A large scale photonic matrix processor enabled by charge accumulation.

Abstract in brief

• The article discusses the development of large-scale photonic matrix processors designed to efficiently run complex artificial neural networks by leveraging the high bandwidth and low loss of photonic structures, while addressing scaling challenges.
• It introduces a novel time multiplexed processing scheme that virtually enlarges the physical photonic array without additional electrical processing, enabling higher matrix sizes and more efficient optical accumulation with high accuracy.
• The proposed architecture could perform over 110 trillion multiply-accumulate operations per second, handling matrix-vector multiplications involving up to 16,000 matrices, all within a compact area of 512mm², representing a significant advancement in photonic neural network hardware.