Measured vs estimate
How we test
A number on AMDai is either a validating amdai-bench/run@1 card or a labeled estimate. Empty stays not measured. We do not scrape leaderboards or invent tok/s.
Benchmarks Run cards amdai-bench Contribute Stack Lab Recipe gates
How measured numbers are produced
Stack Lab and overnight local desktop drops emit one JSON object per line (schema amdai-bench/run@1) under data/runs/. A card records hardware, software pins, GPU layout (single, split, dual-instance), the exact launch command, prefill (pp) and decode (tg) tok/s, VRAM per card / total, and peak system RAM when sampled. Official 35B tables use llama.cpp Vulkan, f16 KV, contexts 4K / 32K / 128K / 256K, one R9700 versus a split. llama.cpp first. The overnight 35B sweep recorded llama-bench, not llama-server.
Measured, cited, and estimate
Measured = a validating amdai-bench/run@1 card. Cited = a public doc with an https URL and an as-of date. Estimate = calculator math or a community listing, always labeled. Empty stays not measured.
How estimates are labeled
Calculator totals (weights, KV, overhead, DIMM-kit rounding) are estimates unless a tested badge cites a card. Official 35B tables print the measured Q6, Q5, and Q4 cards from local desktop. 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. Q3 stays on the table as that estimate. Q1/Q2 stay off, including draft and MTP examples. House daily driver is Q5 when the file fits. We do not treat Q3 as the pick.
Calculator lines Minimum system RAM needed and RAM needed at full context use a measured peak when a card recorded one; otherwise a labeled kit estimate. Open the calculator.
Citations and as-of dates
Each measured cell links its run card. Pages that reprint a figure keep the card date or “Measured on our rig, 2026-10.” External docs get an https citation and a retrieved or as-of date. Version pins are dated snapshots, not a forever matrix.
Corrections
If a card or sentence is wrong, send a sourced recipe or note through Contribute — attach the log, do not invent a replacement tok/s. Official cards can be withdrawn. The 2026-10-12 35B sweep is labeled measured, pending re-verify after the DDR5-5200 oops. Q6_K 1-GPU 4K uses the 23:55 card (103.23 tok/s); the earlier 82.62 card stays on /results and is not averaged.
How to send a result
Contribute a recipe JSON plus a log. Stack Lab exports a lab-notes-report/1 pack — each row is an amdai-bench/run@1 card. Ingest with npx tsx src/scripts/amdai-bench.ts ingest pack.json. Verified Q4+ measured rows merge into data/runs/. Q3 stays a labeled estimate. Do not invent a tok/s to fill a cell.
Pack schema lab-notes-report/1 · row schema amdai-bench/run@1 · how-to Import from Stack Lab.