A new publication from Opto-Electronic Technology; DOI 10.29026/oet.2025.250011 , discusses integrated photonic synapses, neurons, memristors, and neural networks for photonic neuromorphic computing.
As large-scale models and edge intelligence rapidly proliferate, computing systems are increasingly constrained by a triple bottleneck: insufficient bandwidth, excessive power consumption, and slow data movement between memory and computation. In the conventional von Neumann architecture, where memory and computing units are physically separated, data must be shuttled back and forth repeatedly across and within chips, causing latency and energy costs to grow sharply with system scale. Against this backdrop, neuromorphic photonics—which uses light as the information carrier—has attracted substantial attention. Light inherently offers ultra-high bandwidth, low latency, and massive parallelism, and can execute core operations such as matrix-vector multiplication during propagation. This opens a path to transforming AI inference from a major power consumer into a "light-speed, low-power" paradigm.
In this review, the authors focus on integrated photonic neural networks (IPNNs) as a key hardware route and systematically summarize three foundational device classes: photonic synapses for weight storage and loading, photonic neurons for nonlinear activation, and photonic memristors for short-term and long-term memory. The article further surveys multiple IPNN architectures and the latest advances, and discusses the critical issues that must be addressed for practical deployment—namely calibration and stability, photonic-electronic co-integration, programmable general-purpose architectures, and efficient training strategies—thereby providing a roadmap toward photonic AI chips that are low-energy, scalable, and trainable.
A research team led by Academician Gu Min at the University of Shanghai for Science and Technology (USST) was invited to publish a review article entitled "Integrated photonic synapses, neurons, memristors, and neural networks for photonic neuromorphic computing" in Opto-Electronic Technology (2025, Issue 3). The review centers on the critical chain from devices to systems for IPNNs, spanning device principles—network architectures—engineering translation. It examines in depth three key building blocks-photonic synapses, photonic neurons, and photonic memristors-and their roles in neuromorphic computing, systematically summarizes major technological progress in the field, and outlines future directions.
The article introduces photonic synapse devices like MRRs, MZIs, and PCMs (Fig. 1), which store neural-network weights. MRRs are compact and energy-efficient, while non-volatile PCMs excel in weight storage. These components enable low-power, highly parallel photonic computation.
It then explores photonic neurons for nonlinear activation, favoring all-optical schemes for efficiency. Photonic memristors, categorized as non-volatile (long-term storage) and volatile (fast temporal processing), provide optical memory. The review details four IPNN architectures: coherent networks (e.g., MZI meshes) enabling on-chip training; parallelized IPNNs using multiplexing for high throughput; integrated diffractive networks for low-latency inference; and reservoir computing for dynamic signal processing. Large-scale deployment hinges on both device performance and system-level capabilities like stability and trainability.
The review concludes that IPNNs hold great potential for low-power, high-performance computing. However, commercialization faces challenges in stability, integration density, and system-level engineering. Innovations in optoelectronic hybrid integration and programmable platforms (Fig. 2) are expected to overcome these hurdles, enabling future adoption in edge computing, autonomous driving, and intelligent manufacturing.
Overall, IPNNs are forming a clear trajectory in which devices are feasible, architectures are scalable, and systems are trainable. Nevertheless, achieving general-purpose, low-power photonic AI will require breakthroughs in low-energy nonlinearities and storage, calibration stability for large-scale arrays, photonic-electronic co-packaging, and programmable general-purpose architectures. As phase-change materials, microcombs, multidimensional multiplexing, and chiplet-based integration continue to advance, photonic neuromorphic computing is poised to be deployed first in edge-intelligence and real-time inference scenarios.
Keywords: neuromorphic photonics, integrated photonic neural networks (IPNNs), photonic integrated circuits (PICs)
