image: Evolution timeline of core computing architecture in photonic neural networks
Credit: ©Science China Press
With the rapid development of artificial intelligence (AI), applications such as large models, intelligent sensing, autonomous driving, and edge computing continue to emerge. Computing systems are facing unprecedented challenges in computing power, bandwidth, and energy efficiency. Electronic computing systems based on the traditional von Neumann architecture are increasingly limited by interconnect bandwidth and energy consumption bottlenecks, making it difficult to continuously meet the demands of AI for high speed, low power consumption, and highly parallel computing. Against this background, integrated photonic neural networks (PNNs) have emerged as a promising computing paradigm for overcoming existing computing bottlenecks.
Recently, a joint research team published a review article titled “Emerging integrated photonic neural network technologies for artificial intelligence: from devices to systems”. The article systematically reviews the development landscape, key architectures, and system implementation paths of integrated PNNs for AI applications. The review adopts a cross-layer analytical framework and provides a comprehensive summary of key technological advances in integrated PNNs for AI applications. Covering physical devices, computing core architectures, on-chip integration strategies, and advanced system packaging, the team analyzes the key hardware requirements and mainstream technical routes of PNNs. The team also summarizes the challenges and future potential of these technologies in scalability, computational precision, energy efficiency optimization, and optoelectronic system coordination.
The review points out that, for on-chip PNNs, high-performance passive and active photonic devices form the physical foundation of system operation, while different network architecture schemes determine the efficiency, throughput, and scalability of the computing core. On this basis, the article further discusses the evolution of integration platforms, from relatively mature hybrid integration to rapidly developing heterogeneous and monolithic integration, and then to advanced packaging technologies for system-level applications. These developments indicate that integrated PNNs are gradually forming a technical path to address the computing and data transmission bottlenecks in AI.
Looking ahead, further performance improvement of PNNs will benefit from the coordinated evolution of advanced material platforms, computing core designs, and optoelectronic system architectures. New photonic materials and integrated platforms are expected to bring new capabilities in modulation speed, integration density, and transmission loss. At the system level, the development of PNNs will rely on deep optoelectronic integration and collaborative optimization design. Photonic technologies can take advantage of their strengths in large-scale linear operations, high-bandwidth interconnects, and low-latency parallel processing, while electronic technologies can provide support in nonlinear operations, memory management, and precise control. Through coordinated design of optical and electronic functions at the architecture level, on-chip PNNs are expected to further improve overall computing efficiency, energy efficiency, and engineering feasibility. On this basis, as manufacturing processes become more standardized and advanced packaging technologies continue to mature, on-chip PNNs based on optoelectronic integration platforms are expected to evolve from experimental demonstration systems into co-integrated accelerator modules for next-generation AI infrastructure.
Overall, this review follows the complete chain from devices, architectures, integration, and packaging to systems. From the perspectives of hardware foundations, architecture design, integration processes, system packaging, and industrialization pathways, it systematically presents the development logic of integrated PNNs for AI applications. It also discusses the current progress, core challenges, and future trends of emerging integrated PNN technologies, providing a useful reference for researchers to understand the broader technical landscape of integrated photonic intelligent computing and to identify future directions. As the demand for AI computing continues to grow, integrated PNN technologies are expected to become an important part of next-generation AI computing infrastructure and to play a significant role in high-performance computing, data centers, and related fields.