Ielmini Daniele, Pedretti Giacomo
Dipartimento di Elettronica, Informazione e Bioingegneria (DEIB), Politecnico di Milano and IUNET, piazza L. da Vinci 32, 20133, Milano, Italy.
Artificial Intelligence Research Lab, Hewlett-Packard Labs, 820 N McCarthy Blvd, Milpitas, California 95035, United States.
Chem Rev. 2025 Jun 25;125(12):5584-5625. doi: 10.1021/acs.chemrev.4c00845. Epub 2025 May 2.
In the information age, novel hardware solutions are urgently needed to efficiently store and process increasing amounts of data. In this scenario, memory devices must evolve significantly to provide the necessary bit capacity, performance, and energy efficiency needed in computation. In particular, novel computing paradigms have emerged to minimize data movement, which is known to contribute the largest amount of energy consumption in conventional computing systems based on the von Neumann architecture. In-memory computing (IMC) provides a means to compute within data with minimum data movement and excellent energy efficiency and performance. To meet these goals, resistive-switching random-access memory (RRAM) appears to be an ideal candidate thanks to its excellent scalability and nonvolatile storage. However, circuit implementations of modern artificial intelligence (AI) models require highly specialized device properties that need careful RRAM device engineering. This work addresses the RRAM concept from materials, device, circuit, and application viewpoints, focusing on the physical device properties and the requirements for storage and computing applications. Memory applications, such as embedded nonvolatile memory (eNVM) in novel microcontroller units (MCUs) and storage class memory (SCM), are highlighted. Applications in IMC, such as hardware accelerators of neural networks, data query, and algebra functions, are illustrated by referring to the reported demonstrators with RRAM technology, evidencing the remaining challenges for the development of a low-power, sustainable AI.
在信息时代，迫切需要新颖的硬件解决方案来高效存储和处理日益增长的数据量。在这种情况下，存储设备必须大幅发展，以提供计算所需的必要比特容量、性能和能源效率。特别是，已经出现了新颖的计算范式来尽量减少数据移动，而在基于冯·诺依曼架构的传统计算系统中，数据移动是已知的最大能耗来源。内存计算（IMC）提供了一种在数据内部进行计算的方式，具有最小的数据移动以及出色的能源效率和性能。为了实现这些目标，电阻式开关随机存取存储器（RRAM）因其出色的可扩展性和非易失性存储而似乎是一个理想的选择。然而，现代人工智能（AI）模型的电路实现需要高度专业化的器件特性，这需要对RRAM器件进行精心设计。这项工作从材料、器件、电路和应用的角度探讨了RRAM概念，重点关注物理器件特性以及存储和计算应用的要求。突出了内存应用，如新型微控制器单元（MCU）中的嵌入式非易失性存储器（eNVM）和存储级存储器（SCM）。通过参考已报道的采用RRAM技术的演示器，说明了IMC中的应用，如神经网络的硬件加速器、数据查询和代数函数，这也证明了低功耗、可持续AI发展中仍然存在的挑战。
