The AMD Radeon RX 9070 XT is a 16 GB card with 640 GB/s of memory bandwidth. On paper that beats the RTX 5060 Ti 16GB and comes close to the RTX 5070 Ti.
In 2026 it has also become a much easier card to use for AI. ROCm officially supports it on Linux, ROCm 10.1 now runs on Windows, and LM Studio and llama.cpp run models on it through Vulkan with no setup at all.
The bad news is the price. The card launched at $599 but sells for about $900-$960 in November 2026. This guide covers what it runs, real speeds for LLMs and image generation, how to set it up on Windows and Linux, how it compares with NVIDIA cards, and when it’s worth buying.
- Specs: 16 GB GDDR6, 640 GB/s, 64 RDNA 4 compute units (gfx1201), 304 W.
- LLM speed: gpt-oss-20b about 92 t/s in Ollama; Llama 2 7B Q4_0 about 137 t/s on Vulkan vs about 101 t/s on ROCm.
- Image speed: SDXL 1024×1024, 20 steps in about 3.3-3.7 seconds on Windows with ROCm 10.1.
- Software: Vulkan is the easy path (LM Studio, llama.cpp). ROCm works for ComfyUI and PyTorch on Linux, and on Windows since ROCm 10.
- Price: about $959 new and $899 used (14 Nov 2026), against a $599 launch price.
- Verdict: a good 16 GB AI card if you already own one or find it near $700. At $950 it’s poor value next to a 5060 Ti 16GB at about $770.
RX 9070 XT Specs for AI
| Spec | RX 9070 XT | Why it matters |
|---|---|---|
| Architecture | RDNA 4, Navi 48 (gfx1201) | The target name ROCm and llama.cpp builds use |
| Compute units | 64 | Same count as the Radeon AI PRO R9700 |
| VRAM | 16 GB GDDR6, 256-bit | Fits models up to about 20B at Q4, or 14B at Q6-Q8 |
| Bandwidth | 640 GB/s | Sets text generation speed; 43% more than an RTX 5060 Ti |
| Board power | 304 W | Plan for a 750-850 W power supply |
| Launch | 6 March 2025 at $599 | Now sells far above that (see prices) |
The RX 9070 (non-XT) uses the same chip with 16 GB and is a little slower. The Radeon AI PRO R9700 is the same chip with 32 GB for about $1,700-$1,850, which is the AMD card to get if you need bigger models. I cover it in my best workstation GPU for AI guide.
LLM Speed: Real Tokens per Second
I don’t own an RX 9070 XT, so these numbers come from published tests. The table shows each source so you can check it.
| Model | Tool and backend | Generation (t/s) | Source |
|---|---|---|---|
| gpt-oss-20b | Ollama, ROCm (Linux) | 91.9 | LocalAIMaster |
| Qwen3 14B | Ollama, ROCm | 52.2 | LocalAIMaster |
| Qwen3.5 9B Q4_K_M | llama.cpp, Vulkan (Windows) | 92.3 (ROCm: 81.7) | LocalAIMaster |
| Gemma 4 12B Q4_K_M | llama.cpp, Vulkan (Windows) | 64.6 (ROCm: 54.5) | LocalAIMaster |
| Llama 2 7B Q4_0 | llama-bench, Vulkan | 137.1 (ROCm: 101.3) | Community scoreboard via RunAIHome |
| Llama 3.1 8B Q4_K_M | Ollama, ROCm (Linux) | About 56 | RunAIHome |
| Qwen3-Coder-30B-A3B Q4_K_M (spills past 16 GB) | llama.cpp, ROCm | 30.6 (Vulkan: 22.6) | LocalAIMaster |
| Qwen3.5 27B Q4 (about 17 GB) | Ollama, ROCm | 6.3 | LocalAIMaster |
Two patterns stand out:
- Vulkan usually generates text faster than ROCm on this card, by about 10-35% on models that fit in VRAM. ROCm is often faster at reading the prompt (prompt processing), for example 3,609 vs 2,490 t/s on Qwen3.5 9B.
