近似计算新范式在深度学习加速系统中的应用及研究进展  

Application and Research Progress of Approximate Compu

作　　者：龚宇 王丽萍[1,2] 王佑 刘伟强 GONG Yu;WANG Liping;WANG You;LIU Weiqiang(College of Electronic and Information Engineering/College of Integrated Circuits,Nanjing University of Aeronautics and Astronautics,Nanjing 211106,China;Key Laboratory of Aerospace Integrated Circuits and Microsystem,Ministry of Industry and Information Technology,Nanjing 211106,China)

机构地区：[1]南京航空航天大学电子信息工程学院/集成电路学院,南京211106 [2]空天集成电路与微系统工信部重点实验室,南京211106

出　　处：《电子与信息学报》2023年第9期3098-3108,共11页Journal of Electronics & Information Technology

基　　金：国家自然科学基金(62022041);中央高校基本科研业务费(NJ2023020)。

摘　　要：深度学习已经成为当前人工智能技术中最为重要的算法之一。随着应用场景不断扩展,深度学习硬件规模越来越大,计算复杂度呈现数量级提升趋势,对加速系统提出了极高能效需求。后摩尔时代,新型计算范式逐渐取代工艺微缩成为提升能效的有效方案,近似计算以牺牲部分精度的代价换取大幅能效提升,成为最具前景的设计方法之一。该文以深度学习加速系统的不同设计层次为切入,首先介绍了深度学习网络模型的算法特征,围绕算法层的近似计算方案介绍了量化方法的研究进展;其次,围绕硬件架构和电路层调研了当前深度学习加速在图像、语音等多个方向采用的近似电路和架构方案,围绕层次化的设计方法调研了当前近似计算的系统设计方法学及EDA领域的关键问题和研究进展;最后,对该领域方向进行展望,旨在推动近似计算新范式在深度学习加速系统中的应用。Deep learning has emerged as one of the most important algorithms in artificial intelligence.With the increasing application scenarios,the hardware scales for deep learning are becoming larger,and the computational complexity has considerably increased,leading to a high demand for energy efficiency in accelerating systems.In the post-Moore’s Law era,new computing paradigms are gradually replacing process scaling as an effective solution for improving energy efficiency.One of the most promising design paradigms is approximate computing,which sacrifices some precision to improve energy efficiency.This research focuses on different design layers of deep learning acceleration systems.First,the algorithm characteristics of deep learning network models are introduced,and the research progress on quantization methods is presented in view of the approximate computing scheme at the algorithm layer.Second,approximate circuits and architectures employed in various directions such as image and speech recognition in the circuit-architecture layer are surveyed.Furthermore,the current hierarchical design methods for approximate computing as well as critical issues and research progress in Electronic Design Automation(EDA)are investigated.Finally,the future direction of this field is anticipated to promote the application of a new paradigm of approximate computing in deep learning acceleration systems.

关 键 词：近似计算 计算范式 深度学习 加速系统 研究进展 

分 类 号：TN402[电子电信—微电子学与固体电子学] TP183[自动化与计算机技术—控制理论与控制工程]