Fig. 1. Categories of photonic MVM. (a) PLC-based photonic MVM; (b) MZI-based photonic MVM; (c) WDM-based photonic MVM

Fig. 2. Summary of applications in optoelectronic intelligent computing

Fig. 3. In situ training of optical neural networks through BP algorithm. (a) Chip of integrated optical neural networks^[77]; (b) forward propagation of diffractive optical neural networks^[78]; (c) backward propagation of diffractive optical neural networks^[78]

Fig. 4. Online training of optoelectronic intelligent computing chip through SGD algorithm. (a) Brief flow chart of SGD algorithm; (b) multifunctional on-chip polarization processor^[42]; (c) photonic accelerator for Google PageRank algorithm^[43]; (d) self-configuring and fully reconfigurable silicon photonic signal processor^[44]

Fig. 5. Three typical optoelectronic intelligent computing architectures. (a) Coherent MZI mesh^[40]; (b) photonic accelerator based on time-wavelength interleaving^[62]; (c) integrated photonic tensor core based on PCM^[61]

Fig. 6. Factors influencing computing capacity and energy consumption

Table 1. Comparison of computing capacity and energy efficiency of microelectronic chips and optoelectronic chips

Table 2. Computing capacity and energy efficiency of different optoelectronic computing architectures