Researchers proposed VC-Attention to improve low-bit attention for video generation. It uses Value smoothing to group tokens for better quantization. It also uses ExpCast-FP8 to skip the high-precision softmax step. The method works without retraining. It was tested on B200, B300, H200, RTX PRO 6000, and RTX 5090 GPUs. It improved fidelity over low-bit baselines. It sped up attention kernels by 1.46-1.59x on datacenter GPUs. It sped up by 2.3-3.6x on workstation GPUs. It generated clips 1.13-1.19x faster on datacenter hardware. It generated clips 1.36-1.70x faster on workstation hardware.
What changed
Low-bit attention kernels were slower and less accurate due to value outliers and high-precision softmax.
What it unlocks
Faster and more accurate low-bit attention for video generation models on multiple GPU architectures.
- 1.46-1.59x faster on datacenter GPUs
- 2.3-3.6x faster on workstation GPUs
- 1.13-1.19x faster end-to-end on datacenter
- 1.36-1.70x faster end-to-end on workstation
Sources