Optimizing hardware-software co-design based on non-ideality in memristor crossbars for in-memory computing

作     者：Pinfeng JIANG Danzhe SONG Menghua HUANG Fan YANG Letian WANG Pan LIU Xiangshui MIAO Xingsheng WANG 

作者机构：School of Integrated Circuits Huazhong University of Science and Technology Hubei Yangtze Memory Labs 

出 版 物：《Science China(Information Sciences)》 (中国科学:信息科学(英文版))

年 卷 期：2025年

核心收录：

学科分类：080903[工学-微电子学与固体电子学] 0809[工学-电子科学与技术（可授工学、理学学位）] 08[工学] 081201[工学-计算机系统结构] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：supported in part by National Natural Science Foundation of China (Grant Nos. U2341221, 62274070) Hubei Province Science and Technology Major Project (Grant No. 2022AEA001) Interdisciplinary Research Program of Huazhong University of Science and Technology (Grant No. 2023JCYJ042) Hubei Key Laboratory of Advanced Memories 

摘      要：The memristor crossbar, with its exceptionally high storage density and parallelism, enables efficient vector matrix multiplication (VMM), significantly improving data throughput and computational efficiency. However, its analog computing is vulnerable to issues like IR-drop, device-to-device (D2D) variation, and stuck-at-fault (SAF), leading to a substantial decrease in the inference accuracy of neural networks deployed on crossbars. This work presents a hardware-software co-design approach tailored to deal with memristor crossbar non-ideality. We introduce an end-to-end functional array simulator (FAST) for precise and ultra fast end-to-end training, mapping, and evaluation of neural networks on the memristor crossbar. Utilizing the sparsity of the memristor crossbar coefficient matrix, it achieves simulation with low storage and computational resource requirements,dynamically selecting the optimal solution to complete the process. It can also precisely simulate the impact of non-ideal effects such as IR-drop, retention, variation, SAF, and AD/DA precision. Using FAST, we assess memristor crossbar matrix operations under non-ideal conditions, identifying the max throughput and the most energy-efficient crossbar configurations. Additionally,we propose a comparator-based activation function modulation (CAFM) scheme and its corresponding hardware architecture with programmable activation function circuits to address the IR-drop issue, enabling low power and area overheads, resulting in the recovery of neural network accuracy by 54%or more. This is validated within FAST, demonstrating the success of our hardware-software optimization co-design.