DiffLight: Generative Diffusion Model Acceleration With Silicon Photonics

Diffusion models are a powerful class of generative AI, with an inherent capability to generate highly realistic synthetic data. However, their iterative denoising process and complex underlying architecture, typically involving UNets and attention mechanisms, result in high inference latency and computational overheads. To address these challenges, we introduce DiffLight, the first silicon-photonic accelerator for generative diffusion models. Our analysis demonstrates that DiffLight exhibits at least 5.5× higher GOPS and 3× lower energy-per-bit (EPB), compared to state-of-the-art diffusion model accelerators.