Aivora

#pallas

Pallas

1 article

Accelerating Video Diffusion on TPUs: Optimizing Spatio-Temporal Sparse Attention
Google AI DevelopersVideo AI

Accelerating Video Diffusion on TPUs: Optimizing Spatio-Temporal Sparse Attention

Google TPU 影片生成加速:如何透過「時空稀疏注意力」實現 1.69 倍效能提升

This explainer details how optimizing block tiling and memory layouts on TPUs translates theoretical video diffusion sparsity into a 1.69x end-to-end speedup for 2K video generation.

2 min read