to solve numerous problems in academia and industry.However,deploying AI networks on edge devices remains challenging.These scenarios generally have strict limitations on power and size,while also have high requirements for system latency and real-time performance.Thus,building low-power AI computing systems involves making trade-offs among performance,power,and size.This article reviews the current state of low-power AI computing systems,introduces low-power AI computing hardware and software tools,elaborates on existing technical challenges,discusses system evaluation methods and metrics,and looks to future development trends.

关 键 词：嵌入式人工智能 边缘人工智能 低功耗人工智能 深度学习 神经形态 

分 类 号：TP368[自动化与计算机技术—计算机系统结构]