local desktop · llama.cpp first

Compare, from the cards

These pages turn the official AMDai 35B sweep into shopper questions: one R9700 or a split across two, which quant to load, Radeon versus an RTX card, GPU-vs-GPU pages under /compare/gpus, and model-vs-model pages under /compare/models built only from the sheets already on the site. Every tok/s and measured GB is from validated amdai-bench/run@1 cards on the local desktop. We do not invent a speed or a price.

Engine is llama.cpp Vulkan. Q3 is listed only as a labeled estimate. Q1/Q2 stay off. The master grid of every 3.6 / 3.8 file stays on /qwen.

• GPUs

  R9700 vs 9070 XT (and other cards)

  Head-to-head from the published sheets: VRAM, bandwidth, which 35B Q6/Q5/Q4 files fit, a labeled Q3 estimate, and a plain-English pick per job. No invented prices or tok/s.

• Models

  Qwen3.6-35B vs GLM 5.3 Flash (and other pairs)

  Head-to-head from cited /models pages: sizes, official-card context, Q6/Q5/Q4 fit estimates on existing GPU sheets, and a labeled Q3 estimate. No invented prices or tok/s.

• This card

  Radeon AI PRO R9700

  One cited page for the 32 GB SKU: what it is for, Start → setup → recipe → Stack Lab, and the measured 35B tables.

• Layout

  One R9700 vs two (split)

  Qwen3.6-35B-A3B on llama.cpp Vulkan: UD-Q5_K_XL, Q5_K_M, Q4_K_M, and Q6_K at 4K / 32K / 128K / 256K. Decode, prefill, VRAM, and system RAM from the overnight cards.

• Quant

  Q6 vs Q5 vs Q4 — plus a Q3 estimate

  Same 35B family, 1-GPU or split. Quality vs speed vs VRAM. Q3 is a labeled estimate (Unsloth UD-Q3_K_XL listing), not a local desktop run.

• Camp

  Radeon vs NVIDIA

  Cited specs and an honest software split for shoppers picking an R9700 / RX 7900-class card versus an RTX 5090, 4090, 3090, 5080, 5070 Ti, or 5060 Ti 16GB. Measured R9700 tok/s; NVIDIA speed stays empty until a cited figure exists.

• Camp

  Radeon vs Mac (Apple Silicon)

  Cited AMD and Apple specs for R9700, 7900 XTX, dual R9700, Strix Halo, and Mac Studio. Measured R9700 35B tok/s; Mac and Strix Halo speed stay not measured here. Q3 is a labeled estimate.

• Backend

  llama.cpp Vulkan vs ROCm (HIP)

  Why the house 35B daily driver is a Vulkan build. HIP is the cited second compile. Measured Vulkan cells only; HIP stays pending until a local desktop card lands. Q3 is a labeled estimate.

• Context

  Long context on Radeon

  128K–262K from local desktop 35B llama.cpp cards. f16 KV default. Empty cells stay not measured yet. Q3 is a labeled estimate. No planted-token recall score.

Also: R9700 measured page · run cards · benchmarks · calculator · two copies vs split · dual / multi-GPU · Compare GPUs · Compare models · Radeon vs NVIDIA · Radeon vs Mac · Vulkan vs ROCm · glossary · 35B model page · Flash model page · gpt-oss 20B · Gemma 3 27B · Llama 3.3 70B · GLM 5.3 Flash · Qwen3.6 27B · Qwen3.8 27B · long context