Short answer
Radeon RX 7000 and RX 9000 cards run ComfyUI on Windows through ComfyUI Desktop or the AMD portable build, both on AMD’s ROCm, and on Linux with ROCm 7.2. RX 6000 cards need a community build. What fits is set by the memory, as on NVIDIA; the speed, and how well ComfyUI juggles memory, still trail.
An RX 9070 XT or 7900 XTX is a good ComfyUI card. An RX 6600 is a project.
Which install for your Radeon.
| Card | Windows | Linux |
|---|---|---|
| RX 9000 (RDNA 4) | ComfyUI Desktop, or the AMD portable build | PyTorch for ROCm 7.2 |
| RX 7000 (RDNA 3) | ComfyUI Desktop, or the AMD portable build | PyTorch for ROCm 7.2 |
| Ryzen AI Max (Strix Halo) | ComfyUI Desktop | ROCm nightly wheels for gfx1151 |
| RX 6000 (RDNA 2) | Community builds: ComfyUI-ROCm or ZLUDA | ROCm with HSA_OVERRIDE_GFX_VERSION=10.3.0 |
ComfyUI Desktop added ROCm for Windows in version 0.7.0, in January 2026. Its Windows requirements now name NVIDIA and AMD cards.
ComfyUI Desktop has run on AMD cards under Windows since January 2026, on ROCm 7.1.1 [1][2]. ComfyUI also publishes a portable build for AMD, ComfyUI_windows_portable_amd.7z, and its README lists experimental Windows and Linux wheels for RDNA 3, 3.5 and 4 [3].
RX 7000 and 9000 on Windows.
Update the Radeon driver
When ROCm support launched, AMD recommended its ROCm 7.1.1 preview driver. Take the current Adrenalin driver, or a newer ROCm driver if AMD offers one for your card.
Install ComfyUI Desktop
From comfy.org/download. Or unpack the AMD portable build into a plain folder such as
C:\ComfyUIif you’d rather keep everything together.Leave headroom
ComfyUI’s newer memory handling isn’t on for most AMD setups yet, so leave room for Windows. AMD’s own guide uses
--reserve-vram 3on 16 GB cards and--reserve-vram 2 --lowvramon 12 GB cards [4]. In Desktop that’s Reserved VRAM (GB) and VRAM management mode in the server settings.Pick fp16, bf16 or GGUF files
Radeon cards don’t compute fp8. fp8 files still save memory, since ComfyUI unpacks them as it goes, but not time. nvfp4 files are for NVIDIA’s RTX 50 series.
On Linux.
Linux is still where Radeon cards run ComfyUI best, and ROCm has supported them there for longer [8]. Install PyTorch for ROCm, then ComfyUI as usual:
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm7.2AMD’s guide for the RX 9000 series adds these environment settings and flags for stability [4]:
export PYTORCH_HIP_ALLOC_CONF=expandable_segments:True python main.py --reserve-vram 3
ComfyUI’s README adds PYTORCH_TUNABLEOP_ENABLED=1, which can speed things up after a very slow first run.
RX 6000 cards.
AMD’s ROCm doesn’t officially support RDNA 2 in ComfyUI’s builds. It works with workarounds:
- Linux: start ComfyUI with
HSA_OVERRIDE_GFX_VERSION=10.3.0for the RX 6700 and 6600 families, as ComfyUI’s README says. - Windows: the community ComfyUI-ROCm build supports GCN 5 through RDNA 4 [9]. ZLUDA builds work too, more slowly.
- RX 6600: ComfyUI picks bf16 on this card, which runs about half speed. A fix was proposed in September 2026 [7]; keep ComfyUI updated.
DirectML also runs ComfyUI on any Radeon, slowly. It reports a placeholder amount of memory, so tools that read ComfyUI’s memory can’t tell what fits.
What fits.
