# Realtime AI video on AMD GPUs: reading MI355X benchmarks

> Read Windflow's AMD MI355X video inference measurements with the right context: GPU count, resolution, generated windows, edit latency and runtime limits.

Source: https://windflow.online/guides/realtime-video-generation-amd-gpus · Updated 2026-10-11 · Windflow Labs

Realtime video workloads can run on AMD GPUs when their model and serving path support the runtime. Windflow publishes measured examples on **AMD Instinct MI355X**, including video with sound and live video editing. Use the [public measurement report](https://stayinalive181.github.io/world-models-on-mi355x/) for the experiment details, and the [model catalog](https://windflow.online/models) for current API capabilities.

These are measurements of specific workloads and configurations. They do not establish that every video model runs faster on AMD, or that a session you create will use the same GPU. Live readiness and placement are operational state rather than benchmark results.

## Measured workloads

| Workload | Published Windflow measurement | Configuration context |
|---|---|---|
| MiniMax H3 video with sound | A 14.375-second video window generated in 9.30 seconds; approximately 1.5 times realtime | 960 × 544, four AMD MI355X GPUs |
| JoyAI live video editing | 443 ms median camera-frame-to-edited-frame latency; 578 ms p90 | 1248 × 720 at 24 fps, one AMD MI355X GPU |
| LingBot-World v2 | 16.3 generated fps, above its 16 fps clock | 704 × 384, one AMD MI355X GPU |

The rows describe different jobs. Window throughput measures how quickly a block of generated audio and video completes. Camera edit latency measures the delay between an incoming frame and its edited result. World-model FPS measures the output cadence. Dividing or ranking those numbers against each other would not produce a useful comparison.

## What faster than realtime means

For a window-generating model, compare the output duration with generation time. If a 14.375-second window takes 9.30 seconds to generate, it can replenish playback faster than that window is consumed under the measured configuration.

That ratio does not make the first frame immediate. The first window still has to be generated, and a prompt change can wait until the next supported boundary. A product can have smooth sustained playback and a substantial control delay at the same time.

## Evaluate ROCm support at the model level

Before choosing a GPU, identify the exact code revision, dependencies, attention implementation, precision, compilation settings and decoder path. Record GPU count and output resolution. Repeat long enough to catch recompilation, memory growth and quality degradation.

A port that loads weights is only a starting point. Your acceptance criteria should also include stable video, supported controls, sustained output and successful recovery. The [model intake page](https://windflow.online/bring-a-model) asks authors for a reproducible baseline and hosting permission before integration.

## Compare against NVIDIA fairly

Use the same model revision, resolution, input sequence, generation settings, output duration and quality criteria. Report the number of GPUs and the client path. If one side uses interpolation or a different serving tier, label it explicitly rather than calling delivered frame rate native generation speed.

This page does not publish a blanket H100-versus-MI355X ranking. Follow the report's specific configurations and measure the experience you plan to ship. The [latency measurement guide](https://windflow.online/guides/measure-realtime-video-latency) explains how to separate generation, transport and visible response.

## Use the API without coupling your product to a GPU

Windflow exposes the session contract and each model's controls to the application. Choose the model by its input and output requirements, then validate readiness, responsiveness and cost. For a custom GPU or serving profile, discuss the workload through [Bring a model](https://windflow.online/bring-a-model).

## FAQ

### Can realtime video generation run on AMD MI355X?

Windflow has published specific MI355X video-generation and editing measurements. Support depends on the exact model and serving configuration; a benchmark is not a guarantee for every model or session.

### Is the 1.5 times realtime measurement a first-frame latency?

No. It compares a generated 14.375-second H3 window with 9.30 seconds of generation time on four MI355X GPUs. First-frame and control-to-visible latency need separate measurements.

### Where are the benchmark sources?

The [Windflow public measurement report](https://stayinalive181.github.io/world-models-on-mi355x/) contains the underlying performance write-up. Current hosted capabilities and rates are on the model and pricing pages.