the von Neumann architecture’s energy efficiency bottleneck. Shuttling data between physically separated memory and processing units consumes excessive power, hindering scalability for complex tasks like real-time pattern recognition and large-scale simulation. This has prompted a shift in research paradigms from solely relying on process miniaturization towards exploring non-von Neumann architectures and novel physical carriers. Neuromorphic computing, inspired by the brain’s massively parallel, event-driven, and co-located memory/compute structure, offers a revolutionary path forward. Moreover, optical computing has emerged as a particularly compelling platform for neuromorphic implementation due to its inherent ultrahigh speed (enabling terahertz operations), massive parallelism (through wavelength and spatial multiplexing), and low power consumption (minimal resistive losses during signal propagation). In recent years, by combining the advantages of optical computing and neuromorphic computing, specialized optical processors have demonstrated a potential far exceeds that of electronic processors in tasks such as mathematical operations and signal processing.

Nevertheless, existing optical artificial intelligence accelerators still exhibit significant limitations: on one hand, they are functionally singular, tailored only for specific neural network architectures or tasks, lacking versatility. On the other hand, the model complexity and experimental performance of optical neural networks remain relatively low, with key metrics such as classification accuracy still lagging behind those of electronic neural networks. This stems primarily from the inherent physical properties of photons: the difficulty in efficiently storing and modulating optical signals makes fully optical, general-purpose neuromorphic computing still elusive.

Against this backdrop, optoelectronic-integrated neuromorphic computing has emerged as the most viable technological pathway. By organically combining the flexibility of electronic control with the high bandwidth, low latency, and other advantages of photonic transmission, it utilizes the high-speed parallel nature of light for rapid data transmission and processing, reducing data transfer delays. Simultaneously, it leverages the precise control and storage capabilities of electronics to implement complex logic operations and information storage. This integration can break through limitations of traditional computing architectures, significantly enhancing computational speed and energy efficiency to meet the high-performance computing demands of fields like artificial intelligence and big data.

Neuromorphic computing fundamentally necessitates the deployment of sophisticated algorithms onto dedicated neuromorphic hardware. This requirement inherently demands a profound level of co-design and optimization, a process that must span both the architectural blueprint and the functional implementation of the algorithms and the neuromorphic devices. When considering the hardware implementation level specifically, the simulation of spiking neural networks typically adopts a modular design approach. Within this approach, a key separation occurs: the computational function performed by the neuronal soma is deliberately segregated from the weighted connections managed by the synapses.

The application of various optical neuromorphic devices within this neuron simulation paradigm is sequentially illustrated in Fig. 2. Utilizing these devices as foundational components, researchers have successfully achieved significant milestones: optical simulations accurately replicating neuronal spiking dynamics, simulations effectively modeling synaptic plasticity, and implementations enabling reconfigurable weight configurations. Leveraging this intrinsic “propagation-as-computation” characteristic, along with the ultra-strong parallel processing capabilities inherent to photonics, becomes crucial. Consequently, photonic computing technology targeted for artificial intelligence demonstrates substantial advantages?specifically computational speed and energy efficiency advantages measured to be several orders of magnitude higher than those achievable with conventional electronic computing.

Most importantly, the fundamental functional units underlying neuromorphic computing possess distinct properties categorized as linear computation and nonlinear computation. The implementation of these units relies upon the use of either active devices or passive devices. Based precisely on these fundamental unit properties and their implementation dependencies, the recent progress made in developing three different hardware architectures specifically for optoelectronic neuromorphic integrated computing is systematically introduced. To provide a detailed comparative analysis, Table 2 further examines and contrasts the performance characteristics of these three hardware architectures, which encompass critical metrics across several dimensions: nonlinear capability, power efficiency, maximum response speed, scalability potential, associated cost factors, and physical integration area requirements.

Despite demonstrating immense potential in speed and power consumption, optical neuromorphic computing still faces numerous critical challenges on its path to achieving high scalability, density, and performance—moving from specialized computing to general-purpose architectures, and transitioning from discrete experimental verification to practical applications—when compared to biological neural systems with their highly structured, parallel, adaptive, and data fault-tolerant capabilities. The development of optoelectronic integrated neuromorphic computing relies on multi-dimensional, deep-level collaborative breakthroughs. This requires not only the continuous evolution of optoelectronic integration technology and innovation in neuromorphic computing algorithmic models, but more critically, deep co-design and optimization between the optoelectronic integrated hardware systems and the algorithmic models.