Integrated Photonic Convolution Accelerator Empowered by Thin-Film Lithium Niobate Modulators

Abstract

The current optical convolution architectures are facing challenges related to limited scalability, excessive data redundancy and restricted processing bandwidth. In this work, we introduce an integrated photonic convolution accelerator (IPCA) empowered by high-speed thin-film lithium niobate (LN) modulators. Consequently, data replication redundancy is free and Fourier transform is avoided, which paves the way for highly efficient convolution scaling. We implement convolution in an optical frequency spacing of 8 GHz with a power consumption of only 52.9 mW. Optical neural networks of different parameter sizes are constructed across various complex tasks. Given its advantages to address energy consumption and computing power challenges inherent to current AI advancements, our method heralds a pivotal shift in upcoming optical computing hardware architectures.