Memory decides the model list the same way as on NVIDIA. Match your card to the NVIDIA guide with the same memory:
| Memory | Cards | Guide |
|---|---|---|
| 8 GB | RX 7600, 9060 XT 8 GB, 6600 family | 8 GB |
| 12 GB | RX 7700 XT, 6700 XT, 6750 XT | 12 GB |
| 16 GB | RX 9070 XT, 9070, 9060 XT 16 GB, 7900 GRE, 7800 XT, 7600 XT, 6800 to 6950 XT | 16 GB |
| 20 to 24 GB | RX 7900 XT, 7900 XTX | 24 GB |
| 32 GB | Radeon AI PRO R9700 | 32 GB |
Two differences: nvfp4 files are NVIDIA-only, and without ComfyUI’s dynamic memory handling a model that just fits on NVIDIA may need a size down on AMD.
How long it takes.
| Card | Setup | Time |
|---|---|---|
| Radeon AI PRO R9700 | SDXL, 1024 px, 20 steps, ROCm 7.1, Linux | 5.2 s[4] |
| RX 9070 | SDXL, 1024 px, 20 steps, ROCm 7.0 | 1.8 to 2.1 it/s[5] |
| RX 6800 XT | SDXL, 1024 px, 20 steps, ROCm 6.3, Linux | 12 to 16 s[5] |
| RX 7600 XT | SDXL, 1024 px, 20 steps, ZLUDA | 20.5 s[5] |
| RX 6700 XT | SDXL, 1024 px, 20 steps, ZLUDA | 23.5 s[5] |
| RX 6600 | Anima, per step, before and after the fp16 fix | 13.4 s / 6.8 s[7] |
| Intel Arc A770 | Flux.1 Dev fp8, 20 steps | 47 to 53 s[6] |
For comparison, an RTX 4070 does the same SDXL image in 7.1 seconds. There are no reliable numbers for the RX 9070 XT with the newer image and video models yet.
Video on Radeon.
Wan 2.2 runs, with more friction than on NVIDIA. On an RX 7900 GRE with 32 GB of RAM, the high-noise half of Wan 2.2 14B stayed in RAM while the low-noise half loaded, and the system ran out [10]. Start with Wan 2.2 5B, keep 64 GB of RAM for the 14B, and add --reserve-vram.
Intel Arc.
ComfyUI publishes a portable build for Intel cards (ComfyUI_windows_portable_intel.7z), and on Linux and Windows it runs on PyTorch’s own xpu builds. The Arc B580 has 12 GB and ran SDXL at about 4 it/s in the benchmark thread [5]. Use the 12 GB guide for what fits.
Questions.
Does ComfyUI Desktop work with AMD cards now?
Yes. Since version 0.7.0 in January 2026, ComfyUI Desktop runs on Radeon cards under Windows through AMD’s ROCm. The RX 7000 and 9000 series are the supported ones.
ZLUDA or ROCm?
ROCm, where your card supports it: it’s AMD’s own stack and what ComfyUI builds for. ZLUDA builds are for older cards on Windows, and they’re slower.
Why is my RX 6600 so slow?
ComfyUI chooses bf16 on that card, which it computes about half as fast as fp16. A fix was proposed in September 2026. Keep ComfyUI updated.
Does DirectML work?
It runs, slowly, on almost any Radeon. It reports a placeholder amount of memory, so nothing can tell what fits. ROCm is the better route on any card that supports it.
RX 9070 XT or RTX 5070 Ti for ComfyUI?
Both have 16 GB. The 5070 Ti computes fp8 and nvfp4, gets ComfyUI’s newer memory handling, and most guides and custom nodes assume NVIDIA. The 9070 XT works well on ROCm. For ComfyUI alone, NVIDIA is the easier road.
Sources: [1] Official AMD ROCm support in ComfyUI Desktop, Comfy blog, [2] ComfyUI Desktop for Windows, Comfy docs, [3] ComfyUI README, [4] ComfyUI on Radeon RX 9000, AMD ROCm blog, [5] SDXL GPU benchmark thread, ComfyUI discussion #2970, [6] Flux Dev fp8 GPU benchmark thread, ComfyUI discussion #9002, [7] RX 6600 runs in bf16, ComfyUI issue #16535, [8] ComfyUI and an AMD Radeon card, ComfyUI discussion #11209, [9] ComfyUI-ROCm for Windows, patientx, [10] Wan 2.2 on an RX 7900 GRE runs out of RAM, ComfyUI issue #12332, [11] HEISS UI hardware notes.
