Reconfigurable photonic neural networks: distribution-aligned calibration and dynamic resource allocation for...

Area of Science:

• Photonics
• Artificial Intelligence
• Computer Engineering

Background:

• The increasing demand for generative and large language models drives up computational needs and energy costs in data centers.
• Photonic computing offers a solution with high bandwidth, low energy use, and low latency for AI.
• Current photonic AI research often uses fixed parameters, limiting system efficiency and performance.

Purpose of the Study:

• To address the limitations of fixed parameter settings in photonic AI systems.
• To enhance energy efficiency and compute density in photonic computing for AI.
• To introduce dynamic configuration schemes for optical computing accelerators.

Main Methods:

• Proposed a distribution-alignment calibration (DAC) algorithm for photonic convolution.
• Implemented dynamic power allocation (DPA) and dynamic dimension allocation (DDA) schemes.
• Validated the algorithms using a calibrated photonic-computing simulator and hardware experiments.

Main Results:

• DAC improved inference accuracy from 12.93% to 75.77% at 5 dBm input power.
• DDA reduced power consumption by up to 20.9% for ResNet models.
• DPA increased compute density by up to 28.19% and achieved significant optical power savings.

Conclusions:

• Dynamic adjustment of optical power and computational scale optimizes photonic computing systems.
• The proposed DAC, DPA, and DDA schemes enable more flexible, energy-efficient, and high-performance AI.
• Photonic computing with dynamic configurations is a viable path for future AI infrastructure.