Photonics and microwaves merge to improve computing flexibility

Abstract

In artificial neural networks, data structures usually exist in the form of vectors, matrices, or higher-dimensional tensors. However, traditional electronic computing architectures are limited by the bottleneck of separation of storage and computing, making it difficult to efficiently handle large-scale tensor operations. The research team has developed a photonic tensor processing unit based on a single microring resonator, which performs tensor convolution operations in multiple dimensions of time, wavelength, and microwave frequency by precisely adjusting the operating state of multi-wavelength lasers. This innovative design increases the photonic computing density to 34.04 TOPS/mm², significantly surpassing the performance level of existing photonic computing chips.