SOFTWARE-HARDWARE CODESIGN FOR EFFICIENT NEURAL NETWORK ACCELERATION
SOFTWARE-HARDWARE CODESIGN FOR EFFICIENT NEURAL NETWORK ACCELERATION
复制标题
DOI：
10.1109/mm.2017.39
复制
发表时间：
2017-03-01
期刊：
影响因子：
3.6
通讯作者：
Yang, Huazhong
中科院分区：
文献类型：
--
作者：
Guo, Kaiyuan;Han, Song;Yang, Huazhong
让深度学习更高效的开发者不能仅仅依靠硬件。诸如模型压缩等创新的固件优化技术可显著节省功率并提高性能。本文概述了DEEPHI的技术流程，包括压缩、编译和硬件加速。两个加速器通过卷积和递归神经网络为客户端和数据中心应用程序实现极高的能效。
DESIGNERS MAKING DEEP LEARNING COMPUTING MORE EFFICIENT CANNOT RELY SOLELY ON HARDWARE. INCORPORATING SOFTWARE-OPTIMIZATION TECHNIQUES SUCH AS MODEL COMPRESSION LEADS TO SIGNIFICANT POWER SAVINGS AND PERFORMANCE IMPROVEMENT. THIS ARTICLE PROVIDES AN OVERVIEW OF DEEPHI'S TECHNOLOGY FLOW, INCLUDING COMPRESSION, COMPILATION, AND HARDWARE ACCELERATION. TWO ACCELERATORS ACHIEVE EXTREMELY HIGH ENERGY EFFICIENCY FOR BOTH CLIENT AND DATACENTER APPLICATIONS WITH CONVOLUTIONAL AND RECURRENT NEURAL NETWORKS.