ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2019, Vol. 56 ›› Issue (2): 240-253.doi: 10.7544/issn1000-1239.2019.20170852

• 综述 • 上一篇    下一篇

硬件加速神经网络综述

陈桂林,马胜,郭阳   

  1. (国防科技大学计算机学院 长沙 410073) (cglnudt@163.com)
  • 出版日期: 2019-02-01
  • 基金资助: 
    国家自然科学基金项目(61672526);国防科技大学科研计划项目(ZK17-03-06)

Survey on Accelerating Neural Network with Hardware

Chen Guilin, Ma Sheng, Guo Yang   

  1. (College of Computer ,National University of Defense Technology, Changsha 410073)
  • Online: 2019-02-01

摘要: 人工神经网络目前广泛应用于人工智能的应用当中,如语音助手、图像识别和自然语言处理等.随着神经网络愈加复杂,计算量也急剧上升,传统的通用芯片在处理复杂神经网络时受到了带宽和能耗的限制,人们开始改进通用芯片的结构以支持神经网络的有效处理.此外,研发专用加速芯片也成为另一条加速神经网络处理的途径.与通用芯片相比,它能耗更低,性能更高.通过介绍目前通用芯片和专用芯片对神经网络所作的支持,了解最新神经网络硬件加速平台设计的创新点和突破口.具体来说,主要概述了神经网络的发展,讨论各类通用芯片为支持神经网络所作的改进,其中包括支持低精度运算和增加一个加速神经网络处理的计算模块.然后从运算结构和存储结构的角度出发,归纳专用芯片在体系结构上所作的定制设计,另外根据神经网络中各类数据的重用总结了各个神经网络加速器所采用的数据流.最后通过对已有加速芯片的优缺点分析,给出了神经网络加速器未来的设计趋势和挑战.

关键词: 机器学习, 神经网络, 通用芯片, 专用加速芯片, 体系结构

Abstract: Artificial neural networks are widely used in artificial intelligence applications such as voice assistant, image recognition and natural language processing. With the rise of complexity of the application, the computational complexity has also increased dramatically. The traditional general-purpose processor is limited by the memory bandwidth and energy consumption when dealing with the complex neural network. People began to improve the architecture of the general-purpose processors to support the efficient processing of the neural network. In addition, the development of special-purpose accelerators becomes another way to accelerate processing of neural network. Compared with the general-purpose processor, it has lower energy consumption and higher performance. The article aims to introduce the designs from current general-purpose processors and special-purpose accelerators for supporting the neural network. It also summarizes the latest design innovation and breakthrough of the neural network acceleration platforms. In particular, the article provides an overview of the neural network and discusses the improvements made by various general-purpose chips to support neural networks, which include supporting low-precision operations and adding a calculation module to speed up neural network processing. Then from the viewpoint of the computational structure and storage structure, the article summarizes the customized designs of special-purpose accelerators, and describes the dataflow used by the neural network chips based on the reuse of various types of the data in the neural network. Through analyzing the advantages and disadvantages of these solutions, the article puts forward the future design trend and challenge of the neural network accelerator.

Key words: machine learning, neural network, general-purpose processor, special-purpose accelerator, architecture

中图分类号: