LightPro: a linear photonic processor with full programmability.

Area of Science:

• Photonics
• Artificial Intelligence
• Hardware Acceleration

Background:

• Conventional photonic hardware for deep neural networks faces limitations in scalability, optical losses, and footprint.
• Mach-Zehnder interferometers are commonly used but contribute to these limitations.

Purpose of the Study:

• To present LightPro, a novel programmable linear photonic processor.
• To optimize scalability, power efficiency, and area footprint for deep neural network acceleration.

Main Methods:

• Integrated a neural architecture search and pruning framework with tunable phase-change material directional couplers.
• Dynamically adjusted coupling coefficients by thermally modulating phase-change material states.
• Validated phase-change material devices using multiphysics simulations and compact models against experimental data.

Main Results:

• LightPro architectures achieved up to 85% footprint reduction and over 50% decrease in power consumption.
• Network scaling evaluations showed less than 5% inference accuracy degradation.
• Experimental prototyping validated the computational accuracy and scalability of LightPro.

Conclusions:

• LightPro offers a scalable and efficient pathway for next-generation photonic artificial intelligence accelerators.
• The architecture overcomes fundamental physical scaling limitations of current photonic hardware.