Nunchux on AMD MI355X: 5s MiniMax-H3 Videos in 1.3s (www.nunchux.ai)

Nunchux has achieved a breakthrough in video generation using AMD MI355X GPUs, processing a 5-second video in just 1.33 seconds and a 15-second video in 5.39 seconds, significantly faster than existing solutions. This development showcases the effectiveness of low-precision attention techniques, as previously introduced with their VC-Attention framework. On a configuration of eight MI355X GPUs, Nunchux demonstrated speedups of 21.8× to 26.7× compared to the SGLang baseline, effectively generating video clips at a rate faster than real-time playback. The significance of this advancement lies in Nunchux's ability to optimize multimodal content generation across various hardware platforms. Utilizing the MI355X's robust computing power and large memory capacity, Nunchux's custom MXFP6 kernel operates approximately 13× faster than traditional libraries. This combination of software and hardware optimization means that Nunchux can effectively handle large models and generate high-quality video content efficiently, paving the way for faster, more responsive AI-driven media generation across diverse applications.