Flexible testbed for prototyping photonic processors

The demand for neuromorphic photonic systems is hindered by the high development costs and inflexibility of traditional photonic integrated circuits (PICs) once fabricated. This work introduces a reconfigurable testbed for prototyping photonic processors, featuring modular, hot-swappable components and standardized connectivity. The system uses fiber-coupled architecture with MEMS-based variable optical attenuators (VOAs) for rapid weight adjustments, balanced photodetectors for signal integration, and wavelength-division multiplexing (WDM) for parallel processing. Calibration ensures 5.85 bits of precision, enabling high-fidelity operations. Experimental validation shows the testbed's versatility, achieving low root mean square error (RMSE) in finite impulse response (FIR) filtering and 2D image convolutions, and 75% accuracy on ImageNet subsets (comparable to digital) when integrated into a ResNet-18 model. The platform's scalability and cost-effectiveness make it a useful addition to accelerating photonic processor design and short-term deployments of real-time signal processing and AI inferencing.