AMD says Nunchux AI video generation outpaces playback on MI355X GPUs
AMD reports Nunchux-optimized MiniMax-H3 generates video faster than real-time on Instinct MI355X hardware.
Why it matters
The demonstration shows full-stack optimization on AMD hardware can achieve real-time video generation latency, which could accelerate deployment of multimodal AI inference workloads. Significance is limited to a single vendor-reported proof of concept without independent verification.
Key facts
AMD published a blog post on October 19, 2026 describing a proof of concept with Nunchux AI for MiniMax-H3 video generation on Instinct MI355X GPUs
Reported by Blockchain.News
Nunchux optimized MiniMax-H3 on a single server with eight AMD Instinct MI355X GPUs
Reported by Blockchain.News
Reported 1.33 seconds of generation latency for a five-second video
1.33 secondsReported by Blockchain.News
Reported 5.39 seconds of generation latency for a 15-second video
5.39 secondsReported by Blockchain.News
With four or eight MI355X GPUs every tested clip length completed faster than it plays
Reported by Blockchain.News
Improvement rose from 14.9x for a five-second clip to 23.2x for a 15-second clip
Reported by Blockchain.News
What happened
AMD published a blog post on October 19, 2026 describing a proof of concept with Nunchux AI showing MiniMax-H3 video generation on Instinct MI355X GPUs completing faster than a clip plays. The company said Nunchux optimized the model on a single server with eight AMD Instinct MI355X GPUs, reporting 1.33 seconds of generation latency for a five-second video and 5.39 seconds for a 15-second video. AMD stated that with four or eight MI355X GPUs, every tested clip length completed faster than it plays. The company compared the Nunchux-optimized stack against an SGLang baseline on the same eight GPUs, reporting the optimized solution reduced 15-second MP4 production from 150.5 seconds to 6.49 seconds, a reported 23.2x speedup, with improvement rising from 14.9x for a five-second clip to 23.2x for a 15-second clip. AMD attributed gains to co-design across model and runtime including custom kernels, memory movement, and execution strategy on AMD ROCm software, and said Nunchux's custom MXFP6 kernel ran about 13x faster than the rocm-libraries MXFP6 baseline, with denoising latency falling by 26.3x to 35.4x.
How the story developed
- First report
First report by Blockchain.News
AMD says Nunchux video AI ran faster than playback on MI355X
- Update
AMD publishes blog post with Nunchux AI proof of concept results
AMD reported MiniMax-H3 video generation faster than playback on Instinct MI355X GPUs
Affected entities
Original sources
| Publisher | Report | Role | Published |
|---|---|---|---|
| Blockchain.News Crypto media | AMD says Nunchux video AI ran faster than playback on MI355X | Original |