The verdict
3.9 /5
The hardware is not the problem. 24 GB at 960 GB/s is the best you can buy new at this tier for text generation, and it fills a gap NVIDIA leaves open between 16 GB and 32 GB. The cost is specific rather than general: diffusion has friction, fine-tuning is genuinely limited, and Linux is effectively required. Inside those bounds it is a straightforwardly good card.
Scored against our methodology
Buy it if
Running text models through Ollama or llama.cpp on Linux, when you want 24 GB in a new card with a warranty and the CUDA ecosystem is not something your workflow depends on.
Skip it if
Image generation or fine-tuning, where the toolchain assumes NVIDIA; anyone not comfortable on Linux; or anyone who wants 24 GB for the least money, where a used 3090 is cheaper.
Strengths
- 24 GB at 960 GB/s — more capacity than any 16 GB NVIDIA card, matching the RTX 5080's bandwidth
- The only new card near this price with 24 GB; NVIDIA goes 16 GB then 32 GB
- Reaches 32B-class models where 16 GB cards stop at 14B
- A supported path for llama.cpp, Ollama and LM Studio rather than a workaround
- New and warrantied, against a used 3090 with five years of unknown history
- DisplayPort 2.1, ahead of the RTX 40 generation
- The Vulkan back-end is a genuine vendor-neutral fallback
Trade-offs
- No CUDA, and a large amount of AI tooling assumes it
- Fine-tuning is the weakest case here — the memory-efficient libraries are CUDA-first
- Image generation works, with recurring friction in the community tooling around it
- Linux is effectively required; Windows ROCm support has consistently lagged
- 355 W with worse performance per watt than the current NVIDIA generation
- Costs more than a used RTX 3090 with the same capacity
This is the first card reviewed on this site that does not run CUDA, and the temptation is to make the review about that. It would be the wrong emphasis, because the hardware here is genuinely excellent for local language models — and it occupies a position in the market that nothing from NVIDIA currently does.
What it costs you is specific, nameable, and confined to workloads that are not text generation. That is the review.
| Identity | |
|---|---|
| Manufacturer | AMD |
| Product family | Radeon RX 7000 Series |
| Model | RX 7900 XTX |
| Form factor | Graphics card |
| Release year | 2022 |
| Price class | High-end ($2,000–$3,500) |
| Compute | |
| GPU | Navi 31 RDNA 3, 96 compute units |
| GPU architecture | RDNA 3 (Navi 31) |
| VRAM (GB) | 24 |
| VRAM type | GDDR6, 384-bit |
| Memory bandwidth (GB/s) | 960 |
| Compute capability note | ROCm and Vulkan back-ends; no CUDA |
| Connectivity & physical | |
| Power draw | 355 W total board power; 800 W system PSU recommended |
| Workload suitability | |
| Local LLMs | Good |
| Ollama | Good |
| Stable Diffusion / ComfyUI | Workable with caveats |
| Fine-tuning | Limited |
| Homelab | Workable with caveats |
| Developer workstation | Good |
| Largest comfortable model | 32B at Q4 under llama.cpp or Ollama |
The gap in NVIDIA’s lineup that this card sits in
Look at what is available new, at anything like this price, arranged by the number that decides which models you can run.
| Card | VRAM | Bandwidth | New? |
|---|---|---|---|
| RTX 5070 Ti | 16 GB | 896 GB/s | Yes |
| RTX 5080 | 16 GB | 960 GB/s | Yes |
| RX 7900 XTX | 24 GB | 960 GB/s | Yes |
| RTX 3090 | 24 GB | 936 GB/s | Used only |
| RTX 5090 | 32 GB | 1,792 GB/s | Yes, far more money |
NVIDIA’s current consumer range goes 16 GB, then 32 GB, with nothing in between. The 24 GB tier — the one that reaches 32B-class models where 16 GB stops at 14B — exists new only here, or second-hand in the form of a five-year-old RTX 3090.
On the two specifications that govern language model generation, this card also beats both 16 GB NVIDIA options outright: more capacity than either, and bandwidth matching the 5080’s. That is not a consolation position. It is the best hardware you can buy new at this tier for running text models, and it is worth saying plainly before the caveats arrive.
| Model class | At Q4_K_M | On 24 GB |
|---|---|---|
| 14B | ~8 GB | Very fast |
| 32B | ~19 GB | Fits, with working context |
| 70B | ~42 GB | No |
What it costs you, specifically
The ecosystem question is covered properly in NVIDIA versus AMD for local AI, and that page’s conclusion holds. This section is narrower: what this particular card does and does not do well.
Text generation: genuinely fine
llama.cpp has mature ROCm and Vulkan back-ends, and Ollama and LM Studio both build on it. For loading a quantised model and generating text, this card works — not as a workaround, but as a supported path. Our product record rates it good for exactly this, and that rating is not charity.
The Vulkan back-end deserves specific mention as the pragmatic escape hatch: it is vendor-neutral, avoids ROCm entirely, and if a ROCm build is giving you trouble it will usually just run.
Image generation: workable, with friction
Diffusion runs. The difficulty is that the ecosystem around it — custom nodes, extensions, the long tail of community tooling — is written and tested against CUDA first. Things work, and then a specific node does not, and you are debugging somebody’s dependency rather than generating images.
If image generation is a serious part of your work, the Stable Diffusion guide does not recommend this card and neither does this review.
Fine-tuning: limited, and that word is doing work
This is where the gap is widest. The libraries that make fine-tuning tractable on consumer hardware — quantised optimisers, memory-efficient attention implementations — are CUDA-first, sometimes CUDA-only. The fine-tuning workstation assumes NVIDIA throughout for this reason.
