Which AMD GPUs Actually Work for Local LLMs in 2026?

Bottom Line (August 2026)

• There are two supported-GPU lists and they do not match. AMD’s ROCm 7.14 page is the narrow one. Ollama’s hardware page is the wide one. For local LLM work, the runtime’s list is the one that decides.
• Safest buys: Radeon AI PRO R9700 32GB ($1,299 MSRP) and RX 7900 XTX 24GB (~$900-950). Both are on both lists.
• Works but unlisted by AMD: RX 6800 / 6800 XT / 6900 XT / 6950 XT (gfx1030) and RX 7600 (gfx1102). Ollama lists both targets; AMD does not.
• Cut, and stay away: the MI50, Radeon PRO VII and Radeon VII (gfx906). AMD marks them unsupported and warns that prebuilt libraries “will cause runtime errors.”
• The detail that explains AMD’s policy: MI100 (gfx908) is still supported while the MI50 (gfx906) is not — an older architecture kept, its successor kept, and the cheap one dropped.

Two Lists, One Question

Ask “does this AMD card work for local LLMs” and you get opinions. Both companies involved publish an answer instead.

AMD’s ROCm 7.14 system requirements list the hardware AMD will support. Here is that list, verbatim in substance:

Ollama’s hardware documentation lists what its own builds run on. Its supported targets include gfx908, gfx90a, gfx942, gfx950, gfx1030, gfx1100, gfx1101, gfx1102, gfx1150, gfx1151, gfx1200 and gfx1201.

Put the two side by side and three targets appear on Ollama’s list that are absent from AMD’s: gfx1030, gfx1102, and the gfx1150/gfx1151 Ryzen AI integrated graphics.

That gap is not an error. It is the difference between “AMD will take your support ticket” and “the software runs.” Almost nobody running a local model needs the first one.

What the Gap Means Card by Card

gfx1030 — the RDNA2 cards that quietly still work

The RX 6800, 6800 XT, 6900 XT and 6950 XT are all Navi 21, and Navi 21 compiles to gfx1030.

AMD does not list a single consumer RDNA2 Radeon as supported. But it does list the W6800 and V620, which are the same gfx1030 target — so ROCm ships gfx1030 binaries regardless. Ollama lists gfx1030 in its own supported targets and names the RX 6800 in its Linux table.

Practical translation: a used RX 6800 XT runs local models today. You have no promise it will still be in the build matrix in two years, and the day AMD drops gfx1030 from the professional side, the consumer cards go with it silently. That is precisely how the MI50 died.

gfx1102 — the RX 7600, listed by the runtime and not by AMD

The RX 7600 is Navi 33, target gfx1102. AMD’s RDNA3 support stops at the 7700 XT (gfx1101). Ollama supports gfx1102.

It works. At 8GB of VRAM it is not a card we would recommend buying for this job anyway — see is 16GB of VRAM still enough for why the current model band has moved past small cards entirely — but if you own one, it is not a dead end.

gfx906 — the one genuine cliff

AMD’s page marks the MI50, Radeon PRO VII and Radeon VII unsupported and warns that prebuilt ROCm libraries “will cause runtime errors.” ROCm 5.7 was the last release with full support and fixes ended in Q2 2024. Ollama does not list gfx906 either. Both lists agree, which is the only time in this article that they do.

Here is the part worth sitting with: MI100 is gfx908, it is a year older in spirit and one generation newer in silicon, and it is still supported in 7.14. AMD kept the data-centre card and cut the one that ended up on eBay for a few hundred dollars. We went through what that means for buyers in the MI50 32GB write-up — 1,024 GB/s of HBM2, more bandwidth than a used RTX 3090, and no vendor behind it.

A support list is a business document. Read it that way and the pattern is obvious: AMD supports what it currently sells and what enterprises currently pay for.

The Escape Hatch, and Its Limits

Ollama documents an environment variable, HSA_OVERRIDE_GFX_VERSION, that forces an unlisted card to identify as a nearby supported target — HSA_OVERRIDE_GFX_VERSION=10.3.0 to run as gfx1030, for instance, with numbered variants for mixed multi-GPU systems.

This is genuinely useful and it is also the reason so much confusion exists. It works when the unlisted card is architecturally close to a supported one. It does not conjure support for hardware whose instruction set has actually diverged, and when it fails it tends to fail as wrong output or a hang rather than a clean error.

Treat it as a way to keep a card you already own working. Do not treat it as a reason to buy one.

What to Buy

Best AMD card for local LLMs — Radeon AI PRO R9700 32GB

32GB of VRAM, gfx1201, on both supported lists, at $1,299 MSRP as of August 2026. That capacity is the whole argument: 32GB runs a 30B-class model with a real context window where a 24GB card runs out. The caveat is stock — Newegg alternates between MSRP and out-of-stock, and secondary listings have run from $1,403 to $2,500, which erases the value case entirely. Buy it at MSRP or do not buy it.

Compare directly against the obvious NVIDIA alternative in R9700 vs RTX 3090.

32GBXFX Radeon AI PRO R9700 32 GB ↗

Best value, and the sanest AMD buy — RX 7900 XTX 24GB

24GB at roughly $900-950 as of August 2026. In a year where the RTX 5090 trades at more than double its MSRP and last-generation NVIDIA cards sell above their own launch prices, the 7900 XTX sitting near its $999 MSRP makes it the best dollars-per-gigabyte in new consumer cards. gfx1100, on both lists, and the card most local-AI ROCm tutorials are actually written against.

24GBXFX Speedster MERC310 RX 7900 XTX 24 GB ↗

The 128GB AMD route — a Ryzen AI Max+ 395 box

Worth naming because it appears on Ollama’s list (gfx1151) and not on AMD’s ROCm list, which is a strange place for AMD’s own flagship AI APU to be. 128GB of unified memory in a mini PC, and the cheapest way to hold a very large mixture-of-experts model at home. Stock is the problem, not silicon — see which Strix Halo mini PC to buy.

128GBRyzen AI Max+ 395 128 GB Box ↗

If the software risk is what you are weighing — the NVIDIA baseline

An honest comparison has to name the thing AMD is being measured against. A used RTX 3090 at $1,000-1,300 has 24GB and CUDA, and no article like this one needs to exist for it. That is the premium NVIDIA charges and the reason it can.

24GBEVGA RTX 3090 24 GB ↗

The Honest Caveat

“Supported” is not one thing, and this article has been careful not to pretend otherwise.

A card can be on AMD’s list and still be awkward — a niche quantization kernel that exists only for CUDA, a fine-tuning library that assumes NVIDIA, a new model whose day-one llama.cpp support lands on ROCm a week later. Being supported means the platform builds for your card. It does not mean every tool in the ecosystem was tested on it.

And a card can be off AMD’s list and work perfectly for years, right up until it does not. The gfx1030 cards are in exactly that position now.

The rule we would give: buy an AMD card on the runtime’s list for what you run today, buy it at a price that already assumes you might get less support than an NVIDIA card, and never buy one whose only argument is that a community fork still builds for it. That last sentence is the MI50, and it is why we told people not to buy one.

See Also

• AMD R9700 vs RTX 3090 — the head-to-head this site ranks for
• AMD MI50 32GB for Local LLMs — the card AMD cut, and why the bandwidth is a trap
• The Cheapest 32GB VRAM GPU — where the R9700 sits against Intel and NVIDIA
• Which Strix Halo Mini PC to Buy — AMD’s 128GB unified route
• Is 16GB of VRAM Still Enough? — the capacity question underneath the vendor question
• Best Local LLM for 96GB of VRAM — where AMD currently has no answer at all