- Speed falls off a cliff once a model spills out of the 16 GB. A 27B dense model at about 17 GB drops to around 6 t/s.
These are single runs from different software versions, so treat small differences as noise. llama.cpp and driver updates change results often.
Best Models for 16 GB
| Model | Size at Q4 | Speed on 9070 XT | Good for |
|---|---|---|---|
| gpt-oss-20b | About 13 GB | About 92 t/s | The best all-round pick |
| Qwen3 14B | About 9.5 GB | About 52 t/s | Reasoning, longer context |
| Gemma 4 12B | About 6.7 GB | About 55-65 t/s | Writing, chat, vision |
| Qwen3.5 9B | About 6 GB | About 58-92 t/s | Fast assistant, room for long context |
| Qwen3-Coder-30B-A3B | About 18 GB | About 30 t/s with a small spill | Coding, if you accept some offload |
| Dense 27B-32B | 17-20 GB | About 6 t/s | Avoid; get a 24 GB+ card |
For a full breakdown of what fits, see my how much VRAM do I need for AI guide.
Image Generation and ComfyUI Speed
Image generation used to be the weak spot for AMD cards on Windows. That changed with ROCm 10 in August 2026 and ROCm 10.1 on 5 October 2026, which bring native PyTorch on Windows for RDNA 4.
| Workload | Setup | Time | Source |
|---|---|---|---|
| SDXL 1024×1024, 20 steps | A1111 (AMD fork), ROCm 10.1, Windows | About 3.7 s per image | Note.com user test, Oct 2026 |
| SDXL, standard settings | Forge Neo with ROCm Flash Attention | About 3.3 s per image | Same test |
ComfyUI runs on the card with PyTorch built for ROCm 10.1, using the gfx1201 wheel. The same tester reports that ROCm 10.1 fixed a hipErrorInvalidValue crash that hit ROCm 10.0 on this GPU.
Some ComfyUI custom nodes are written only for NVIDIA CUDA, and tools like xformers have only experimental ROCm support. If a workflow depends on a CUDA-only node, it won’t run on any AMD card.
For big video models, 16 GB is tight on any card. My out of memory fix guide covers the low-VRAM flags, and most of them apply to AMD too.
What Software Works (Vulkan vs ROCm)
| Tool | Windows | Linux | Notes |
|---|---|---|---|
| LM Studio | Yes (Vulkan by default) | Yes | The easiest start; switch runtimes in settings |
| llama.cpp | Yes (Vulkan or ROCm) | Yes (Vulkan or ROCm) | Build with -DGGML_VULKAN=ON or -DGGML_HIP=ON -DGPU_TARGETS=gfx1201 |
| Ollama | Use its Vulkan backend | Yes, ROCm v7 or Vulkan | Vulkan is in release builds from v0.12.11 |
| ComfyUI | Yes, with ROCm 10.1 PyTorch | Yes, with ROCm PyTorch | Some custom nodes are CUDA-only |
| vLLM | No | Partly | No FP8 on consumer RDNA 4 yet; llama.cpp is faster for one user |
| Fine-tuning (Unsloth, Llama-Factory) | Limited | Works with effort | AMD lists GPU resets in these tools on RX 9070-series cards as a known issue |
How to Set It Up on Windows
Chat models (5 minutes)
- Update the driver. Install the latest AMD Software: Adrenalin Edition for the RX 9070 XT from AMD’s site.
- Install LM Studio and open its runtime settings. It uses the Vulkan llama.cpp runtime on AMD by default.
- Download a model that fits, such as gpt-oss-20b or Qwen3 14B, and set GPU offload to the maximum.
- Check the speed. If you get less than about 40 t/s on a 14B model, the model is probably running on the CPU.
ComfyUI with ROCm 10.1
- Install the ROCm 10.1 driver (AMD Software: Adrenalin Edition for ROCm 10.1, version 26.10.41.05, released 5 October 2026).
