Preprint claims 1.5x faster video diffusion attention on Blackwell and Hopper
single source· 1 articles · confidence: medium · first seen 2026-09-13 20:00 UTC
What this means for you
Nothing to adopt yet. The abstract describes the kernel but links no code and gives no evaluation date, so the 1.5x figure is the authors' own measurement. If your video pipeline currently leans on BF16 FlashAttention-4 on Hopper or Blackwell, this is the number to watch.
A preprint reports VC-Attention, a training-free low-bit attention method for video diffusion transformers that speeds the attention kernel — where every token is compared against every other, and the dominant cost at long video lengths — by 1.46–1.59x over BF16 FlashAttention-4 on datacenter Blackwell and Hopper GPUs, and 2.3–3.6x on workstation cards, with end-to-end clip generation 1.13–1.19x and 1.36–1.70x faster. V-Smooth reorders value tokens by online clustering so a hardware block quantises well; ExpCast-FP8 maps log-domain scores straight to E4M3 probability codes, removing the FP32 exponential. Across Wan2.2, LongCat-Video, HunyuanVideo-1.5 and MiniMax-H3 it beats low-bit baselines on fidelity. The paper gives no evaluation date for those figures.
Key facts
- ·VC-Attention is described as a training-free low-bit attention method for video diffusion transformers, implemented for B200, B300, H200, RTX PRO 6000 and RTX 5090. source
- ·Reported attention-kernel speed-up over BF16 FlashAttention-4 is 1.46–1.59x on datacenter Blackwell and Hopper GPUs. source
- ·On workstation cards the reported attention-kernel speed-up is 2.3–3.6x. source
- ·End-to-end clip generation is reported 1.13–1.19x faster on datacenter GPUs and 1.36–1.70x faster on workstation cards. source
- ·Evaluated across Wan2.2, LongCat-Video, HunyuanVideo-1.5 and MiniMax-H3, where the authors report improved fidelity over low-bit baselines. source
- ·The abstract gives no evaluation date for the reported speed-ups. source
What the sources say
- Hugging Face Daily Papers (research) — Abstract only: kernel and end-to-end speed-ups across five GPUs and four video models, no evaluation date.
Sources
The original reporting. Follow these — they did the work.
- Hugging Face Daily PapersVC-Attention: Value Smoothing and Softmax Casting for Low-bit Attention2026-09-13