AMD said an optimized Nunchux AI system using eight AMD Instinct MI355X GPUs delivered a 15-second AI video as a finished MP4 in 6.49 seconds, cutting the wait for users.
For the comparator, SGLang needed 150.5 seconds to produce an MP4. AMD used the same prompt and eight MI355X GPUs in both setups and calculated a 23.2-fold speed advantage for the optimized system. The Oct. 19 result measures the wait until the encoded file is ready.
Nunchux released separate generation measurements Sept. 23, before the encoding step. Its tuned MiniMax-H3 model generated a 15-second clip in 5.39 seconds; Nunchux clocked a five-second sample at 1.33 seconds. The shorter timings included text-prompt processing and decoding of audio and video, but ended before the clips were saved as MP4 files.
AMD said denoising accounted for more than 90% of the baseline’s latency, making it the main focus of the software changes. Denoising involves calculations used to build the frames that make up a video. For the 15-second sample, denoising took 4.04 seconds with Nunchux’s ROCm-optimized MiniMax-H3 setup, compared with 143.1 seconds under SGLang.
AMD also compared its generation times with Sol-H3 results published by NVIDIA Research using eight NVIDIA B300 GPUs. Across five-, 10- and 15-second videos, AMD calculated that its MI355X setup had 18.5% to 19.4% lower generation latency. Neither set of figures includes final MP4 encoding. The measurements came from separate runs with different inference software, so they cannot establish a GPU-only speed advantage.
AMD reported that its Data Center business generated $6.7 billion in the fiscal second quarter of 2026, up 107% from a year earlier. The company plans to report fiscal third-quarter results after the market closes Nov. 3.
