Float-Fix: An Efficient and Hardware-Friendly Data Type for Deep Neural Network

doi:10.1007/s10766-018-00626-7

	Float-Fix: An Efficient and Hardware-Friendly Data Type for Deep Neural Network
	Han, Dong 1,2; Zhou, Shengyuan 1,2; Zhi, Tian 1; Wang, Yibo 4; Liu, Shaoli 1,3
	2019-06-01
发表期刊	INTERNATIONAL JOURNAL OF PARALLEL PROGRAMMING
ISSN	0885-7458
卷号	47 期号:3 页码:345-359
摘要	Recent years, as deep learning rose in prominence, neural network accelerators boomed. The existing research shows that both speed and energy-efficiency can be improved by low precision data structure. However, decreasing the precision of data might compromise the usefulness and accuracy of the underlying AI. And the existing studies can not meet all AI application requirements. In the paper, we propose a new data type, called Float-Fix (FF). We introduce the structure of FF and compare it with other data types. In our evaluation, the accuracy loss of 8-bit FF is less than 0.12% on a subset of known neural network models, 7x better than fixed-point, DFX and floating-point on average. We implement the hardware architectures of operators and neural processing unit using 8-bit FF data type with TSMC 65nm Gplus High VT library. The experiments show that the hardware cost of convertors converting between 16-bit fixed-point and FF is really small. And the multiplier of 8-bit FF only needs 1188m2 area, which is nearly 8-bit fixed-point. Comparing with the neural processing unit of DianNao, FF reduces 34.3% area.
关键词	Float-Fix Neural network Hardware accelerator Data type
DOI	10.1007/s10766-018-00626-7
收录类别	SCI
语种	英语
资助项目	National Key Research and Development Program of China[2017YFA0700902] ; National Key Research and Development Program of China[2017YFB1003101] ; NSF of China[6147239] ; NSF of China[61432016] ; NSF of China[61473275] ; NSF of China[61522211] ; NSF of China[61532016] ; NSF of China[61521092] ; NSF of China[61502446] ; NSF of China[61672491] ; NSF of China[61602441] ; NSF of China[61602446] ; NSF of China[61732002] ; NSF of China[61702478] ; 973 Program of China[2015CB358800] ; National Science and Technology Major Project[2018ZX01031102] ; Strategic Priority Research Program of Chinese Academy of Sciences[XDBS01050200]
WOS研究方向	Computer Science
WOS类目	Computer Science, Theory & Methods
WOS记录号	WOS:000471644400002
出版者	SPRINGER/PLENUM PUBLISHERS
引用统计	被引频次：1[WOS] [WOS记录] [WOS相关记录]
文献类型	期刊论文
条目标识符	http://119.78.100.204/handle/2XEOYT63/4180
专题	中国科学院计算技术研究所期刊论文_英文
通讯作者	Liu, Shaoli
作者单位	1.Chinese Acad Sci, Intelligent Processor Res Ctr, Inst Comp Technol, Beijing, Peoples R China 2.Univ Chinese Acad Sci, Beijing, Peoples R China 3.Cambricon Tech Ltd, Beijing, Peoples R China 4.Tsinghua Univ, Beijing, Peoples R China
推荐引用方式 GB/T 7714	Han, Dong,Zhou, Shengyuan,Zhi, Tian,et al. Float-Fix: An Efficient and Hardware-Friendly Data Type for Deep Neural Network[J]. INTERNATIONAL JOURNAL OF PARALLEL PROGRAMMING,2019,47(3):345-359.
APA	Han, Dong,Zhou, Shengyuan,Zhi, Tian,Wang, Yibo,&Liu, Shaoli.(2019).Float-Fix: An Efficient and Hardware-Friendly Data Type for Deep Neural Network.INTERNATIONAL JOURNAL OF PARALLEL PROGRAMMING,47(3),345-359.
MLA	Han, Dong,et al."Float-Fix: An Efficient and Hardware-Friendly Data Type for Deep Neural Network".INTERNATIONAL JOURNAL OF PARALLEL PROGRAMMING 47.3(2019):345-359.