PRO W7900 vs RX 6950 XT for local AI

VRAM, bandwidth, gfx

Which 35B quants fit

Qwen3.6-35B-A3B on llama.cpp, f16 KV. Totals are calculator estimates. Speed is measured only on the R9700 cells that already exist; everything else stays not measured. UD-Q3_K_XL is always a labeled estimate.

Weights file 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.. Same GGUF on both cards.

System RAM

When a house 35B Q4 file does not fit, the spill is system RAM. Shopper lines stay Minimum system RAM needed and RAM needed at full context for Qwen3.6-35B-A3B at llama.cpp, Q4_K_M, 32K selected and the model’s full context. Estimate unless a measured R9700 card overlays the kit. If both sides fit, those labels still name the calculator lines if you raise context or a larger quant.

ROCm and Vulkan

1. Note 1

   llama.cpp first

   Consumer RDNA 2 on this list is unofficial on the current Radeon ROCm row (W6800 gfx1030 is the official workstation listing). Prefer Vulkan llama.cpp if HIP is angry. Ollama and vLLM are not the house path on these sheets.

2. Note 2

   Different gfx family packages

   PRO W7900 needs the gfx1100 family package (48 GB workstation. Dual-slot sibling is also listed.) RX 6950 XT needs gfx1030 (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.) Do not mix those pins.

3. Note 3

   Cited support rows

   PRO W7900: official — 48 GB workstation. Dual-slot sibling is also listed. RX 6950 XT: 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 rows: ROCm 10.1.0 compatibility matrix and Linux system requirements. More gfx notes on /rocm/gfx.

Plain-English pick

• Chat

  PRO W7900 holds house 35B Q4 at 4K in the calculator VRAM pool. RX 6950 XT needs RAM offload or a listed UD-Q3_K_XL file — that Q3 is a labeled estimate (Hugging Face listing for Qwen3.6-35B-A3B-UD-Q3_K_XL.gguf (16.8 GB)).

• Coding agents

  Keep llama.cpp first. The side that holds Q4 in VRAM (PRO W7900) is the less painful coding-agent host. House Flash / agent pins stay R9700 / local desktop — they do not transfer. No invented agent speed.

• Long context

  KV grows with context. Q4 at 32K fits PRO W7900 and spills on RX 6950 XT on the calculator. Read the long-context guide for the measured R9700 128K–262K cards. We do not invent a 128K tok/s here.

• Image gen

  Image gen on AMDai is ComfyUI / HIP PyTorch, not llama.cpp. PRO W7900 is the 32 GB discrete class; RX 6950 XT is the 16 GB discrete class. Seconds-per-image stay not yet measured by AMDai. Start smaller (SD 1.5) on 8 / 12 / 16 GB class cards. Flux-class is tighter than SDXL. Size the checkpoint on the image-gen guide — do not copy a tok/s from the LLM table.

Check price

Affiliate hops only. No live or invented street price. Same disclosure as every other sheet.