Reconfigurable photonic neural networks: distribution-aligned calibration and dynamic resource allocation for high-performance optical computing accelerators
Songcheng Zhou, Ziqiang He, Yiheng Zhao, Bo Xu, Peng Zou, Fangchen Hu, Wei Cheng Chu · Optics Express · 2025
Driven by the rapid adoption of generative models and large language models, the computational demand and scale of modern intelligent data centers have surged. Concurrently, the energy consumption and operational costs of AI workloads are rising exponentially. Photonic computing-offering large bandwidth, low energy consumption, and sub-nanosecond latency-has therefore emerged as a promising paradigm for large-scale AI deployment. However, most existing photonic-AI studies investigate only a single architecture with fixed parameter settings, neglecting dynamic configuration of key variables such as optical power and computational scale. This limitation constrains the attainable energy efficiency and compute density of the system. To overcome these constraints, we propose a distribution-alignment calibration (DAC) algorithm for photonic convolution on an optical computing accelerator (OCA), along with dynamic power allocation (DPA) and dynamic dimension allocation (DDA) schemes. Tested on a calibrated photonic-computing simulator, DAC increases inference accuracy from 12.93% to 75.77% at an input optical power of 5 dBm. In addition, DDA reduces power consumption by 18.7% and 20.9% on ResNet-18 and ResNet-50, respectively, while DPA boosts compute density by 24.84% and 28.19% on the same models. The hardware results demonstrate that applying DPA achieves optical power savings of 16.98% on ResNet-18 and 17.83% on ResNet-50, both slightly lower than those obtained in simulation. These results confirm that dynamically adjusting core size and optical power enables a more flexible, energy-efficient, and high-performance photonic computing system.