LightPro: A Linear Photonic Processor with Full Programmability

This article presents LightPro, a fully programmable linear photonic processor aimed at enhancing the scalability, performance, and footprint of photonic matrix-vector multiplication (MVM) for AI hardware accelerators. The proposed processor leverages Silicon Photonics (SiPh) technology and integrates phase-change materials (PCMs) to dynamically adjust the coupling ratios of directional couplers (DCs). This approach allows for a more compact and energy-efficient implementation compared to traditional photonic MVM hardware.

Key points include:

1. LightPro Architecture: The core of LightPro consists of programmable SiPh directional couplers that integrate PCMs. The PCM's phase changes (from amorphous to crystalline states) enable dynamic tuning of the coupling coefficients, allowing LightPro to perform different MVM operations. This architecture reduces both the footprint and power consumption compared to conventional MVM hardware, achieving up to 85% reduction in footprint and 50% improvement in power efficiency.

2. Neural Architecture Search (NAS): To optimize the processor's design for MVM operations, a NAS algorithm is developed. This algorithm searches for the best network configuration by optimizing the topology and the reconfigurable parameters of the photonic devices. The optimized LightPro architecture demonstrates high flexibility, as it can perform unitary transformations based on complex weights trained on datasets.

3. Pruning and Efficiency: A pruning algorithm is employed to remove redundant phase shifters and unnecessary DCs, further improving power efficiency and reducing the number of active devices. After pruning, the system still maintains fidelity in performing MVM tasks, with less than 5% accuracy loss compared to original trained models.

4. Simulation and Experimental Validation: The results are validated through both simulations and experiments. The LightPro network is tested on several datasets (e.g., MNIST), with simulations showing strong agreement with experimental results. The network is implemented using the iPronics SmartLight photonic processor, which uses cascaded MZIs and achieves good agreement with both simulation and experimental results.

5. Advantages over Conventional MZI-based Networks: Compared to traditional Clements networks using Mach-Zehnder interferometers (MZIs), LightPro offers significant improvements in scalability, footprint, and power efficiency. It achieves up to 84% reduction in footprint while maintaining performance.

In conclusion, the paper demonstrates that LightPro, with its programmable SiPh devices and phase-change materials, provides an efficient and scalable solution for photonic AI accelerators. The combination of NAS optimization and pruning allows it to perform complex MVM operations with significantly reduced size and power requirements, making it a promising candidate for next-generation photonic AI hardware.