Technological advancements like autonomous driving and computer vision are driving a surge in demand for computa­tional power. Optical computing, with its high throughput, energy efficiency, and low latency, has garnered considerable attention from academia and industry. However, current optical computing chips face limitations in power consump­tion and size, which hinders the scalability of optical computing networks. Thanks to the rise of nonvolatile integrated photonics, optical computing devices can achieve in-memory computing while operating with zero static power consumption. Phase-change materials (PCMs) have emerged as promising candidates for achieving photonic memory and nonvolatile neuro­morphic photonic chips. PCMs offer high refractive index contrast between different states and reversible transitions, making them ideal for large-scale nonvolatile optical computing chips.

While the promise of non­volatile integrated optical computing chips is tanta­lizing, it comes with its share of challenges. The need for frequent and rapid switching, essential for online training, is a hurdle that researchers are determined to overcome. Forging a path towards quick and efficient training is a vital step on the journey to unleash the full potential of photonic computing chips. Recently, researchers from Zhejiang University, Westlake University and the Institute of Micro­electronics of the Chinese Academy of Sciences achieved a promising result. They developed a 5-bit photonic memory capable of fast volatile modulation and proposed a solution for a nonvolatile photonic network supporting rapid training. This was made possible by integrating the low-loss PCM anti­monite (Sb₂S₃) into a silicon photonic platform.

The photonic memory utilizes the carrier dispersion effect of a PIN diode to achieve volatile modulation with a rapid response time of under 40 nanoseconds, preserving the stored weight infor­mation. After training, the photonic memory utilizes the PIN diode as a microheater to enable multilevel and reversible phase changes of Sb₂S₃, allowing the storage of trained weights in the photonic computing network. This leads to an incredibly energy-efficient photonic computing process.

Using the demonstrated photonic memory and working principle, the research team simulated an optical convolu­tional kernel archi­tecture. Remarkably, they achieved over 95 percent accuracy in recognizing the MNIST dataset, showcasing the feasibi­lity of fast training through volatile modulation and weight storage through 5-bit non­volatile modulation. This work establishes a new paradigm for photonic memory and offers a promising solution for implementing nonvolatile devices in fast-training optical neural networks. With these advance­ments, the future of optical computing looks brighter than ever before. (Source: SPIE)

Reference: M. Wei et al.: Electrically programmable phase-change photonic memory for optical neural networks with nanoseconds in situ training capability, Adv. Phot. 5, 046004 (2023); DOI: 10.1117/1.AP.5.4.046004

Link: State Key Laboratory of Modern Optical Instrumentation, College of Information Science and Electronic Engineering, Zhejiang University, Hangzhou, China