A Reconfigurable and Multifunctional Sensing-Memory-Computing Unit Based on Photodiodes and Memristors for Self-Powered Artificial Vision Perception

Abstract To overcome the limitations of the conventional Von Neumann architecture in energy consumption and processing time, sensing-memory-computing integrated devices inspired by the human visual system demonstrate great potential. Here, a sensing-memory-computing unit based on the hybrid integration of a photodiode and a memristor is proposed, enables flexible hybrid integration of visible or near-infrared photodiodes with digital or analog memristors by a standard semiconductor fabrication process. The storage function is achieved by the large switching ratio of the digital memristor, while the computing function is achieved by the synaptic properties of the analog memristor. A novel self-powered operating scheme based on this device is proposed, in which the photovoltage generated by the photodiode directly drives the memristor, enabling self-powered readout and computation. In image denoising and encoding tasks, the denoising process achieves ultralow training power consumption and self-powered inference, while the encoding process achieves an image classification accuracy of up to 85%. This work promotes the integration of sensing-memory-computing and provides a solid foundation for efficient and biomimetic neuromorphic computing.