- Create a Python 3.12 virtual environment in your ComfyUI folder, for example
python -m venv venv. - Install PyTorch for your GPU from AMD’s wheel repository, with the gfx1201 device tag and pinned versions, for example
torch[device-gfx1201]==2.14.0+rocm10.1.0with the matching torchvision. - Install ComfyUI’s requirements and start it with
python main.py --enable-manager. - Run a basic SDXL workflow to check it works before you add custom nodes.
Pin every PyTorch package version. If you don’t, pip can pick a torchvision build that doesn’t match, and ComfyUI won’t start. Avoid the device-all tag, which downloads libraries for every AMD GPU.
How to Set It Up on Linux
AMD lists the RX 9070 XT as officially supported in ROCm on Ubuntu 24.04 and 22.04 and on RHEL. On Ubuntu 24.04 the steps are:
- Download AMD’s
amdgpu-installpackage for your Ubuntu version from AMD’s ROCm docs and install it withsudo apt install ./amdgpu-install_*.deb. - Run
sudo apt update, thensudo apt install amdgpu-dkms rocm. - Add yourself to the GPU groups with
sudo usermod -a -G render,video $LOGNAMEand reboot. - Check the card with
rocminfo | grep -i gfx. You should seegfx1201. - Install Ollama with
curl -fsSL https://ollama.com/install.sh | shand runollama run gpt-oss:20b. - Run
ollama ps. The PROCESSOR column should say100% GPU.
If you only want chat models, you can skip ROCm completely. The Mesa RADV Vulkan driver ships with most distros, and llama.cpp or Ollama’s Vulkan backend runs on it. You don’t need HSA_OVERRIDE_GFX_VERSION on RDNA 4, since current ROCm supports gfx1201 natively.
My AMD GPU for AI guide covers ROCm, Ollama and ComfyUI on all recent Radeon cards.
RX 9070 XT vs RTX 5060 Ti vs RTX 5070 Ti vs RTX 3090
| Card | VRAM | Bandwidth | Power | Price, Nov 2026 | Best at |
|---|---|---|---|---|---|
| RX 9070 XT | 16 GB | 640 GB/s | 304 W | About $900-$960 | Fast LLMs through Vulkan, no CUDA |
| RTX 5060 Ti 16GB | 16 GB | 448 GB/s | 180 W | About $770 | Cheapest 16 GB CUDA card, every tool works |
| RTX 5070 Ti | 16 GB | 896 GB/s | 300 W | About $1,400 | Fastest 16 GB option, CUDA |
| Used RTX 3090 | 24 GB | 936 GB/s | 350 W | About $1,450 | 27B-32B models, CUDA |
In practice, the RX 9070 XT and RTX 5060 Ti are close on small models. Both run Llama 3.1 8B at about 51-58 t/s in Ollama. On Vulkan the 9070 XT pulls ahead thanks to its extra bandwidth.
The 5060 Ti wins on software, power use (about $54 a year less at 8 hours a day and $0.15/kWh, by RunAIHome’s estimate) and price.
Read the full card guides: RTX 5060 Ti 16GB for AI, RTX 5070 Ti for AI and used RTX 3090 for AI. If you’re also weighing Intel, see my Intel Arc for local AI guide.
Price in November 2026
| When | New (US) | Used (US) |
|---|---|---|
| Launch, March 2025 | $599 MSRP | – |
| June-July 2026 | About $699 (monthly median) | About $630 |
| 14 November 2026 | $959 (Amazon, lowest) | $899 (eBay) |
Prices are from BestValueGPU’s tracker. The jump since the summer follows the GDDR memory shortage that has pushed up almost every GPU. I explain the causes in why GPU prices are so high in 2026.
Used prices are only about $60 below new, so a used card makes little sense right now unless it comes with a warranty.
Should You Buy One for AI?
| Your situation | Verdict |
|---|---|
| You already own an RX 9070 XT | Use it. LM Studio and ComfyUI both work well now. |
| You game and want to try local AI too | Good choice if you find it near $700 |
| You want a 16 GB card mainly for AI at today’s prices | The RTX 5060 Ti 16GB is cheaper and easier |
| You want 27B-32B models | Skip it; get 24 GB+ (RTX 3090, R9700) |
| You rely on CUDA-only nodes, fine-tuning or vLLM | Skip it; buy NVIDIA |
| You run Linux and like tinkering | A solid pick with great Vulkan speeds |
Still undecided? My best GPU for local AI guide ranks every card by budget.
