Intermediate 10 minRadeon RX 6000

Which LLM on Radeon RX 6800 XT / 6900 XT (16 GB) ?

The Radeon RX 6800 XT (November 2020) and 6900 XT (December 2020) have 16 GB of GDDR6 VRAM, 512 GB/s, and RDNA 2. Used, at varying prices, in 2026, they are excellent 16 GB AMD bargains. But beware: official ROCm support on RDNA 2 is limited—Vulkan is often the best backend. This guide details the options and limitations.

Choosing a machine? Our picks by budget →

By Mohamed Meguedmi·Update 2026-08-27·Tested on Windows, macOS, and Linux
Recommended hardware

Buying alternative for this guide: Radeon RX 9070 XT 16GB (ASUS Prime OC).

Why this choice? Our complete guide on Radeon RX 9070 XT 16GB (ASUS Prime OC) →

Compare all options by budget, from €800 to €3,500 →

On the go: which laptop for local AI →

Affiliate links — possible commission at no extra cost to you. As an Amazon Associate, BestLLMfor earns from qualifying purchases.

#RX 6800 XT / 6900 XT

RX 6800 XT
RDNA 2 Navi 21, 4,608 SPs, 512 GB/s, 300 W TDP. Used, price varies.
RX 6900 XT
Full Navi 21, 5,120 SPs, 512 GB/s, 300 W TDP. Used, variable price.
Shared VRAM
16 GB GDDR6, 16 Gbps.

#1. ROCm on RDNA 2

Limited official support
AMD prioritized RDNA 3 for a long time. ROCm 6.x supports RDNA 2, but with some kernel restrictions.
Vulkan recommended
llama.cpp Vulkan works very well on RDNA 2 and delivers 90% of ROCm’s performance. Easier to deploy.
DirectML
On Windows, DirectML works but is slower (-30% vs Vulkan).

#2. Compatible models

RX 6800 XT / 6900 XT 16 GB — Vulkan
ModelQuantTokens/sec
Qwen 3.5 9BQ5_K_M48-58 t/s
Granite 4.2 8BQ5_K_M42-50 t/s
Gemma 4 12BQ5_K_M28-34 t/s
gpt-oss 20BMXFP430-38 t/s
Mistral Small 24BQ4_K_M16-22 t/s

#3. Benchmarks

RX 6900 XT Qwen 3.5 9B Q4: 56 tok/s under Vulkan. ~30% slower than a modern RX 7800 XT.

#2026 verdict

A good used deal at a bargain price
16 GB AMD for cheap. Good option if Linux + Vulkan is OK.
Choose RX 7800 XT if your budget allows
RDNA 3, +30% speed, same 16 GB.

#Frequently asked questions

RX 6800 XT or RX 6900 XT for an LLM?+
~10% speed difference, with a moderate price gap. At the same budget, get the 6900 XT. Otherwise, the 6800 XT is a good deal.
Does ROCm work on RX 6800 XT in 2026?+
Officially supported on ROCm 6.x, but with restrictions. In practice, llama.cpp Vulkan delivers 90% of the performance without ROCm dependencies.
How many tokens/sec does Qwen 3.5 9B get on RX 6900 XT?+
52-58 tok/s in Q4_K_M under Vulkan. Honest, but clearly behind RDNA 3 (RX 7800 XT: 68 tok/s).
Did this guide help you?

Feedback, an error, or a clarification? Let us know—it improves the guide for everyone.