AMD says Nunchux video AI ran faster than playback on MI355X

News

AMD reported Nunchux optimized MiniMax-H3 video generation on eight Instinct MI355X GPUs to finish clips faster than playback.

AMD said a proof of concept with Nunchux AI showed MiniMax-H3 video generation on Instinct MI355X GPUs could complete faster than a clip plays, highlighting how full-stack optimization can reduce latency in multimodal inference. The company published the results in a blog post dated Oct. 19, 2026.

Evidence and context

In AMD’s account, Nunchux optimized MiniMax-H3 on a single server with eight AMD Instinct MI355X GPUs and reported 1.33 seconds of generation latency for a five-second video and 5.39 seconds for a 15-second video. AMD said that with 4 or 8 MI355X GPUs, every tested clip length completed faster than it plays, which the company framed as real-time video generation because the model stays ahead of the playhead once playback begins.

AMD also said the same prompt was run through the Nunchux-optimized stack and an SGLang baseline on the same eight MI355X GPUs. On that basis, the company said the optimized solution reduced the time to produce a finished 15-second MP4 from 150.5 seconds with the baseline to 6.49 seconds, a reported 23.2x speedup. AMD said the improvement rose from 14.9x for a five-second clip to 23.2x for a 15-second clip.

The company attributed the gains to co-design across the model and runtime, including custom kernels, memory movement and execution strategy on AMD ROCm software. AMD said Nunchux’s custom MXFP6 kernel ran about 13x faster than the rocm-libraries MXFP6 baseline in its testing, and that denoising latency fell by 26.3x to 35.4x across the measured clip lengths. For the 15-second clip, AMD said denoising time dropped from 143.1 seconds to 4.04 seconds.

AMD positioned the result as an example of what its accelerator and software stack can support for deployable AI services, not as a general claim about all workloads. The post said diffusion-based video models put heavy pressure on compute and memory, and noted that MI355X was selected for large high-bandwidth memory, low-precision datatype support and an open ROCm software stack that allowed deeper customization.

The same post also compared Nunchux’s MI355X result with NVIDIA Research’s separately published Sol-H3 results on eight NVIDIA B300 GPUs. AMD said the Nunchux solution reported 18.5% to 19.4% lower generation latency across 5-, 10- and 15-second outputs, using a shared scope of 1344×768 video at 24 frames per second with text encoding, diffusion transformer denoising, and video and audio VAE decoding. AMD said final MP4 encoding was excluded from both sets of results for that comparison.

The post follows other recent AMD benchmark and technical articles on MI355X, including a Sept. 24, 2026 report on 469 tokens per second per user in Million-Token Agent Sessions and a Sept. 16, 2026 MLPerf Inference 6.1 submission. Those earlier reports covered different workloads and should not be read as direct evidence for video generation performance.