Multiple-Precision Floating-Point Dot Product Unit for Efficient Convolution Computation

Li, Kai1; Mao, Wei1; Xie, Xinang1; Cheng, Quan1; Xie, Huan2; Dong, Zhenjiang2; Yu, Hao1

10.1109/AICAS51828.2021.9458534

2021

IEEE 3rd International Conference on Artificial Intelligence Circuits and Systems (AICAS)

978-1-6654-3025-8

2021 IEEE 3RD INTERNATIONAL CONFERENCE ON ARTIFICIAL INTELLIGENCE CIRCUITS AND SYSTEMS (AICAS)

1-4

JUN 06-09, 2021

null,null,ELECTR NETWORK

345 E 47TH ST, NEW YORK, NY 10017 USA

IEEE

In the computational process of the convolutional neural network (CNN) and high-performance computing, the convolution operation dominates the hardware costs and performance of the overall system. To improve the hardware efficiency, the dot-product units (DPU) have been used for convolution computation. Besides, multiple-precision floating-point (FP) support is essential for the accuracy requirement of various applications. In this work, multiple-precision FP many-term DPU is designed with single instruction multiple data (SIMD) structure. The proposed design also supports multiple-precision operations with configurable multiplier and alignment shifter. FP16 twenty-term, FP32 five-term or FP64 one-term dot product operations can be executed in two successive clock cycles without idle multiplication resources. To speed up the summation process, a carry-select adder (CSLA) is designed with excellent area-delay product (ADP) and power-delay product (PDP) performance. The proposed design is realized in UMC 55-nm process with experiment results. Compared with the state-of-the-art multiple-precision work, the proposed design achieves maximum 3.76 times power performance improvement for FP16 operations. Compared with previous CSLA designs, the proposed work improves ADP and PDP performance by 4.7% and 3.91%, respectively.

multiple-precision floating-point radix-4 booth multiplier CSLA dot product

第一 ; 通讯

英语

CPCI-S ; EI

Science and Technology Innovation Committee Foundation of Shenzhen[JCYJ20180504170454381]

Computer Science ; Engineering

Computer Science, Artificial Intelligence ; Computer Science, Hardware & Architecture ; Engineering, Electrical & Electronic

WOS:000722241000061

20213410817462

Convolution ; Convolutional neural networks

Information Theory and Signal Processing:716.1 ; Computer Theory, Includes Formal Logic, Automata Theory, Switching Theory, Programming Theory:721.1

Web of Science

1.Southern Univ Sci & Technol, Sch Microelect, Shenzhen, Peoples R China
2.Huawei Technol Co Ltd, Shenzhen, Peoples R China

Li, Kai,Mao, Wei,Xie, Xinang,et al. Multiple-Precision Floating-Point Dot Product Unit for Efficient Convolution Computation[C]. 345 E 47TH ST, NEW YORK, NY 10017 USA:IEEE,2021:1-4.