Buying a used Radeon for local AI

Which generations make sense

gfx is the LLVM target ROCm and HIP care about. llama.cpp Vulkan (GGML_VULKAN) is the first path on every row.

• R9700 (32 GB, measured)
• RX 9070 XT (16 GB)
• RX 9070 (16 GB)
• RX 9070 GRE (12 GB)
• RX 9060 XT (16 GB)
• RX 7900 XTX (24 GB)
• RX 7900 XT (20 GB)
• RX 7800 XT (16 GB)
• RX 7900 GRE (16 GB)
• RX 7700 XT (12 GB)
• RX 6800 XT (16 GB)
• RX 6950 XT / 6900 XT (16 GB)
• RX 6700 XT (12 GB)
• PRO W6800 (32 GB, official gfx1030)
• gfx1031 HSA_OVERRIDE note

VRAM-first picker

Same calculator as the GPU sheets: Qwen3.6-35B-A3B on llama.cpp, f16 KV, one GPU, 4K context shown. Q6_K / Q5_K_M / Q4_K_M plus UD-Q3_K_XL (Hugging Face listing for Qwen3.6-35B-A3B-UD-Q3_K_XL.gguf (16.8 GB)) labeled estimate. Speed stays not measured. Open the sheet for the full table.

RAM lines where offload appears

Minimum system RAM needed and RAM needed at full context for the representative card in each class at Q4_K_M, 32K selected. Used-card kits stay labeled estimate. The 64 GB shopper default on consumer sheets is an estimate, not a measured peak.

Vet the card before you pay

1. 1 · Cited, not measured on the local desktop

   Confirm the exact model and VRAM before money moves

   Listings lie about XT vs XTX, 12 vs 16 vs 20 vs 24 GB, and “RX 9700.” On a live Linux box the site already documents: lspci -nn | grep -Ei 'vga|display|3d' for the PCI name, rocminfo | grep -E 'Marketing Name|gfx' for marketing name plus gfx, rocm-smi --showproductname --showmeminfo vram --showuse or amd-smi for published VRAM, and vulkaninfo for the RADV device. Match those strings to the /gpus sheet before you pay. A 7900 XT is 20 GB; the XTX is 24 GB; the 6700 XT is 12 GB gfx1031, not the 16 GB 6800 XT.

   • Verify ROCm
   • Install ROCm
   • GPU directory
   • gfx IDs

2. 2 · Cited, not measured on the local desktop

   Pick the generation for the engine you will actually run

   llama.cpp first. Vulkan (GGML_VULKAN, Mesa RADV) is the always-works consumer path on RDNA4, RDNA3, and RDNA2. Official HIP / ROCm follows the 10.1 matrix: gfx1201 / gfx1200 / gfx110x are listed; consumer gfx1030 / gfx1031 are not the official W6800 row. Community HSA_OVERRIDE_GFX_VERSION (for example 10.3.0 on gfx1031) is unofficial — not an AMD how-to. If HIP is angry, stay on Vulkan. Ollama and vLLM are not the house path for a used card.

   • llama.cpp
   • What is ROCm
   • Vulkan vs HIP
   • gfx1031 HSA_OVERRIDE
   • Ubuntu setup (measured)

3. 3 · Cited, not measured on the local desktop

   Buy VRAM first — then open the sheet, do not copy it

   Shop 12 / 16 / 20 / 24 / 32 GB using the picker on this page. Those totals are the same calculator estimates already on the /gpus sheets for Qwen3.6-35B-A3B on llama.cpp, f16 KV, one GPU. Q6_K / Q5_K_M / Q4_K_M plus UD-Q3_K_XL (Hugging Face listing for Qwen3.6-35B-A3B-UD-Q3_K_XL.gguf (16.8 GB)) labeled estimate. Nothing below Q3. Speed cells stay not measured on used SKUs. The measured 32 GB recipes stay on /gpus/r9700.

   • Calculator
   • Can my Radeon run it?
   • Compare GPUs
   • R9700 measured page
   • Quants

4. 4 · Cited, not measured on the local desktop

   Read both RAM lines wherever the file spills

   Minimum system RAM needed is the kit estimate rounded to a real DIMM size. RAM needed at full context is the llama.cpp KV line at the model’s full context. Both stay labeled estimate on used cards (64 GB shopper default on the consumer sheets). GB here is Linux GiB printed as GB. Offload is not free speed — decode usually drops. Do not invent a kit from a forum screenshot.

   • Calculator
   • Flash RAM offload
   • Long context
   • Budget

5. 5 · Cited, not measured on the local desktop

   Short VRAM and stability check in the return window

   After the card is in the box, follow /tutorials/verify-rocm: rocminfo prints an HSA GPU agent, rocm-smi / amd-smi shows the marketing name and VRAM near the sticker, then a short llama.cpp generate while the meter moves. That is identity plus a live kernel — not a burn-in suite and not a failure-rate claim. If the process never sees /dev/kfd or Vulkan lists only the iGPU, stop and use the return window. We do not invent a loop count or a maximum temperature.

