Recent progress in neuromorphic computing from memristive devices to neuromorphic chips

This review first revisits the theoretical background and developmental history of neuromorphic computing. It then briefly introduces the working mechanisms of memristive devices and how they can mimic synaptic plasticity and neuronal activation, followed by a detailed analysis of their performance in terms of energy consumption, integration density, stability, and resistance-state tuning precision. On this basis, the review introduces typical array structures built from memristive devices, such as crossbar arrays, and discusses their applications in constructing artificial neural networks, including perceptrons, convolutional neural networks, and spiking neural networks. Next, the review explores the architectures and performance metrics of neuromorphic chips based on memristive arrays and peripheral circuits, focusing on energy efficiency, area efficiency, training efficiency, and inference accuracy. Finally, it examines the intrinsic links among memristive devices, memristive arrays, and neuromorphic chips, emphasizing the importance of collaborative optimization from device-level to system-level in enhancing overall system performance.

Against the backdrop of rapid development in artificial intelligence and big data, neuromorphic computing offers a novel approach to overcoming the energy bottleneck faced by traditional computing. Leveraging the advantages of memristive devices in emulating neural functions and providing high-density storage, combined with the scalable structure and parallel computing capabilities of memristive arrays, neuromorphic computing chips exhibit energy efficiency far exceeding that of traditional chips in specific applications. However, challenges remain in applying neuromorphic computing to complex tasks. Future efforts need to further investigate the intrinsic links among memristive devices, memristive arrays, and neuromorphic chips to drive the continuous optimization and key breakthroughs of the technologies. With ongoing convergence and innovation in materials science, microelectronics, and computing architecture, memristive-device based neuromorphic chip is poised to become a foundational element of next-generation intelligent computing systems, providing support in scientific computing, edge computing, and the Internet of Things.