Artificial intelligence (AI) compute demands outstrip electronic hardware scaling, creating a critical “AI compute gap.” Photonic computing promises speed, low latency, and parallelism yet faces bottlenecks in core array scalability and optoelectronic bandwidth efficiency. We propose a time-domain parallelism (TDP) computing architecture that temporally multiplexes compact cores, enabling large-scale convolutional operations while optimizing bandwidth efficiency. Supporting this framework, we develop indium tin oxide (ITO)–based photonic devices demonstrating a record 9-bit dynamic reconfigurability at 60 kilohertz and a 3-decibel bandwidth of 311 kilohertz. These devices feature compact phase shifter lengths ( L π = 20 micrometers) and exceptional endurance (>10 10 cycles with <0.1-decibel extinction ratio degradation). Validation of the TDP system using a 2 × 2 ITO photonic array achieves an accuracy of 84.91% on Fashion-MNIST classification, comparable to the full-precision (FP32) software baseline (84.66%) and outperforming a 5-bit benchmark (71.94%). Notably, the system maintains this high-fidelity performance while operating at 1694 frames per second. This work establishes a scalable pathway toward energy-efficient photonic computing for AI workloads.
Che et al. (Thu,) studied this question.