Self-compensation tensor multiplication unit for adaptive approximate computing in low-power CNN processing  

作　　者：Bo LIU Zilong ZHANG Hao CAI Reyuan ZHANG Zhen WANG Jun YANG 

机构地区：[1]National ASIC System Engineering Center,Southeast University,Nanjing 210096,China [2]Nanjing Prochip Electronic Technology Co.Ltd.,Nanjing 210001,China

出　　处：《Science China(Information Sciences)》2022年第4期283-284,共2页中国科学（信息科学）（英文版）

基　　金：supported by the National Science and Technology Major Project(Grant No.2018ZX01031101-005);National Natural Science Foundation of China(Grant No.61904028).

摘　　要：Dear editor,Approximate multiplication is an emerging circuit design technique for AI-based IoT devices using deep neural networks which can reduce energy consumption with acceptable accuracy loss.Digital signal processing with low power consumption is very important for battery-powered devices,such as real-time speech recognition[1]and cuffless blood pressure monitoring[2].We propose a hybrid self-compensation approximate tensor multiplication unit for low-power convolutional neural network(CNN)processing.The proposed architecture encompasses three advantages:(a)a positive-negative hybrid compensation encoding scheme is utilized in partial products generation and the partial products are reorganized;(b)a self-compensation addition tree structure with a staged configuration method is proposed to process the accumulation of reorganized partial products;(c)the proposed unit is further optimized by using a lower supply voltage in the imprecision parts of the addition tree.

关 键 词：APPROXIMATE power EDITOR 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]