   • Verify ROCm
   • Run a local LLM
   • Troubleshoot
   • ROCm problems

6. 6 · Cited, not measured on the local desktop

   Compare a short llama-bench or amdai-bench run to the site

   Used-card sheets stay not measured on the local desktop. After a short llama.cpp llama-bench or an amdai-bench/run@1 card on your box, compare against the measured R9700 numbers on /benchmarks and /results — do not paste those tok/s onto a 6800 XT listing. Empty speed cells stay empty. Contribute a recipe if you have a log. We do not invent a used-card tok/s to fill the hole.

   • amdai-bench
   • Benchmarks (measured R9700)
   • Run cards
   • Contribute a recipe
   • Wanted

7. 7 · Cited, not measured on the local desktop

   Check board power and connectors from AMD — not a forum watt

   Typical board power (TBP) and any published PSU or connector string live on /power and the AMD product page the GPU sheet already cites. Empty connector cells stay not recorded — we do not invent an 8-pin count. 7900 XTX / XT and 7800 XT TBP figures are on those sheets and /power. Size the PSU for the whole box; dual-card honesty is /power plus /multi-gpu. No prices.

   • Power and PSU
   • Hardware
   • BIOS (ReBAR / SAM)
   • Affiliate disclosure

8. 8 · Cited, not measured on the local desktop

   Watch temps and fans while it generates — not at the wallpaper

   rocm-smi / amd-smi show temperature and fan while llama.cpp is actually generating. Listen for a dead fan or a grinding bearing before the return window closes. Partner coolers change acoustics; we do not publish a failure rate or a “safe °C.” Leave the adjacent slot empty on a blower. Case airflow is /hardware.

   • Verify ROCm
   • Hardware
   • Power and PSU

9. 9 · Cited, not measured on the local desktop

   Two cards are not automatic 2× decode

   A second used Radeon is a /multi-gpu decision, not a speed coupon. Hands-on llama.cpp flags are /setup/multi-gpu. Overnight 35B split numbers are one-R9700-plus-one-R9700 cards — they do not transfer to a used 7900 pair. Mixing RDNA generations or PCIe widths is not measured. Electrical honesty for two boards is /power. Combo-preset tok/s is not measured — use amdai-bench.

   • Dual / multi-GPU
   • Multi-GPU llama.cpp how-to
   • Power and PSU
   • One vs two GPU
   • 35B on two cards

Linux identity commands

Same tools as /tutorials/install-rocm and /tutorials/verify-rocm. amd-smi is the ROCm 10.1 successor to rocm-smi.

lspci -nn | grep -Ei 'vga|display|3d'

rocminfo | grep -E 'Marketing Name|gfx'

rocm-smi --showproductname --showmeminfo vram --showuse

amd-smi

vulkaninfo

Check price — no prices on this page

Existing /go/ routes only. Buttons say “Check price.” We do not print a live or stale number. Affiliate disclosure sits with the buttons and on /disclosure.

FAQ

• Should I buy a used 7900 XTX or a new R9700?

  VRAM first, then gfx. A used 24 GB 7900 XTX is gfx1100; the measured 32 GB house card is the Radeon AI PRO R9700 (gfx1201). Fit estimates for the XTX stay on /gpus/rx-7900-xtx. Measured 35B recipes stay on /gpus/r9700. We do not invent a street price or a tok/s delta between those SKUs. Size the file on the calculator, then use /go/ check-price buttons — no prices on this page.

  • RX 7900 XTX sheet
  • R9700 measured page
  • Budget
  • Compare GPUs

• Is a used RX 6800 XT or 6700 XT still worth it for llama.cpp?

  Yes as a VRAM buy, with honest ROCm status. 6800 XT and 6950 XT / 6900 XT are 16 GB gfx1030; 6700 XT is 12 GB gfx1031. Consumer RDNA 2 is unofficial on the current Radeon ROCm row. llama.cpp Vulkan is the first path. HSA_OVERRIDE_GFX_VERSION on gfx1031 is unofficial. House 35B Q6/Q5/Q4 files expect RAM offload on these cards — both RAM lines stay labeled estimate. Open the sheet instead of this paragraph.

  • RX 6800 XT
  • RX 6950 XT / 6900 XT
  • RX 6700 XT
  • gfx1031 HSA_OVERRIDE

• What should I run before the return window closes?

  Confirm lspci, rocminfo (marketing name + gfx), rocm-smi or amd-smi VRAM, and vulkaninfo. Then a short llama.cpp generate while the meter moves (/tutorials/verify-rocm), and a short llama-bench or amdai-bench run compared to the site’s measured R9700 cards — not as a used-card tok/s. Check TBP and connectors on /power. Watch temps and fans while generating. We do not publish a failure rate.

  • Verify ROCm
  • amdai-bench
  • Power and PSU
  • Hardware

• Can I buy a 12 GB card and just run Q3?

  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 and not a promise that Q3 “makes 12 GB enough.” Q6 / Q5 / Q4 stay on the table. Nothing below Q3. Size the file on the calculator and the 12 GB sheet. Minimum system RAM needed and RAM needed at full context print when that file spills.

  • Quants
  • Calculator
  • RX 6700 XT
  • Can my Radeon run it?

Sources

• AMD — ROCm 10.1.0 compatibility matrix
• AMD — ROCm Linux system requirements
• AMD — ROCm documentation
• llama.cpp — build.md (GGML_VULKAN / GGML_HIP)
• llama.cpp repository
• AMD — Radeon AI PRO R9700 product page