This work proposes a novel, scalable photonic convolutional neural network architecture (MMORR-PCNN) for high-throughput and low-latency inference in edge artificial intelligence (AI) systems. By integrating 4 × 4 multi-mode interference (MMI) couplers with multi-operand microring resonators (MMORRs), the architecture exploits both wavelength-division multiplexing (WDM) and mode-division multiplexing (MDM) to achieve 192 parallel multiply-accumulate (MAC) operations per photonic cycle. Each MMORR unit performs 8 MACs at 7 GHz which achievie a total of 8192 MACs/cycle and a compute density of 13.44 tera-operations per second per square millimeter (TOPS/mm 2 ) across a 12.8 mm 2 chip. The architecture delivers sub-nanosecond latency ( ∼ 140 ps) and operates within a power budget of 8.8–15 W. An entropy-regularized training method mitigates inter-channel crosstalk, ensuring reliable operation under 2–8-bit quantization. Structured sparsity reduces photonic complexity by 43.75% without degrading accuracy. The proposed architecture generalizes across MNIST, Fashion-MNIST (FMNIST), and CIFAR-10 datasets, and scales to deeper models such as VGG16 and VGG19, achieving up to 98.8% accuracy. A rigorous coupled-mode theory (CMT) analysis validates the resonance behavior and mode-selective weight encoding of the MMORR unit, while comprehensive FDTD simulations confirm the physical realizability of the proposed architecture under realistic silicon photonic fabrication constraints. These results validate MMORR-PCNN as a compact, energy-efficient, and CMOS-compatible platform for real-time photonic AI inference.
Do et al. (Fri,) studied this question.