Google DeepMindAnalysisAI ModelsSeptember 30, 2026

Google details Sparse VideoGen attention optimization for TPUs

Read original source →developers.googleblog.com

Google Developers Blog describes implementing Sparse VideoGen (SVG) on TPUs, routing attention heads to spatial or temporal sparse masks to cut video diffusion latency. Attention's share of per-layer latency grows from 55.5% to 88.2% scaling 720p to 1440p, with 81 frames of 720p spanning 50K-400K sequence length.

1 source

More stories today

Open the live feed