SpawnItems › Guides
RX 9070 XT crashes in ComfyUI and AI video renders – cause and fix
"unspecified launch failure" after 15 minutes of rendering, while every game runs fine? Here is why, and how to fix it for good.
unspecified launch failure or a driver timeout – often after many minutes and often in the frame-interpolation step – the most common cause is an undervolt or VRAM overclock that is stable in games but not under long AI load. Reset to factory to confirm, update the driver, then re-tune with a test that includes long AI load.The symptoms
- The render runs for 10–20 minutes, then ComfyUI disconnects or the screen flickers and the driver resets.
- The log shows
HIP error: unspecified launch failure,hipErrorLaunchFailureor "device lost". - It often happens when a new model is loaded mid-job – for example the frame-interpolation model (FILM or RIFE) after the video has been sampled.
- Games, 3DMark and short stress tests run fine with the same settings.
Why it happens
AI video jobs combine everything that exposes a marginal undervolt: VRAM nearly full for a long time, sudden load jumps between model stages, and memory-heavy convolution work. I had exactly this on my own RX 9070 XT: a 16-minute video render crashed twice at the interpolation step, with settings that had passed every game test.
The fix, step by step
- Confirm the cause: reset GPU tuning to factory and run the same render once. If it finishes, the tuning was too tight.
- Update the driver: Adrenalin 26.11.2 (18 Nov 2026) adds ROCm 10.2 support and stability fixes.
- Re-tune with an AI-grade test: the Sweet-Spot Tuner ends every run with an AI soak – near-full VRAM, FP16 matrices, load steps, RAM↔VRAM transfers and a video-interpolation phase – and adds reserve if a round fails. Use the Rock-Solid preset if you render long videos.
- Protect long renders: save the sampled video before the interpolation step, so a crash there does not cost the whole render. (My local AI studio now does this automatically and falls back from FILM to RIFE.)
Bonus: ROCm 10.2 is noticeably faster on RDNA 4
With driver 26.11.2 I moved ComfyUI from a ROCm 7.13 preview build to PyTorch with ROCm 10.2. Same card, same tuning:
| Operation (RX 9070 XT) | ROCm 7.13 | ROCm 10.2 |
|---|---|---|
| Attention (core of image and video models) | 16.9 ms | 9.3 ms (−45 %) |
| Convolution (VAE, interpolation) | 2.7 ms | 2.2 ms (−19 %) |
| Matrix multiply FP16 / BF16 | 92 / 93 TFLOPS | 97 / 99 TFLOPS |
| Full Wan 2.2 14B video test | 265 s | 234 s (−12 %) |
Wheels from AMD's index stable.repo.amd.com/rocm/whl-next/ (torch 2.14.0+rocm10.2.0, Python 3.12). Install into a copy of your environment first, so you can switch back.
Frequently asked questions
What does 'unspecified launch failure' mean on an AMD GPU?
The GPU stopped executing a compute kernel and the driver reset it. On tuned cards the usual cause is a voltage or memory clock that is too tight for long AI load; on stock cards it can be a driver bug.
Is my RX 9070 XT broken?
Almost certainly not. If the render finishes at factory settings, the card is fine – the tuning needs more reserve.
Does ROCm 10.2 work on Windows with the RX 9070 XT?
Yes. With Adrenalin 26.11.2 I run PyTorch 2.14 with ROCm 10.2 on Windows on an RX 9070 XT without errors.
Which undervolt preset should I use for AI work?
Rock-Solid. It keeps the largest reserve and tests 5 minutes of AI load per round.
RDNA 4 Sweet-Spot Tuner
Free, open source (MIT), Windows 10/11 · Linux: command line