ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2019, Vol. 56 ›› Issue (1): 7-22.doi: 10.7544/issn1000-1239.2019.20180693

• 综述 • 上一篇    下一篇

智能芯片的评述和展望

韩栋1,2,周聖元1,2,支天1,陈云霁1,2,陈天石1,3   

  1. 1(中国科学院计算技术研究所智能处理器中心 北京 100190); 2(中国科学院大学 北京 100049); 3(上海寒武纪信息科技有限公司 上海 201203) (handong2014@ict.ac.cn)
  • 出版日期: 2019-01-01
  • 基金资助: 
    国家重点研发计划项目(2017YFA0700902,2017YFB1003101);国家自然科学基金项目(61472396,61432016, 61473275, 61522211, 61532016, 61521092, 61502446, 61672491, 61602441, 61602446,61732002,61702478);国家“九七三”重点基础研究发展计划基金项目(2015CB358800);国家科技重大专项基金项目(2018ZX01031102);中国科学院战略性先导科技专项(B类)(XDB32050200)

A Survey of Artificial Intelligence Chip

Han Dong1,2, Zhou Shengyuan1,2, Zhi Tian1, Chen Yunji1,2,Chen Tianshi1,3   

  1. 1(Intelligent Processor Research Center, Institute of Computing Technology, Chinese Academy of Sciences, Beijing 100190); 2(University of Chinese Academy of Sciences, Beijing 100049); 3(Shanghai Cambricon Information Technology Co., Ltd., Shanghai 201203)
  • Online: 2019-01-01

摘要: 近年来,人工智能技术在许多商业领域获得了广泛应用,并且随着世界各地的科研人员和科研公司的重视和投入,人工智能技术在传统语音识别、图像识别、搜索/推荐引擎等领域证明了其不可取代的价值.但与此同时,人工智能技术的运算量也急剧扩增,给硬件设备的算力提出了巨大的挑战.从人工智能的基础算法以及其应用算法着手,描述了其运算方式及其运算特性.然后,介绍了近期人工智能芯片的发展方向,对目前智能芯片的主要架构进行了介绍和分析.而后,着重介绍了DianNao系列处理器的研究成果.该系列的处理器为智能芯片领域最新最先进的研究成果,其结构和设计分别面向不同的技术特征而提出,包括深度学习算法、大规模的深度学习算法、机器学习算法、用于处理二维图像的深度学习算法以及稀疏深度学习算法等.此外,还提出并设计了完备且高效的Cambricon指令集结构.最后,对人工神经网络技术的发展方向从多个角度进行了分析,包括网络结构、运算特性和硬件器件等,并基于此对未来工作可能的发展方向进行了预估和展望.

关键词: 人工智能, 加速器, FPGA, ASIC, 权重量化, 稀疏剪枝

Abstract: In recent years, artificial intelligence (AI)technologies have been widely used in many commercial fields. With the attention and investment of scientific researchers and research companies around the world, AI technologies have been proved their irreplaceable value in traditional speech recognition, image recognition, search/recommendation engine and other fields. However, at the same time, the amount of computation of AI technologies increases dramatically, which poses a huge challenge to the computing power of hardware equipments. At first, we describe the basic algorithms of AI technologies and their application algorithms in this paper, including their operation modes and operation characteristics. Then, we introduce the development directions of AI chips in recent years, and analyze the main architectures of AI chips. Furthermore, we emphatically introduce the researches of DianNao series processors. This series of processors are the latest and most advanced researches in the field of AI chips. Their architectures and designs are proposed for different technical features, including deep learning algorithms, large-scale deep learning algorithms, machine learning algorithms, deep learning algorithms for processing two-dimensional images and sparse deep learning algorithms. In addition, a complete and efficient instruction architecture(ISA) for deep learning algorithms, Cambricon, is proposed. Finally, we analyze the development directions of artificial neural network technologies from various angles, including network structures, operation characteristics and hardware devices. Based on the above, we predict and prospect the possible development directions of future work.

Key words: artificial intelligence, accelerators, FPGA, ASIC, weight quantization, sparse pruning

中图分类号: