Coherent-Lightning: A Photonic-Electronic AI Accelerator Facilitating Dynamic Real/Complex-Valued Matrix Multiplication

The rapid expansion of artificial intelligence (AI) technologies, particularly in fields such as autonomous driving, image generation, and natural language processing, has driven the need for higher computational performance. Traditional digital circuits face limitations due to the slowdown of Moore's Law, necessitating alternative approaches to meet the growing computational demands. Photonic computing has emerged as a promising solution, offering fast and efficient operations with ultra-low power consumption and high bandwidth (up to pJ/MAC). However, current photonic architectures are often inflexible and limited in supporting dynamic matrix multiplications, which is crucial for AI models like Transformers. In this work, we propose a coherent photonic-electronic computing core (PECC) that overcomes these limitations by enabling operations in both amplitude and phase dimensions of the optical signal. PECC allows 2-D vector dot productions on a single wavelength and supports complex-valued multiplication operations. This core is scalable through wavelength division multiplexing (WDM) and crossbar arrays, forming the basis of a photonic-electronic tensor core (PETC). Multiple PETCs can be integrated into a Photonic-Electronic Streaming Multiprocessor (PESM) for large-scale AI computations. We experimentally demonstrate the effectiveness of the PECC and numerically simulate the performance of the PESM. For PECC, we experimentally achieved an accuracy over 5 bit at a computing speed of 32 GOPS and an accuracy of 90.0% and 91.2% on the Fashion-MNIST dataset and comparable accuracies on the Cifar-10 dataset using it as the real-valued and complex-valued neural networks. For PESM, we achieved an inference accuracy of 89.47% for the Cifar-100 dataset using Deit and could achieve a theoretical computing speed of up to 1048 TOPS in a simulated environment.