Fig. 1. Evolutionary phases of large-scale computational hardware architecture^[9-26]

Fig. 2. Coherent optical computing architecture based on MZI array. (a) Triangular MZI array^[43]; (b) square MZI array^[44]; (c) optical neural network based on MZI array^[27]; (d) complex-valued optical neural network based on MZI array^[48]; (e) real-valued matrix based on unitary array^[49]; (f) butterfly-topology MZI array^[50]; (g) extended double-layer-like topology MZI array^[51]; (h) ultra-compact computing array based on inverse-designed devices^[52]; (i) hybrid computing architecture of MZI array and diffractive units^[53]; (j) time-division multiplexed complex-valued optical computing array^[54]

Fig. 3. WDM optical computing architecture based on MRR array. (a) MRR-based weight bank^[55]; (b) MRR-based optical neural network^[56]; (c) feedback-enabled MRR weight unit^[57]; (d) digital-analog hybrid computing architecture implemented with MRR^[58]; (e) optical frequency comb-based optical computing architecture^[60]; (f) cross-programmable MRR weight array^[59]; (g) WDM/TDM-based MRR weight array^[61]; (h) computation module based on optical frequency comb and MRR array^[62]; (i) WDM/MDM-based MRR weight array^[63]

Fig. 4. Optical computing architectures based on on-chip color units. (a) Multi-port MMI-based computation module^[64]; (b) multi-port coupler-based operation module^[65]; (c) matrix multiplication using on-chip diffractive units^[66]; (d) MMI-implemented convolutional neural network^[67]; (e) subwavelength slot structure-based optical neural network^[69]; (f) programmable diffractive optical computing on III-V platform^[71]; (g) field-programmable diffractive optical neural network^[72]; (h) in-memory computing architecture based on diffractive units^[70]

Fig. 5. Development status of opto-electronic signal conversion nonlinear activators. (a) Balanced detection-based MRR nonlinear activator^[81]; (b) on-chip photodetector-based nonlinear activator^[82]; (c) programmable electro-optic nonlinear activator^[83]; (d) cascaded computing network using optoelectronic hybrid nonlinear activator^[84]; (e) wavelength-selective nonlinear activator^[85]

Fig. 6. Development status of all-optical nonlinear activators. (a) SOA-based all-optical nonlinear activator^[86]; (b) silicon-based all-optical nonlinear activator^[87]; (c) silicon-germanium-based all-optical nonlinear activator^[88]; (d) 2D material-based nonlinear activator^[89]; (e) phase-change material-based nonlinear activator^[90]; (f) multi-layer diffractive optical neural network implemented with all-optical nonlinear activator^[91]

Fig. 7. Hardware deployment methods of computing models for integrated optoelectronic computing chips and systems. (a) MZI array matrix decomposition^[43]; (b) MRR unit-by-unit configuration^[98]; (c) diffractive structure matrix mapping^[66]; (d) in-situ backpropagation method^[33]; (e) resonant state monitoring^[104]; (f) asymmetric training method^[105]; (g) perturbation iterative optimization^[106]; (h) network training based on genetic algorithm^[108]; (i) network training based on particle swarm optimization^[109]

Fig. 8. Development status of integrated optoelectronic computing systems for high performance computing models. (a) Cascaded optical neural networks^[84]; (b) spatial-on-chip optoelectronic hybrid neural networks^[110]; (c) monolithically integrated multi-layer optical neural networks^[107]; (d) monolithically integrated recurrent optical neural networks^[85]; (e) PACE universal photonic AI processor^[35]; (f) Envise universal photonic AI processor^[36]

Fig. 9. Development status of integrated optical computing systems for practical applications. (a) Fiber-optic communication nonlinear compensation based on optical computing^[111]; (b) wireless signal blind source separation based on optical computing^[112]; (c) signal equalization based on optical reservoir computing^[113]; (d) fiber-optic sensing based on optical computing^[114]

Fig. 10. Development framework of key technologies for integrated optical computing