{"slug": "nunchux-on-amd-mi355x-5s-minimax-h3-videos-in-1-3s", "title": "Nunchux on AMD MI355X: 5s MiniMax-H3 Videos in 1.3s", "summary": "Nunchux generated a 5-second MiniMax-H3 video in 1.33 seconds on a single server with eight AMD MI355X GPUs, a 21.8× speedup over SGLang on the same hardware, according to the company's benchmark post. Nunchux reported 21.8× to 26.7× speedups across clip lengths of 5.2, 10.1, and 15.0 seconds on one, two, four, and eight MI355X GPUs, with the 15-second clip taking 5.39 seconds on eight GPUs and 10.52 seconds on four. The SGLang baseline ran MiniMax-H3 FL2VA at BF16 precision with 50 Euler steps and AITER dense attention, and timings included text encoding, DiT denoising, and video/audio VAE decoding but excluded final MP4 encoding.", "body_md": "[← Back to blog](https://www.nunchux.ai/blog)Research\n\n# Nunchux on AMD MI355X:\n\n5s MiniMax-H3 Videos in 1.3s\n\nIn our [previous post](https://www.nunchux.ai/blog/attention-is-the-video-bottleneck), we introduced VC-Attention and showed how low-precision attention can accelerate video generation while preserving fidelity. Attention is one part of the Nunchux inference stack. In this post, we show the full stack running MiniMax-H3 on AMD MI355X.\n\nOn a single server with eight MI355X GPUs, Nunchux generates a **5-second video in 1.33 seconds** and a **15-second video in 5.39 seconds**. With four or eight GPUs, every clip length we tested generates faster than it plays.\n\n## Up to 26.7× faster on MI355X\n\nWe measured Nunchux and SGLang on one, two, four, and eight MI355X GPUs. On eight GPUs, Nunchux generates clips in 1.33 to 5.39 seconds, a **21.8× to 26.7× speedup** over SGLang on the same hardware. Four GPUs also generate every tested clip faster than playback, including a 15-second video in 10.52 seconds.\n\nThe SGLang baseline runs MiniMax-H3 FL2VA with BF16 precision, 50 Euler steps, and AITER dense attention.\n\nTiming includes text encoding, DiT denoising, and video/audio VAE decoding. Final MP4 encoding is excluded.\n\n## View all measured results\n\n| Generation latency in seconds for SGLang and Nunchux on AMD MI355X, by clip length and GPU count. |  |  |  |  | \n|---|---|---|---|---|\n| Clip | GPUs | SGLang | Nunchux | Speedup | \n|---|---|---|---|---|\n| 5.2 s | 1 | 182.25 | **7.30** | 25.0× | \n| 5.2 s | 2 | 122.70 | **5.46** | 22.5× | \n| 5.2 s | 4 | 55.89 | **2.50** | 22.4× | \n| 5.2 s | 8 | 28.93 | **1.33** | 21.8× | \n| 10.1 s | 1 | 535.64 | **19.34** | 27.7× | \n| 10.1 s | 2 | 317.52 | **12.80** | 24.8× | \n| 10.1 s | 4 | 148.68 | **6.01** | 24.7× | \n| 10.1 s | 8 | 76.22 | **3.01** | 25.3× | \n| 15.0 s | 1 | 1090.20 | **36.16** | 30.1× | \n| 15.0 s | 2 | 597.82 | **22.65** | 26.4× | \n| 15.0 s | 4 | 283.67 | **10.52** | 27.0× | \n| 15.0 s | 8 | 144.02 | **5.39** | 26.7× | \n\n## Same prompt, both stacks\n\nEach pair uses the same prompt on 8× MI355X GPUs: SGLang on the left, Nunchux on the right. The times below show how long each video took to generate. Across these examples, Nunchux delivers comparable visual quality at much lower latency.\n\n### 5s video\n\n## Show the prompt\n\nIntegrated multimodal description\n\n[Shot 1] Cinematic, MCU, the camera performs a slow Tilt Up from the neck of a vibrant pink flamingo with delicate, slender feathers to its head. The flamingo is positioned in the midground against the soft-focus bokeh of a harbor at twilight, where a distant ship's deck is illuminated by warm, glowing practical lamplight reflecting off the dark water. The flamingo abruptly tilts its head once to the side, peering curiously into the lens with a piercing black eye, before leveling its gaze back to the horizon. In the background, wooden pilings and thin, shimmering ripples of light complete the serene, humid coastal atmosphere.\n\nOverall soundscape\n\nThe faint, rhythmic lapping of gentle harbor water against wooden piers is audible in the foreground, accompanied by the distinct, subtle click-clack of the flamingo's beak shifting as it tilts. A distant, muffled nautical bell tolls once, adding depth to the ambient coastal environment.\n\nNon-diegetic music\n\nA solo, minimalist cello melody plays at a slow, deliberate tempo, utilizing long, sustained low notes that fade into the ambient background.\n\n**SGLang** · 28.93 s\n\n**Nunchux** · 1.33 s\n\n(21.8× faster)\n\n### 10s video\n\n## Show the prompt\n\nIntegrated multimodal description\n\n[Shot 1] Cinematic, a medium tracking shot moving steadily forward through a vast field of tall, feathery wild grass at blue hour. The lighting is dominated by a deep, cool ambient twilight, casting a serene indigo hue across the landscape with soft, diffused shadows and a very faint, lingering cool glow on the horizon. A delicate, translucent silver silk scarf drifts gracefully through the air in the center of the frame, carried by a continuous gentle breeze. The fabric undulates smoothly, folding and unfolding onto itself, its sheer material catching the faint residual light from the sky above. The scarf dances just above the grass stalks, dipping slightly to brush against the swaying, pale green tips before catching an updraft and lifting slightly higher. The tall grass sways rhythmically in unison beneath it. The camera fluidly follows the scarf's horizontal and vertical movements, maintaining the silver fabric in the midground as it glides further into the darkening, windswept field.