RX 6950 XT detected as ROCm-compatible, but ROCm installation is not possible

Software Siril
1

Hello,

I would like to ask whether anyone has experience with DirectML GPU acceleration on an AMD RX 6950 XT.

System

  • Windows 10 22H2 (Build 19045)
  • AMD Radeon RX 6950 XT
  • Siril 1.4.3
  • SyQon Siril Edition 2.0.1

ROCm recommendation

The GPU Framework Manager identifies my GPU as:

Primary GPU: amd_rocm
GPU Name: AMD Radeon RX 6950 XT
ROCm Compatible: True

and recommends the ROCm backend.

However, the recommended ROCm installation fails because no matching PyTorch package can be found:

ERROR: Could not find a version that satisfies the requirement torch
ERROR: No matching distribution found for torch

As far as I understand, ROCm on Windows appears to be limited to newer AMD hardware, so I switched to DirectML instead.

DirectML results

After rebuilding the Python environment and installing:

  • torch 2.4.1
  • torchvision 0.19.1
  • torch-directml 0.2.5.dev240914

DirectML is detected correctly and GPU inference works.

The PyTorch GPU test reports:

Using DirectML
Testing GPU execution...
OK: GPU execution successful!

Problem

While GPU processing works, the output is corrupted.

I tested several SyQon models:

  • Prism Deep → black vertical tile/stripe
  • Axiom / Starless → black tile artifact
  • Parallax → circular corruption artifacts

When I enable CPU mode, all artifacts disappear and the same images process correctly.

For Prism Deep, the width of the black stripe changes with tile size, which may indicate a DirectML, PyTorch DirectML, or tile-processing related issue.

Questions

  • Is the ROCm recommendation for an RX 6950 XT on Windows 10 expected?
  • Is DirectML currently the preferred GPU backend for RX 6000 series cards?
  • Has anyone with an RX 6800 XT, RX 6900 XT or RX 6950 XT experienced similar DirectML artifacts?

Thanks!

Sabine

2

It is probably best to seek support for proprietary software from the vendor themselves.

3

Thanks for the reply.

I have already opened a separate support ticket with SyQon regarding the DirectML image artifacts.

My main question here is actually about the GPU Framework Manager itself.

The manager detects my RX 6950 XT as ROCm-compatible and recommends the ROCm backend, but the recommended ROCm installation fails because no matching PyTorch package can be found.

I was therefore wondering:

  • Is the ROCm recommendation for an RX 6950 XT on Windows 10 expected?
  • Is DirectML currently the preferred GPU backend for RX 6000 series cards under Windows?
  • Or is the ROCm detection/recommendation perhaps too optimistic for this hardware/OS combination?

I am mainly trying to understand whether the behavior of the GPU Framework Manager is expected.

Thanks!

4

The GPU manager’s detection might well be inaccurate here. Torch-ROCm on Windows is not in a great place at present, with only very new cards supported in an experimental version that AMD distribute directly outside of pypi. This is an issue that AMD needs to resolve (there are suggestions that things may improve, but we will just have to wait and see). The GPU manager mostly relies on detection routines that are baked into the gpuhelper submodule of sirilpy: at first I was not aware of the issues with that combination (as I don’t use Windows or an AMD GPU myself), so the detection is currently handled by a hacky bit of the GPU manager script and may well not be 100% reliable especially as the situation evolves.

However if there is a torch-directml backend and your GPU supports DirectML then the script should work. (I don’t know if there are any known issues with torch-directml on AMD hardware though.)

Have you tried the CosmicClarity-Native or the SCUnet denoise scripts? Those also use Torch, so if you test those with your GPU and your torch-directml installation then it would give an idea of if it’s likely to be a model bug (if the other scripts work OK) or if it’s likely to be a torch-directml bug (if all the scripts produce garbage).

Hopefully as things settle down with Windows ROCm support I will be able to update the GPU_Manager script to handle this situation a bit better.

But in answer to your question, it appears that at the present time yes, DirectML should be the appropriate backend for your RDNA2 card.

5

Thanks for the clarification.

I did some additional testing.

I used and tested torch-directml via CC as suggested.

  • Cosmic Clarity Denoise v6.6 (ONNX/DML backend)

The script runs correctly on my RX 6950 XT using DirectML, processes the image very quickly, and produces a clean output without any artifacts.

The log reports:

“Loaded mono model → backend=ONNX/DML”

and the resulting image looks completely normal.

By contrast, the SyQon models still show GPU-related corruption:

  • Prism Deep → vertical black stripe/tile artifacts
  • Starless (Axiom 2.1) → black tile artifacts
  • Parallax → circular corruption artifacts

When CPU mode is enabled, the same SyQon models work correctly.

This makes me suspect that DirectML itself is functioning correctly on my system and that the issue may be related to the specific SyQon models, their tiling implementation, or their use of PyTorch DirectML.

I have already opened a support ticket with SyQon and provided screenshots and logs.

So at this point it looks like:

  • ROCm is currently not usable on Windows 10 with my RX 6950 XT.
  • DirectML works in general.
  • The corruption appears to be specific to the SyQon GPU implementation rather than a general DirectML failure.

Thanks again for the explanation regarding ROCm detection.