Radeon RX 7900 XTX for local AI
The RX 7900 XTX is AMD's top consumer graphics card of its generation and the obvious pick if you want a lot of memory without buying NVIDIA. It suits people who are comfortable on Linux, happy to read setup notes, and mainly interested in running text models. For the rest of the cards see the GPU index, and for choosing a machine overall start at the hardware page.
Memory
The 7900 XTX ships with 24 GB of GDDR6 on a 384-bit bus. That capacity is the reason to consider it: 24 GB of VRAM is the same figure the most popular NVIDIA cards for local AI offer, so any model that fits on one of those fits here too.
GDDR6 gives it high memory bandwidth, well ahead of any system memory and in the same class as competing cards of its generation. Memory speed is rarely the thing that holds this card back. Software is.
As with any discrete card, the practical limit is weights plus KV cache plus working buffers. A model whose parameters only just fit at a given quantization will run out of room once you ask for a long context window, so size down a step and keep headroom.
What it runs well
- Text: comfortable. Running LLMs is where the card is strongest. With 24 GB you cover the mainstream open-weight models at sensible quantization levels, and inference speed is competitive.
- Image: workable. Image generation works and is fast enough to iterate with, but expect to spend time on setup and to lose some CUDA-only speedups and custom extensions.
- Audio: workable. Most audio work runs. Transcription and text to speech are generally fine, while the more research-flavoured music tools are hit and miss depending on how much CUDA-specific code they carry.
- Video: tight. Video generation is the roughest area. Memory is not usually the blocker, support is. Assume workarounds.
Runtime support
There are two paths, and it is worth knowing which one you are on.
ROCm is AMD's compute stack and the higher-performance route. It provides a HIP backend that PyTorch and the main runtimes build against, so most CUDA-targeting Python code runs with little or no change. The caveats are practical rather than theoretical: official support is centred on specific Linux distributions, driver and kernel combinations matter, Windows support lags what Linux gets, and projects sometimes ship ROCm builds later than their CUDA ones.
Vulkan is the fallback and is far easier to get going. llama.cpp has a solid Vulkan backend that runs GGUF models on this card across operating systems with no compute stack to install, and Ollama and LM Studio both offer AMD-capable builds. It is the sane starting point for text generation.
The honest summary versus CUDA: everything mainstream works eventually, but you are not the default target. Tutorials assume CUDA, some libraries used for fine-tuning and LoRA training are CUDA-only or have less tested AMD forks, and when something breaks the answer is usually in a GitHub issue rather than the README.
Trade-offs
This is a large, power-hungry desktop card. It needs a case with airflow, a power supply with headroom, and it will be audible under sustained load. That is the normal cost of desktop performance, and it is also the opposite trade to Apple Silicon, where you give up speed and ecosystem for silence and capacity. See the M3 Max page for that side.
Against NVIDIA at the same capacity, the RTX 3090 is the direct rival: same 24 GB, older, and on the software stack everything is written for. The RTX 4090 and RTX 5090 are faster and, in the 5090's case, larger. Buy the 7900 XTX if you value the card as a graphics card too, want to avoid NVIDIA, and treat setup time as acceptable. If you want the shortest path to a working stack, it is not this one.