基于Radix-4 Booth编码的12位乘累加运算单元设计  

Design of a 12-bit MAC Operation Unit Based on Radix-4 Booth Encoding

作　　者：吴秀龙[1,2] 王光辰 WU Xiu-long;WANG Guang-chen(Anhui University;Anhui High Performance Integrated Circuit Engineering Research Center)

机构地区：[1]安徽大学 [2]安徽高性能集成电路工程研究中心

出　　处：《中国集成电路》2025年第3期55-62,共8页China lntegrated Circuit

基　　金：国家自然科学基金项目(62274001);安徽省重点研发计划项目(2022a05020044);安徽省高校协同创新项目(GXXT-2023-013);安徽省高校协同创新项目(GXXT-2023-003)。

摘　　要：乘累加(MAC)运算作为卷积神经网络(CNN)中的主体运算,在人工智能(AI)技术等方面得到了大量使用。然而CNN中的MAC运算消耗大量功耗,给硬件设备带来严峻挑战。鉴于该问题,本文提出一种高能效的MAC运算单元以适用于CNN计算。其特点包括通过Radix-4 Booth编码以减少乘法部分积数量,设计了规则化的生成方案对乘法部分积进行约束以简化后续累加过程,在累加阶段使用了基于4-2压缩和3-2压缩的混合加法树结构以提高压缩效率,引入流水结构以提高吞吐量。在0.5 V下,提出的结构能效可以达到15.04 TOPS/W,相比使用行波进位加法器进行累加的MAC结构优化约13.4%。The multiply-accumulate(MAC)operation,as the main operation in convolutional neural networks(CNN),is heavily used in artificial intelligence(AI)techniques and other aspects.However,the MAC operation in CNN consumes a large amount of power and poses a serious challenge to hardware devices.Given the consideration,an energy-efficient MAC operation unit is proposed for CNN computation.Its features include reducing the number of partial products in multiplication through Radix-4 Booth encoding,designing a regularized generation scheme to constrain the multiplication partial products to simplify the subsequent accumulation process,using a hybrid adder tree structure based on 4-2 compression and 3-2 compression in the accumulation stage to improve compression efficiency,and introducing a pipeline structure to enhance throughput.At 0.5 V,the energy efficiency of the proposed structure can reach 15.04 TOPS/W,which is optimized by about 13.4%compared to the MAC using ripple carry adder for accumulation.

关 键 词：乘累加 Radix-4 Booth编码 加法树 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]