Cited, not measured by AMDai
Sourced from @iamMrDuncan’s X Article on 2026-10-12. Everything on this page is Cited, not measured by AMDai. We do not invent tok/s or prices. llama.cpp first. Q3 stays a labeled estimate. Nothing below Q3.
Primary source: How I run my Lab. House cards: /results. Attribution: /credits.
Cited, not measured by AMDai
Lab-wide agent and machine access. Official Herdr is a terminal workspace for several coding agents at once — detach without stopping work, list which pane is working or blocked, drive panes from a CLI. Shannon’s write-up is why it appears on this site. Not used on the local desktop.
Cited, not measured by AMDai
Self-hosted artifacts and dashboards. Official Coolify is an open-source control plane for deploying Docker workloads on servers you own. It is not an OpenAI-shaped /v1 chat client. Shannon’s write-up is why it appears next to the harness catalog.
Cited, not measured by AMDai
Next to the house Hermes on the local desktop (Strata :8083 and llama-server :8080), Shannon describes a multi-profile Hermes setup and Hermes desktop bot mode. That layout is his. We have not reproduced it. Official Hermes docs stay the attach path. llama.cpp first.
Cited, not measured by AMDai
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.
Cited, not measured by AMDai
Shannon’s 2026-10-05 post cites Qwen 3.8 27B at about 33 tok/s with MTP on a P100. That is NVIDIA, cited context only. It is not a Radeon figure and not a local desktop card. Do not paste it into a Radeon table.
Cited, not measured by AMDai
Shannon’s lab write-up describes an isolated command enclave behind its own router, Proxmox VMs that split the agent host from the knowledge-base host, a guardrail / prompt-injection layer in front of agents, models on a separate NVMe, and offsite backup. Cited layout only. We have not built this on the local desktop.
Cited, not measured by AMDai
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.
Cited, not measured by AMDai
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.
P100 cite (NVIDIA, context only): Qwen 3.8 27B on a P100 (Shannon, 2026-10-05) — Qwen 3.8 27B ~33 tok/s with MTP on a NVIDIA P100. NVIDIA cited context only. Not a Radeon figure. Not measured by AMDai.
How we test — Measured cells cite official amdai-bench/run@1 cards. Estimates stay labeled. Q3 is estimate-only. Q1/Q2 stay off.