InfraSep 30, 2026 · 23:521 source

Accelerating Spatio-Temporal Attention for Video Diffusion on TPUs

From developers.googleblog.com

Image: developers.googleblog.com
AI;DR

Developers implemented Sparse VideoGen (SVG) to address the quadratic latency bottleneck of self-attention in high-resolution video diffusion models by routing attention heads to structured spatial or temporal sparse masks. Optimizations to the Splash Attention kernel on TPUs, including bypassing empty memory tiles and permuting token memory layouts into temporal-major order, achieved up to a 1.69x end-to-end inference speedup for 1440p video generation. The approach aligns sparse masks with hardware tile execution to reduce wasted matrix operations.

Read the sourceSign in to suggest

Sources

#google#inferenceMarkdown
Content history (1)
  • 2026-10-04 · Title · vi · Wording fix
    Tối Ưu Hóa Attention Thời Gian Không Gian Cho Video Diffusion Trên TPU
    Tối ưu hóa attention thời gian không gian cho video diffusion trên TPU

One ranked edition, every morning

The stories that matter today. Free, cancel anytime.

Telegram @aidr_today