\n\nOverall soundscape\n\nA constant, soothing rustling of tall grass stalks rubbing against one another, blending with a smooth, sweeping wind blowing steadily in the background. A subtle, soft fluttering sound of lightweight fabric is clearly heard in the foreground each time a stronger gust catches the silk.\n\nNon-diegetic music\n\nA slow, ethereal ambient synthesizer pad playing sustained, low-register chords, accompanied by a very soft, sparse cello drone underneath.\n\n**SGLang** · 76.22 s\n\n**Nunchux** · 3.01 s\n\n(25.3× faster)\n\n### 15s video\n\n## Show the prompt\n\nIntegrated multimodal description\n\n[Shot 1] Cinematic, Medium Shot, the camera pedestal up slowly to capture a middle-aged potter (S1) with gray-streaked hair and dust-covered hands, wearing a coarse, charcoal-colored linen apron over a faded navy t-shirt. He stands before a rustic wooden kick-wheel in a dimly lit, high-ceilinged industrial studio, his body framed against a backdrop of raw, textured gray concrete walls. Natural, harsh top-down sunlight filters through a high, barred window, casting sharp, dramatic shadows across his focused face and the spinning lump of wet, malleable clay on the wheel. He centers the clay with rhythmic, deliberate pressure, his knuckles white against the dark mud. [Shot 2] At 00:07.500, the camera cuts to a Close-up, static shot, focusing on the potter's (S1) hands as he expertly pulls the clay upwards, the wet surface reflecting the harsh light. He pinches the rim, forming a slender, elegant neck, while the sound of the wheel's wooden frame groaning fills the space; he maintains his steady, rhythmic pace, his breath audible and controlled as he shapes the damp vessel.\n\nOverall soundscape\n\nThe rhythmic, heavy creak of a wooden kick-wheel dominates the space, accompanied by the wet, squelching sound of clay being squeezed and shaped by human hands. A soft, continuous ambient hum of wind blowing against the concrete exterior is clearly heard in the background, punctuated by the sharp, metallic clatter of a metal tool dropped onto a wooden work surface.\n\nNon-diegetic music\n\nN/A\n\n**SGLang** · 144.02 s\n\n**Nunchux** · 5.39 s\n\n(26.7× faster)\n\n## Video you can steer as you watch\n\nThe [demo](#demo) above runs on eight MI355X GPUs, and Nunchux generates video faster than it plays. After the first segment is ready, the demo keeps generating the next while you watch. You can change the prompt during playback to decide what happens next. The next segment follows your input, so the story can change direction as you watch.\n\n## Built for speed, from models to kernels\n\nNunchux combines post-training, model optimization, and inference engineering to accelerate video generation across GPU platforms. We develop the **Nunchux Model Optimizer** and the **Nunchux Inference Engine** together, adapting the model and runtime to the hardware.\n\nWe chose MI355X for its strong compute performance, 288 GB of HBM3E memory, and native MXFP6 support. MXFP6 delivers [up to 10.1 PFLOPS of dense matrix compute per GPU](https://www.amd.com/en/products/accelerators/instinct/mi350/mi355x.html), roughly twice the FP8 peak. This combination makes MI355X a good fit for large multimodal models.\n\nOur custom MXFP6 kernel runs **about 13× faster** than the [rocm-libraries](https://github.com/ROCm/rocm-libraries/blob/rocm-7.2.0/projects/hipblaslt/library/src/amd_detail/rocblaslt/src/rocroller/rocroller_host.cpp#L314-L324) MXFP6 baseline in our tests. This contributes to the video results above alongside our other model and inference optimizations.\n\n## One stack, across hardware\n\nMI355X is one example of how Nunchux combines model optimization and inference systems to make frontier multimodal models fast across hardware. From AMD and NVIDIA GPUs to edge accelerators, our goal is the same: generate content faster than it is consumed.\n\nWe are grateful to [Victor Robles, PhD](https://rocm.blogs.amd.com/authors/victor-robles.html) at AMD for provisioning and supporting our MI355X environment throughout the POC.\n\n[Try the models in Nunchux Modelverse](https://www.nunchux.ai/modelverse). Free access to MiniMax-H3 arrives very soon. Join the waiting list today to be among the first to use it.\n\nIf you run visual models at scale, contact [sales](https://www.nunchux.ai/enterprise#contact). We are hiring. Please visit our [careers page](https://www.nunchux.ai/careers) for more details.", "url": "https://wpnews.pro/news/nunchux-on-amd-mi355x-5s-minimax-h3-videos-in-1-3s", "canonical_source": "https://www.nunchux.ai/blog/video-generation-on-amd-mi355x", "published_at": "2026-09-23 16:34:14+00:00", "updated_at": "2026-09-23 17:00:53.832156+00:00", "lang": "en", "topics": ["generative-ai", "ai-infrastructure", "ai-chips", "ai-research"], "entities": ["Nunchux", "AMD MI355X", "MiniMax-H3", "SGLang", "VC-Attention", "AITER"], "alternates": {"html": "https://wpnews.pro/news/nunchux-on-amd-mi355x-5s-minimax-h3-videos-in-1-3s", "markdown": "https://wpnews.pro/news/nunchux-on-amd-mi355x-5s-minimax-h3-videos-in-1-3s.md", "text": "https://wpnews.pro/news/nunchux-on-amd-mi355x-5s-minimax-h3-videos-in-1-3s.txt", "jsonld": "https://wpnews.pro/news/nunchux-on-amd-mi355x-5s-minimax-h3-videos-in-1-3s.jsonld"}}