Reconfigurable parallel photonic matrix-vector multiplication processor based on multi-dimensional multiplexing

Area of Science:

• Photonics
• Optical Computing
• Signal Processing

Background:

• Matrix-vector multiplication (MVM) is crucial for data processing and artificial neural networks.
• Existing photonic MVM architectures face limitations in parallelism and flexibility.
• Growing demand for computing power necessitates advanced MVM solutions.

Purpose of the Study:

• To propose a novel reconfigurable parallel MVM (RP-MVM) processor architecture.
• To enhance the parallelism and flexibility of photonic MVM systems.
• To demonstrate the feasibility of the RP-MVM for high-performance computing tasks.

Main Methods:

• Incorporation of wavelength division multiplexing (WDM) and digital subcarrier multiplexing (DSM) into a photonic MVM architecture.
• Development of a reconfigurable parallel MVM (RP-MVM) scheme.
• Dynamic adjustment of input data channels without altering hardware scale.

Main Results:

• The RP-MVM scheme increases parallelism by N times compared to traditional WDM-MVM.
• Achieved parallel computing of eight MVMs with a speed of 128 GOPs.
• Demonstrated a root mean square error (RMSE) of 1E-3 for 6-bit data and parallel processing of four grayscale images for edge extraction.

Conclusions:

• The proposed RP-MVM processor offers a significant advancement in photonic computing.
• The architecture provides a highly parallel and reconfigurable solution for large-scale MVM.
• This scheme presents a viable alternative for future high-performance optical computing systems.