Cited, not measured by AMDai
Autoresearch idea
Idea for amdai-bench, cited from Shannon’s layout — one container per GPU, automatic experiments, Prometheus / Grafana for results, a bot that steers the next run. House cards stay amdai-bench/run@1 with an exact launch command. This page is an idea, not a shipped runner.
Cited from How I run my Lab. House schema stays amdai-bench/run@1. llama.cpp first. Q3 is a labeled estimate unless a card is measured. Nothing below Q3.
Cited shape
-
One container per GPU
Isolate the experiment so two cards do not share a process unless the run is a split on purpose. House multi-GPU notes stay on /multi-gpu.
-
Automatic experiment
Sweep engine or runtime knobs, then write an
amdai-bench/run@1line with the exact launch command. Do not invent tok/s. -
Prometheus / Grafana results
Shannon’s write-up names Prometheus and Grafana for the result stream. We have not shipped those dashboards on AMDai. Empty cells stay empty.
-
Bot steering
A bot picks the next experiment. That is his layout. A house overnight loop, when we have one, still has to pass the same how we test rules.
Thermite
Shannon wrote that Thermite, the harness that steers those experiments, is intended to be open-sourced. There is no public repo URL on this site. Do not invent a GitHub path. Not used on the local desktop.
Radeon VII note
Shannon uses a Radeon VII for automated engine and runtime autoresearch aimed at prefill and decode. He did not publish a VII tok/s in the sources we cite. Marked not measured on AMDai. Do not invent a house number. Q3 stays a labeled estimate if someone later contributes a recipe. Nothing below Q3.
Cheap-card context: /hardware. NVIDIA P100 cite stays on Shannon’s lab — not a Radeon table.
How we test — Measured cells cite official amdai-bench/run@1 cards. Estimates stay labeled. Q3 is estimate-only. Q1/Q2 stay off.