DOTA: A Dynamically-Operated Photonic Tensor Core for Energy-Efficient Transformer Accelerator
share
The wide adoption and significant computing resource consumption of attention-based Transformers, e.g., Vision Transformer and large language models, have driven the demands for efficient hardware accelerators. While electronic accelerators have been commonly used, there is a growing interest in exploring photonics as an alternative technology due to its high energy efficiency and ultra-fast processing speed. Optical neural networks (ONNs) have demonstrated promising results for convolutional neural network (CNN) workloads that only require weight-static linear operations. However, they fail to efficiently support Transformer architectures with attention operations due to the lack of ability to process dynamic full-range tensor multiplication. In this work, we propose a customized high-performance and energy-efficient photonic Transformer accelerator, DOTA. To overcome the fundamental limitation of existing ONNs, we introduce a novel photonic tensor core, consisting of a crossbar array of interference-based optical vector dot-product engines, that supports highly-parallel, dynamic, and full-range matrix-matrix multiplication. Our comprehensive evaluation demonstrates that DOTA achieves a >4x energy and a >10x latency reduction compared to prior photonic accelerators, and delivers over 20x energy reduction and 2 to 3 orders of magnitude lower latency compared to the electronic Transformer accelerator. Our work highlights the immense potential of photonic computing for efficient hardware accelerators, particularly for advanced machine learning workloads.
READ FULL TEXT