Deep learning (DL) has become a powerful driving force in the era of intelligence and has been widely used in computer vision, speech recognition, natural language processing, etc. However, more than 80% of calculations in DL are matrix-matrix multiply-accumulate (MM-MAC) operations. Large-scale MM-MAC operations result in a large number of memory access requirements when the algorithm is converted into CPU-executable code. Limited by the von Neumann architecture of electronic computers and the physical constraints of the interconnection limit of copper wires on a chip, the training efficiency and speed of deep neural networks (DNNs) are severely restricted. According to the research of de Lima et al., the computing power required to train state-of-the-art DNNs doubles approximately every 3.5 months, far exceeding the computing power supply of electronic integrated circuits (EICs) that follow Moore’s Law.

Compared to traditional electronic computing, optical computing is expected to build artificial intelligence (AI) accelerators with high computing power and energy efficiency ratio owing to the high parallelism, high speed, and low power consumption of photons. Currently, various optical computing architectures have demonstrated advantages in terms of high computing power and energy efficiency ratio, and their development routes can be divided into two types. One route is to realize dedicated optical information processing based on multidimensional optical signal modulation and to primarily focus on analog optical computing, such as multiply-accumulate (MAC) operations, convolutions and correlations, differentiation and integration, Fourier transform, and optical neural networks (ONNs). The other route is to use the conception of electronic computers to design digital optical computers, such as optical transistors, optical logic devices, optical directed logic operations, space-time parallel coding, and ternary optical computers. Additionally, some important supporting technologies such as optical interconnections, optoelectronic copackaging, optoelectronic heterogeneous integration, and three-dimensional advanced packing have been widely used to improve the performance of electronic computers.

In general, owing to the lack of efficient and reliable weak-light nonlinear optical effects and optical logic devices, it is difficult for photons to realize general digital logic computers like electrons. In addition, the technology that uses photons to store information has not been proven effectively, which implies that photons cannot independently complete the entire process between memory and computing, and all-optical signal processing is still challenging to achieve. Therefore, from electronic information storage to photonic information loading or from photonic information loading to electronic information storage, high-precision and high-speed parallel electronic control systems and analog-to-digital conversion circuits are still required to fully utilize the parallelism of optical computing. Currently, optical computing is primarily based on linear analog computing, and its computing accuracy is sufficient to build practical high-performance AI accelerators. Moreover, by developing suitable coding schemes, parallel algorithms, and architectures to further fully utilize the parallelism of each dimension of photons, photons are expected to provide a computing power density and energy efficiency ratio that exceed those of electrons on the same footprint. Furthermore, a high computing accuracy can be realized even with error-sensitive photonic devices and optical systems. Although binary electronic logic computing can simulate various practical physical scenarios with a sufficiently high computational accuracy, the ever-increasing computational load will significantly increase power consumption. Optical computing is expected to build high-performance and high-energy-efficiency fuzzy parallel computing systems similar to the human brain with limited computing precision.

This review analyzes and discusses mainstream optical computing technologies from the perspective of analog and digital optical computing, aiming to guide the development of optical computing. First, we introduce three main technical paths for solving the bottleneck of computing power supply and power consumption in the post-Moore era (Fig.1). Additionally, we indicate that optoelectronic computing or all-optical computing is the most promising method for building the next generation of human-like fuzzy parallel computing systems with high computing power and energy efficiency (Fig.2). We then summarize the advantages and disadvantages of analog and digital optical computing (Table 1) and discuss the main progress and representative achievements of optical computing at different stages, including early optical computing (Fig.3), integrated optical computing (Fig.4), free-space interconnected optical computing (Fig.5), and multi-imaging-casting architecture (Fig. 6). On this basis, we describe the limitations and key technical bottlenecks facing the further development of optical computing. Finally, we discuss future trends and directions of optical computing.

The route of imitating the digital electronic computer to construct a general computer for photonic logic is severely limited by optical logic devices. Currently, the focus of optical computing should be on special application scenarios that take full advantage of optical parallelism and are challenging to solve by electronic computing. From a long-term perspective, in addition to AI technology, the contradiction between the strategic goal of carbon neutrality and the significant energy consumption of data centers that provide high computing power for the rapidly increasing data processing requirements will continue to aggravate. Before the advent of practical quantum computers, the development of energy-saving, highly efficient, and high computing power optoelectronic intelligent computing is the most promising solution.