AMD Radeon RX 6950 XT / 6900 XT

Official identity

Card

AMD Radeon RX 6950 XT / 6900 XT

gfx

gfx1030

Published VRAM

16 GB

Hardware label, not a bench

ROCm matrix

Unofficial

RDNA 2 consumer cards are not on the current Radeon official row the way W6800 is. Same gfx1030 LLVM target as the 6800 XT and the official 32 GB W6800. Community HIP/Vulkan paths exist. Prefer Vulkan llama.cpp if HIP is angry.

Cited support row: ROCm 10.1.0 compatibility matrix and Linux system requirements. We do not invent a support claim past that listing. AMD product page: RX 6950 XT.

HSA_OVERRIDE: Sometimes used (e.g. 10.3.0) when a build omits gfx1030. Unofficial. Prefer Vulkan llama.cpp if HIP is angry.

What fits on this card (256K context)

Shopper verdict at 4K, then follow the pair page for 32K and full context. Q4 and up. No printed prices.

What fits in VRAM

Qwen3.6-35B-A3B on llama.cpp, f16 KV, one GPU. Q6 / Q5 / Q4 rows are calculator estimates. UD-Q3_K_XL is a labeled estimate (Hugging Face listing for Qwen3.6-35B-A3B-UD-Q3_K_XL.gguf (16.8 GB)) — not a local desktop run. Speed stays not measured. Change quant or context in the calculator.

Weights notes: Q6_K — Estimate: total_params × 6.56 bpw (nominal GGUF Q6_K; UD-Q6_K_XL uses the same bucket) / 8. Active params do not shrink the file. Q5_K_M — Hugging Face listing for Qwen3.6-35B-A3B-UD-Q5_K_XL.gguf (26.6 GB). Converted 26.6 × 10⁹ bytes → GiB. Not measured VRAM. Q4_K_M — Estimate: total_params × 4.85 bpw (nominal GGUF Q4_K_M) / 8. Active params do not shrink the file. UD-Q3_K_XL — Hugging Face listing for Qwen3.6-35B-A3B-UD-Q3_K_XL.gguf (16.8 GB). Converted 16.8 × 10⁹ bytes → GiB. Not measured VRAM.

System RAM

Minimum system RAM needed and RAM needed at full context for Qwen3.6-35B-A3B on llama.cpp at Q4_K_M, 32K selected and the model’s full context. Estimate until a first-party amdai-bench/run@1 card exists for this SKU. Rounded up to a real DIMM kit.

llama.cpp-first stack

1. Step 1

   llama.cpp first

   Vulkan is the consumer default on this site. HIP is unofficial on this gfx1030 consumer pair — the current Radeon official row lists W6800, not the 6950 XT or 6900 XT. Prefer Vulkan llama.cpp if HIP is angry. Ollama and vLLM are not the house path for these cards.

2. Step 2

   gfx1030 family package

   gfx1030 is the same LLVM target as the official 32 GB W6800 and the 16 GB 6800 XT. Do not mix a gfx110x (7900 / 7800 / 7700 / 7600) or gfx1201 (R9700 / 9070) install. Re-check the cited matrix before you install.

3. Step 3

   6950 XT and 6900 XT are both 16 GB

   AMD 6950 XT product page: 16 GB GDDR6, 256-bit, up to 576 GB/s, 335 W typical board power. AMD 6900 XT product page: 16 GB GDDR6, 256-bit, up to 512 GB/s, 300 W typical board power. Calculator bandwidth on this sheet is the 6950 XT figure. House 35B Q4–Q6 files do not fit this VRAM — expect RAM offload or a lighter listed Q3 file, both estimates. The 32 GB W6800 is a different listing.

4. Step 4

   Setup, then Stack Lab doctor

   First-run setup is Ubuntu. Open Stack Lab doctor for the install family rocminfo actually prints. House daily and Flash profiles are R9700 / local desktop — they do not apply to this 16 GB pair. Packages are not published yet.

Compare with…

Head-to-head pages use the same VRAM, bandwidth, and calculator fit already on the sheets. Q3 is a labeled estimate. No invented prices. RX 6950 XT against the other published sheets:

• Compare GPUs
• vs PRO W6800
• vs RX 6800 XT

Also: PRO W7700 · PRO W7800 · PRO W7900 · R9700 · RX 6700 XT · RX 7600 · RX 7600 XT · RX 7700 XT · RX 7800 XT · RX 7900 GRE · RX 7900 XT · RX 7900 XTX · RX 9060 XT · RX 9070 · RX 9070 GRE · RX 9070 XT · Strix Halo