The exponential growth of large AI models places heavy demands on computational capacity, energy efficiency, and input/output bandwidth. Conventional electronic computing systems face fundamental limitations due to the physical constraints of electrons. Optoelectronic computing systems, which combine the ultra-high bandwidth, minimal latency, and superior energy efficiency of photons with the logical control and memory of electronics, offer one of the most promising pathways to overcome these bottlenecks in the AI era. In this paper, we establish the strategic significance of optoelectronic computing systems in addressing challenges such as trillion-parameter model training and inference. Using optical domains for high-speed transmission, parallel processing, and specialized computation, alongside electronic domains for general tasks, such systems can meet the multidimensional requirements of next-generation computing infrastructures.

In this paper, the optoelectronic computing system is decomposed into three core parts: data input/output (I/O), data switching, and data processing. Each part is discussed in terms of technical evolution and the latest research progress. In data I/O, the skin effect causes the transmission distance of electrical signals to drop sharply modulation speed increases. A straightforward solution is to place optoelectronic conversion devices closer to chips, as optical domains avoid the speed-distance trade-off. Current research focuses on two approaches to boost I/O bandwidth. One is developing ultra-high-bandwidth platforms, such as 400 Gbit/s demonstrated with electro-absorption modulated laser (EML) and thin-film lithium niobate (TFLN), and greater than 500 GHz with plasmonic modulators. Novel materials such as lanthanum-modified lead zirconate titanate (PLZT, >300 Gbit/s) and BTO platforms (>500 GHz) further expand possibilities. Another approach is to increase the number of channels and wavelengths rather than modulation speed, as shown by Ayar Labs’ 8 Tbit/s and Nubis’ 1.6 Tbit/s demonstrations. Each lane maintains a relatively low modulation speed, such as 32 Gbit/s. In the section on data switching, electronic routers are being replaced by optical circuit switches (OCS). Google’s Apollo OCS system demonstrates significant advantages over conventional systems based on electronic routers, including low latency, high energy efficiency, and cost reduction. After explaining the benefits of OCS over electronic routers, we discuss the latest research on two types of OCS. In the first category, optical switch arrays based on different technologies, including micro-electro-mechanical systems (MEMS), PLZT, liquid crystal on silicon (LCOS), nano-opto-electro-mechanical systems (NOEMS), the thermo-optic effect, the electro-optic effect, and non-Hermitian designs, are summarized in one table. The largest MEMS OCS, from Calient, offers 640 ports, each with low insertion loss and polarization independence. Integrated OCS has fewer ports (128 or 240) with higher insertion loss and polarization dependence, but switches three orders of magnitude faster than MEMS OCS. The second type of OCS, wavelength switching, typically offers a small number of ports, such as 32. In the section on data processing, research in optoelectronic computing systems falls into three categories. One category is free-space-based systems, characterized by ultra-large scale and low propagation loss. Numerous pioneering works have been demonstrated in this category, such as diffractive deep neural networks, reconfigurable reservoir computing systems, and analog iterative machines. The second category is fiber-based systems that extend computational capacity through time and wavelength domains. Typical demonstrations include a temporal-multiplexed Ising machine solving 100000-node models and convolutional neural networks (CNNs) achieving 5.6 TMAC/s. The last type, integrated system, uses CMOS infrastructure and advanced packaging to create compact, efficient optoelectronic computing systems. The largest matrix realized in this category has a size of 512×512.

system scalability, conversion overhead, and bandwidth limitations. For system scalability, optical computing chiplets must work with CPO/OCS technologies, while 2.5D/3D heterogeneous integration enhances photonic-electronic density and mitigates interconnect bottlenecks. In the section on conversion overhead, two specific conversion overheads are explicitly discussed. One exists between the optical domain and the electronic domain, such as in modulators, laser sources, and photodetectors. The other is conversion between analog and digital domains, involving analog-to-digital converters (ADCs), digital-to-analog converters (DACs), and transimpedance amplifiers (TIAs). Based on a systematic analysis and literature review, we find that the ultimate energy efficiency of an optoelectronic computing system is determined by these two conversion overheads. Three reliable methods have been proposed to improve energy efficiency: developing non-volatile memory devices that can significantly reduce energy consumption, utilizing relatively low-precision devices, and designing particular loop structures that can reduce conversion overhead. The last section explains the limitations of system bandwidth. Because of the limited bandwidth of the electronic computing unit, the optical computing core cannot unleash its unprecedented computational capacity. Free-space computing systems and integrated computing systems are discussed in this section as two reliable ways to build optoelectronic computing systems. Finally, a distributed architecture is proposed to meet ultra-high bandwidth requirements.