忆阻器及其存算一体应用研究进展

Review of recent research on memristors and computing-in-memory applications

深度学习的飞速发展带来了巨大的算力需求,然而基于存算分离的“冯·诺依曼架构”的传统硅基芯片面临着“存储墙”等问题,芯片算力增长逐渐陷入瓶颈。为了解决这个矛盾,研究人员从生物大脑的工作模式得到启发,提出了基于忆阻器的存算一体架构。这种全新的架构在处理神经网络等任务时在能效和速度上较“冯·诺依曼架构”有望实现几个数量级的提升,是实现超低功耗、超高算力计算芯片的最有潜力的技术路线之一。本文综述了各种类型忆阻器的工作机理与最新进展,对比了国内外研究团队的器件研究进展;综述了基于忆阻器的存算一体芯片在神经网络、信号处理和机器学习等方向的应用演示的研究进展;总结了基于忆阻器的存算一体芯片目前面临的挑战,并提出中国在该领域进一步发展的建议。

The rapid development of deep learning raises a massive demand for computing power.However,traditional siliconbased chips based on the von Neumann architecture with physically separated memory and computing units,are facing critical issues such as the"memory wall",and hence the increase of chip computing power is gradually hitting a bottleneck.To address this problem,researchers have been inspired by the working mechanism of biological brain and proposed a computing-inmemory architecture based on memristors.This novel architecture is expected to achieve several orders of magnitude improvement in energy efficiency and speed over the von Neumann architecture for tasks such as artificial neural networks.It is one of the most promising technologies to achieve ultra-low power consumption and ultra-high computing power.This article first reviews the working mechanisms of various types of memristors,and summarizes the latest device research internationally.Then,the progress on application demonstrations of memristor-based computing-in-memory chips such as neural networks,signal processing,and machine learning are reviewed.The current challenges in this field and further research directions are concluded in the end.