军事知识服务平台
面向类脑视觉的忆阻器存算一体器件和架构研究进展
Research progress on memristor-based computing-in-memory devices and chips for neuromorphic vision
摘要 类脑视觉模型在数据处理规模、并行计算能力与实时推理性能方面对硬件平台提出了严苛要求,而传统冯·诺依曼架构存算分离的固有特性所引发的带宽受限与高能耗问题,已成为制约类脑视觉系统性能进一步提升的核心瓶颈.受生物大脑并行信息处理机制启发,忆阻器基存算一体架构可在存储单元内实现原位运算,有效规避了频繁的数据搬运过程,为降低运算开销、提升系统能效提供了新型技术路径.本文针对面向类脑视觉任务的忆阻器存算一体架构开展了系统性综述,重点梳理分析了器件的性能特性、阵列结构类型、系统架构设计、软硬件协同优化及典型应用场景等关键内容.在此基础上,总结归纳了该架构在运算精度、系统可靠性、外围电路开销及多模块协同适配等方面存在的技术挑战,并结合当前技术发展趋势,对忆阻器存算一体类脑视觉系统的后续研究与发展方向进行了展望.
Abstract:
Brain-inspired visual models impose stringent hardware requirements in terms of data scale,parallel computing,and real-time processing.The inherent bandwidth and energy consumption bottlenecks caused by the separation of storage and computation in the traditional von Neumann architecture severely restrict the performance improvement of visual processing systems.Inspired by the parallel information processing mechanism of biological brains,memristor-based computing-in-memory architectures enable in-situ computation within memory cells,effectively reducing data movement overhead and providing a promising approach to high-efficiency visual computing.This paper systematically reviews the recent advances of memristor-based computing-in-memory architectures for brain-inspired visual tasks,with a focus on device characteristics,array configurations,system architecture design,hardware-software co-optimization,and typical application scenarios.Furthermore,the major technical challenges regarding computing accuracy,system reliability,peripheral circuit overhead,and multi-module coordination are summarized.The future development trends and potential research directions of memristor-based computing-in-memory systems for brain-inspired visual processing are finally prospected.
年,卷(期): 2026, 58(3)
分类号: TP303
机标分类号: TP393 TN919.81 TP11
在线出版日期: 2026年10月15日
基金项目: 2023YFB4402303:国家科技计划项目52472149:国家自然科学基金20260201080GX:吉林省科技发展计划重点研发项目,