RDNA 3 does have matrix instructions in hardware. What it does not have is the depth of software written to use them.
Linux is effectively required
ROCm on Linux is the supported, documented path. Windows support has consistently lagged, and building a workflow on the weaker platform is how an afternoon of setup becomes a recurring cost.
If you are not comfortable on Linux, subtract most of the value of this card. That is not a criticism of the hardware; it is an honest statement of what buying it commits you to.
Against the RTX 3090
For anyone shopping the 24 GB tier, this is the actual decision, and it is closer than the CUDA-versus-ROCm framing suggests.
| RX 7900 XTX | RTX 3090 | |
|---|---|---|
| VRAM | 24 GB | 24 GB |
| Bandwidth | 960 GB/s | 936 GB/s |
| Condition | New, warrantied | Used, five-plus years old |
| Board power | 355 W | 350 W, with worse transients |
| Thermal history | None | Hardened pads, hot GDDR6X |
| Ecosystem | ROCm and Vulkan | CUDA |
| Price | Higher | Lower |
On hardware and condition the AMD card wins: marginally more bandwidth, a warranty, no five years of unknown history, no thermal pads to replace, no card that might have been mined on. On software the NVIDIA card wins, and for some workloads it wins decisively.
If your work is text generation on Linux, this is a straightforwardly better product than a used 3090 and you are paying for that. If your work touches diffusion or fine-tuning, the 3090’s software advantage is worth more than everything in the left column.
On the score
This card totals 3.9, which places it above the RTX 3090 at 3.7 and level with the RTX 5070 Ti — and that is worth explaining rather than leaving to look like enthusiasm.
It is a new card in known condition. It scores better than the 3090 on build quality, thermals, noise, power efficiency and connectivity, because a five-year-old used flagship is genuinely worse at all of those. The 3090 takes AI performance (4.0 against 3.7, the CUDA ecosystem) and value (4.5 against 4.0, because nothing is cheaper per gigabyte).
Against the 5070 Ti the trade is the other way round: this card has 24 GB against 16 and reaches a tier of model the NVIDIA card cannot load, and gives back the ecosystem in exchange.
As always, read the criteria rather than the total. AI performance at 3.7 already carries four times the weight of an ordinary row, and it is still the number that should decide this for you.
Who should buy one
Buy it if you run text models through Ollama or llama.cpp, you work on Linux, and you want 24 GB in a new card with a warranty. In that specific case this is the best hardware available at the price and the software objection largely does not apply to you.
Buy a used RTX 3090 instead if you want the same 24 GB for less and are willing to take a five-year-old card, or if anything you do touches CUDA.
Buy an RTX 5070 Ti instead if your models are 14B and below and you would rather have the ecosystem than the capacity. Less memory, fewer problems.
Do not buy it for image generation or fine-tuning. The card is capable and the software is not there, and no amount of memory bandwidth compensates for a toolchain that assumes a different vendor.
Frequently asked questions
Is the RX 7900 XTX good for local AI?
For running text models, yes — genuinely. 24 GB at 960 GB/s reaches 32B-class models where every 16 GB card stops at 14B, and llama.cpp, Ollama and LM Studio all run on it as a supported path rather than a workaround. For image generation it is workable with friction, and for fine-tuning it is the weakest option here. The card is good; what varies is whether the software for your workload was written with it in mind.
RX 7900 XTX or a used RTX 3090?
Closer than the CUDA framing suggests. The AMD card has slightly more bandwidth, a warranty, no five years of unknown history and no thermal pads to replace — it is the better product on hardware and condition. The 3090 has CUDA and costs less. If your work is text generation on Linux, buy the AMD; if anything you do touches diffusion or fine-tuning, the 3090’s software advantage outweighs everything else.
Do I need Linux to use it for AI?
Effectively, yes. ROCm on Linux is the supported, documented path, and Windows support has consistently lagged behind it. You can make things work on Windows and you will spend time doing so. If you are not comfortable on Linux, that is a genuine reason to buy NVIDIA instead — not because the hardware is worse, but because the platform you would be using it on is the weaker one.
What is the Vulkan back-end and why does it matter?
It is a vendor-neutral compute path in llama.cpp that avoids ROCm entirely. It matters because it is the pragmatic escape hatch: if a ROCm build is fighting you, the Vulkan one will usually just run. It is not always the fastest option, and having a fallback that does not depend on the vendor stack is worth more than it sounds when you are trying to get something working on a weekday evening.
Can it run a 70B model?
No. A 70B model at Q4 needs roughly 42 GB before any context, so 24 GB does not hold it — the same limit that applies to an RTX 3090 or 4090. It runs 32B-class models comfortably, which is the tier this capacity exists for. Reaching 70B means two cards, a 96 GB professional card, or unified memory.
Why does it score above the RTX 3090?
Because it is a new card in known condition, and it scores better on build quality, thermals, noise, power efficiency and connectivity — a five-year-old used flagship is genuinely worse at all of those. The 3090 wins the two rows that matter most: AI performance 4.0 against 3.7 for the CUDA ecosystem, and value 4.5 against 4.0 because nothing is cheaper per gigabyte. Read those two rows rather than the total.
Is ROCm support going to improve?
It has been improving, and we are not going to predict a timeline. What we can say is what to do about the uncertainty: check that the specific tools you rely on support this card before buying, rather than trusting a general impression of the ecosystem — including this page’s. Support moves, and a review is a snapshot.
Did AI Gear Stack test this card?
No. This assessment is drawn from manufacturer specifications and published architectural detail, as stated at the top of the page. Throughput figures follow from the memory bandwidth specification rather than measurement, and the ecosystem judgements here reflect how the tooling is documented rather than our own bench experience with this hardware. We say so rather than guessing.