RX 9070 XT for Local LLM (2026): ROCm Support Is Real

Bottom Line

• ROCm support is official and verified. AMD’s ROCm 7.14 system requirements list the RX 9070 XT as gfx1201, along with the RX 9070 and 9070 GRE. No HSA_OVERRIDE_GFX_VERSION needed.
• That is the whole story, and it is a real one. For years the AMD answer was “it works if you lie to the driver”. This card does not need the lie.
• 16GB is the ceiling and it bites. gpt-oss 20B fits at 12.8 GiB. Qwen3.6 27B (17GB) and Qwen3.8 27B (18GB) do not.
• 640 GB/s of bandwidth, from 16GB of GDDR6 on a 256-bit bus at 20 Gbps. 64 compute units, 304W board power.
• Buy the 7900 XTX instead if you want the 27B tier. 24GB on older silicon beats 16GB on newer silicon for inference. Capacity wins.

Ready to buy? See the tested hardware list with current prices.

The Two-List Problem, and Which List This Card Is On

Every AMD local-AI recommendation runs into the same trap. There are two lists of cards, and most guides do not tell you which one they are quoting from.

List one: cards AMD officially supports. These appear by name in the ROCm system requirements with an LLVM target. Driver updates are tested against them. When something breaks, it is a bug.

List two: cards that work anyway. These run by setting HSA_OVERRIDE_GFX_VERSION to impersonate a card on list one. This works, often well. It also means nobody at AMD tested your configuration, and a driver update can end it without warning.

We covered this split in which AMD GPUs actually work for local LLMs. The short version is that the gap between the lists has been the single biggest reason to buy NVIDIA instead.

The RX 9070 XT is on list one. AMD’s ROCm 7.14 system requirements page names it directly:

Read from AMD’s ROCm 7.14 documentation on 2026-09-07.

Notice what is absent from that table: the entire RX 6000 series. Not one consumer RDNA2 card appears. If you own an RX 6800 XT or a 6900 XT, you are on list two permanently.

The Specifications That Decide Anything

The bandwidth figure is arithmetic, not a marketing number: 256 bits at 20 Gbps is 640 GB/s exactly.

For inference, bandwidth sets your generation speed and capacity sets what you can run at all. 640 GB/s is a perfectly good number — it is roughly a third of an RTX 5090 and comfortably ahead of any unified-memory Mac in the same price class. The card is not slow.

It is just small.

What It Runs

Model sizes read from the Ollama library and OpenAI’s Hugging Face repo on 2026-09-07.

gpt-oss 20B is the reason to buy this card and also where it stops. One caveat carries over from our 16GB VRAM guide: gpt-oss 20B’s weights fit, but its full 128K context does not. The KV cache costs 3.0 GiB at fp16, which takes you to 15.8 GiB. Run it with a Q8 cache and the full window fits.

# Ollama on a supported RDNA4 card — no override needed
ollama run gpt-oss:20b

If you have used AMD cards for inference before, the absence of an export line above is the entire point of this article.

The Honest Comparison Against the 7900 XTX

This is where the recommendation stops being automatic.

Both cards are on list one, so the 9070 XT’s headline advantage evaporates in this specific comparison. What remains is newer architecture and lower power against 8GB more capacity.

For inference, take the capacity. The 8GB gap is the difference between running Qwen3.8 27B and not running it, and no architectural improvement compensates for a model that will not load. RDNA4’s second-generation AI accelerators help with prompt processing, but they cannot conjure VRAM.

The 9070 XT wins when you are buying a card that also plays games, when 304W versus 355W matters for your PSU or your room, or when gpt-oss 20B is genuinely all you need. Those are real cases. They are just narrower than the newer-is-better instinct suggests.

If you want to skip the argument entirely, the Radeon AI PRO R9700 puts 32GB on the ROCm-supported list.

On Price

We are not quoting one. Every launch-era figure for this card is stale — the 2026 DRAM shortage pushed several last-generation cards above their own launch prices, and any number we publish today would mislead you next month. Check current listings before you decide, and compare the per-gigabyte cost against the 7900 XTX rather than the sticker.

See Also

• Which AMD GPUs actually work for local LLMs — the two-list problem in full, across every Radeon generation
• Best local LLM for 16GB VRAM — exactly what this card can load, with the KV-cache caveat
• Best local LLM for the RX 7900 XTX — the 24GB AMD tier this card sits below
• Cheapest 32GB VRAM GPU in 2026 — where the R9700 lands against the Intel and NVIDIA options
• Is 16GB of VRAM still enough in 2026? — the capacity question that decides this purchase