Nunchux AI launches vc-attention: a new low-bit attention kernel for video diffusion models

Nunchux AI has unveiled VC-Attention, a novel training-free low-bit attention kernel designed to enhance the efficiency of video diffusion transformers (DiTs). The technology addresses two key challenges: value quantization error and the slow softmax stage in attention computations. By leveraging low-bit tensor cores, VC-Attention accelerates the matrix products QK and PV, which are critical for attention operations. The method introduces V-Smooth, which reduces block energy in sequence order, and ExpCast-FP8, which directly writes bytes from the log-domain score with a fused multiply-add operation. These innovations significantly improve performance, with VC-Attention achieving 6.02× speedup over SageAttention2 on B200 and 1.16× on H200. The technique also maintains high fidelity, with PSNR scores of 20.2 dB on MiniMax-H3. Nunchux Attention, the company's proprietary extension, further boosts performance on B200 and B300. The method is compatible with sparse attention and multi-GPU execution, and the company is offering free access to MiniMax-H3 through its Modelverse waitlist.