Troubleshooting
Ollama runs on the CPU
Run ollama ps. If it doesn’t show 100% GPU, update Ollama. On Linux, check that rocminfo lists gfx1201. You can also enable the Vulkan backend with OLLAMA_VULKAN=1, set in the Ollama service rather than your shell.
ROCm doesn’t see the card on Linux
Make sure your user is in the render and video groups and that you rebooted after installing amdgpu-dkms. Don’t set HSA_OVERRIDE_GFX_VERSION on RDNA 4.
ComfyUI crashes with hipErrorInvalidValue
This affected ROCm 10.0 on gfx1201. Upgrade to the ROCm 10.1 driver and PyTorch wheels, or start ComfyUI with --use-quad-cross-attention as a workaround.
Ollama says “unknown model architecture”
Older Ollama builds can’t load some new GGUF files from Hugging Face. Update Ollama or pull the model from the Ollama library instead.
A model is very slow
It is probably larger than 16 GB and spilling into system RAM. Pick a smaller quant or a smaller model, or lower the context length.
Frequently Asked Questions
Is the RX 9070 XT good for AI?
Yes, for models that fit in 16 GB. It runs gpt-oss-20b at about 92 tokens per second and SDXL images in about 3.3-3.7 seconds. Its weak points are the lack of CUDA and its high November 2026 price.
Does the RX 9070 XT support ROCm?
Yes. ROCm supports RDNA 4 (gfx1201) natively on Linux, and ROCm 10 added Windows 11 support in August 2026, with ROCm 10.1 following on 5 October 2026.
Can I run Stable Diffusion and ComfyUI on the RX 9070 XT?
Yes. ComfyUI runs with ROCm PyTorch on Linux, and on Windows with the ROCm 10.1 gfx1201 wheels. Some custom nodes are CUDA-only and won’t work.
Is Vulkan or ROCm faster on the RX 9070 XT?
Vulkan usually generates text faster, for example about 137 vs 101 t/s on Llama 2 7B. ROCm is often faster at prompt processing and is needed for PyTorch tools like ComfyUI.
RX 9070 XT or RTX 5060 Ti 16GB for AI?
At November 2026 prices, the RTX 5060 Ti 16GB at about $770 is better value and works with every tool. The RX 9070 XT is faster on Vulkan but costs about $900-$960.
What is the largest model the RX 9070 XT can run?
About 20B parameters at Q4 entirely on the GPU, such as gpt-oss-20b. Larger MoE models like Qwen3-Coder-30B-A3B run at about 30 t/s with some offload, but dense 27B models drop to about 6 t/s.
Does LM Studio work with the RX 9070 XT?
Yes. LM Studio uses the Vulkan runtime on AMD cards by default, so it works on Windows without installing ROCm.
Final Thoughts
The RX 9070 XT has become a genuinely capable AI card in 2026. Vulkan gives it some of the fastest text generation of any 16 GB card, and ROCm 10.1 finally makes ComfyUI on Windows practical.
The problem is the price. At about $950, it costs more than an RTX 5060 Ti 16GB with the same VRAM and full CUDA support. If you already have one, or you find one near $700, it’s a great card for local AI. Otherwise, the NVIDIA card is the safer buy for most people.
Sources
- LocalAIMaster: AMD RX 9070 XT for local AI (ROCm setup, tok/s, model picks)
- RunAIHome: RDNA4 Vulkan vs ROCm 7.2 for local LLMs (July 2026)
- RunAIHome: RX 9070 XT vs RTX 5060 Ti 16GB for local AI
- Note.com: RX 9070 XT with ROCm 10.1.0 on Windows (October 2026)
- AMD: Adrenalin Edition for ROCm 10.1 driver release notes
- AMD: ROCm release history
- WindowsForum: ROCm 10 adds Windows 11 support
- BestValueGPU: RX 9070 XT price history (November 2026)
