Which LLM on Radeon RX 7700 XT (12 GB) ?

#What a RX 7700 XT does in 2026

A RX 7700 XT is suitable for a mid-sized local LLM: its 12 GB can accommodate Llama 3.1 8B, Gemma 4 12B, or Phi-4 14B in Q4 with room for context. Its 432 GB/s bandwidth, 12.5% higher than a RX 6700 XT, sets the generation ceiling: about 113 tokens per second on the scoreboards' reference 7B, versus 163 for a RX 7800 XT. It has a clear advantage over the 6700 XT: ROCm and Ollama officially list it, including under Windows. Its drawback is capacity, which rules it out for 20- to 24-billion-parameter models, where the 7800 XT's 16 GB becomes valuable.

Architecture

RDNA 3, a Navi 32 chip assembled from chiplets (1 GCD and 3 MCDs), launched on September 6, 2023.

Compute

3,456 stream processors, 54 compute units.

Memory

12 GB GDDR6, 192-bit bus, 432 GB/s.

Power consumption

245 W of board power.

Pricing

Not listed here: check /prix-gpu-ia for the lowest price recorded each week.

#ROCm, Ollama, and Windows: official support

This is what distinguishes it from the previous generation. AMD’s compatibility table lists the RX 7700 XT under RDNA 3, target gfx1101, like the RX 7800 XT. The Ollama documentation lists it among the supported cards on Linux and Windows. One nuance affects the choice of operating system: according to AMD, the Radeon RX 7000 and 9000 are supported only on Ubuntu 24.04.4, Ubuntu 22.04.5, RHEL 10.1, and RHEL 9.7. Another distribution may work, but without guarantees. Vulkan remains the safety net: Ollama enables it by default when the backend is installed.

#What fits in 12 GB of VRAM

On a 12 GB card, allow about 11 GB for the model and its KV cache if it isn't driving the display. The listed weights come from the QuelLLM catalog; the KV cache is added on top and grows with conversation length.

To go beyond 14B without changing cards, there are two options: quantize the KV cache to recover one or two gigabytes, or choose a mixture-of-experts (MoE) model in which only a few parameters are active per token. The occupied memory remains the model’s full size, so a 20B MoE does not fit any better in 12 GB, but it generates faster if it fits.

• Which LLM for 12 GB of VRAM, across all GPUs
• Quantize the KV cache to reclaim VRAM
• VRAM calculator for your model and context

#Throughput: what we know and what we estimate

Let’s be precise about the evidence: the two public discussions in the llama.cpp repository, one for Vulkan and the other for ROCm, which list AMD cards using the same model (Llama 2 7B in Q4_0), contain no entry for the RX 7700 XT at the time of writing. We did not measure anything ourselves. What we can do is place it between two measured cards: the RX 6700 XT at 83.88 tokens per second in generation (384 GB/s bandwidth) and the RX 7800 XT at 118.27 (624 GB/s), both using Vulkan.

Bandwidth-based reasoning gives us an upper bound: 432 GB/s divided by 3.82 GB of weights yields about 113 tokens per second on the reference 7B model. The AMD cards in the same discussion reach between 59% and 83% of that (83% for the 6700 XT, 72% for the 7800 XT). On that basis, a value between 65 and 95 tokens per second seems plausible for the 7700 XT: this is a working hypothesis, not a result, to be confirmed with llama-bench.

For conversation, anything above roughly ten tokens per second reads comfortably, so these estimates remain well above the practical threshold, even with a significant margin of error.

#Compared with its direct neighbors: 6700 XT, 7800 XT, and RTX 4070

The choice often comes down to three closely matched cards. Measurements published under Vulkan help position the RTX 4070, one of the 12 GB Ada-generation cards discussed here: 92.29 tokens per second for generation and 3,179 for prompt processing. Its 12 GB, like the 7700 XT’s, limits both cards to the same models; the difference comes down to the software ecosystem, prompt processing, power efficiency, and price. The RX 7700 XT’s throughput remains to be measured; the other three have published figures.

How to read this table: moving from the 7700 XT to the 7800 XT doesn’t just add speed; it adds 4 GB of VRAM, giving you access to 20B to 24B models. This is the only criterion that truly changes what you can run. If you use an 8B or 12B model, the 7700 XT isn’t a limitation; if you’re targeting 24B coding models, it is.

• Radeon RX 7800 XT, the same chip with 16 GB
• Radeon RX 6700 XT, the previous generation with 12 GB
• RTX 4070: the NVIDIA alternative at 12 GB

#Getting started on Linux

1. 01
   Check the system

   Make sure your distribution is one of Ubuntu 24.04.4, Ubuntu 22.04.5, RHEL 10.1, or RHEL 9.7, the only versions AMD lists for this card.
2. 02
   Install the ROCm driver

   Ollama requires the ROCm v7 driver on Linux. Install it with the amdgpu-install utility described in AMD's documentation.
3. 03
   Install Ollama, then a model

   Download Gemma 4 12B in Q4_K_M or an 8B model, then run it with ollama run.
4. 04
   Check that the GPU is working

   Run ollama ps: the processor column should indicate execution on the GPU. If the card is not detected, add the ollama user to the render group.
5. 05
   Measure and compare

   Run llama-bench with the GGUF from the public discussion to compare your card with others' measurements and publish yours.

ollama ps
rocminfo | grep -i gfx
./llama-bench -m llama-2-7b.Q4_0.gguf -ngl 100 -fa 0,1

On Windows, the Ollama documentation specifies a ROCm v7 or HIP7 stack for Radeon RX 7000; installation is limited to updating the Radeon driver and then running Ollama. The guide on Ollama and AMD GPUs explains the environment variables useful for selecting a card when the system has several.

• Ollama and AMD GPUs: configure Radeon RX 6000 to 9000

#When 12 GB is no longer enough

The threshold is clear: above 11 GB of weights, the model no longer fits entirely in VRAM. Ollama and llama.cpp then offload some layers to the processor, and speed collapses because system memory is much slower than GDDR6. Three situations come up repeatedly: a 24B coding model, a long context on a 14B model, or a second model loaded at the same time for RAG. Each one is solved with VRAM, not computing power.

Before buying it used, check three specific points. The card’s 245 W power draw requires a suitable power supply and two properly connected PCIe power connectors. The exact model matters little for an LLM because all manufacturers use the same 12 GB of memory; only the cooling quality changes, which matters during long generation sessions. Finally, test the card with a real 9–10 GB model before approving the purchase: it’s the only test that uses most of the 12 GB of memory and the continuous weight rereading an LLM performs, which has nothing to do with a video game.

#2026 verdict

If you already have it

Keep it: it covers everything up to 14B with official support and no workaround.

If you buy used

Compare its price with that of a RX 7800 XT on /prix-gpu-ia: if the difference is modest, the 7800 XT’s additional 4 GB is worth it.

If you are targeting 24B models

Go straight to 16 GB or more: capacity, not throughput, is the limiting factor.

• Local AI graphics card prices, recorded twice a week
• Ollama documentation: supported AMD GPU
• AMD ROCm compatibility table
• Vulkan scoreboard for the llama.cpp repository

#Frequently asked questions