一种基于闪存的温度补偿型向量矩阵乘法器  

A Temperature-Compensated Vector-Matrix Multiplier Based on Flash Memory

作　　者：蒋明峰 方毅[2] 黄鲁[1,2] JIANG Mingfeng;FANG Yi;HUANG Lu(Department of Electronic Science and Technology,Univ,of Science and Technology of China.Hefei 230027,P.R.China;Information Science and Practice Center,Univ,of Science and Technology of China,Hefei 230027,P.R.China)

机构地区：[1]中国科学技术大学电子科学与技术系,合肥230027 [2]中国科学技术大学信息科学实验中心,合肥230027

出　　处：《微电子学》2020年第3期344-348,共5页Microelectronics

基　　金：国家自然科学基金资助项目(61874102)。

摘　　要：基于XMC 65 nm浮栅工艺,设计了一种基于闪存的存算一体向量矩阵乘法器。进行模拟计算的闪存单元工作在线性区,采用两个闪存单元以差分的方式存储一个权重,提高了计算线性度,减小了权重的温度系数分布。设计了电流减法与缓冲电路,精确得到了乘累加的计算结果,并通过负阻并联型闪存负载实现了良好的温度补偿。模拟计算的精度为4位。仿真结果表明,各权重的计算结果为:在-40℃~85℃范围内,温度系数小于0.1%/℃;在85℃~130℃范围内,温度系数小于0.25%/℃。Based on XMC FG 65 nm process,a vector-matrix multiplier(VMM)using flash with processing in memory(PIM)architecture was designed.The flash cells worked in the linear region.By storing one weight in two flash cells differentially,the variation of different weight’s temperature coefficient(TC)could be reduced while increasing the calculation linearity.The current subtractor and buffer were designed to obtain the multiply accumulate(MAC)results accurately,and good temperature compensation was achieved by a negative impedance parallelled flash load.The analog calculation accuracy was 4 bit.The simulation results showed that the temperature coefficient of the stored weights was less than 0.1%/℃from-40℃to 85℃and was less than 0.25%/℃from 85℃to 130℃.

关 键 词：存算一体化 向量矩阵乘法器 闪存 温度补偿 

分 类 号：TN432[电子电信—微电子学与固体电子学] TP183[自动化与计算机技术—控制理论与控制工程]