p620 — AMD Workstation¶

The primary machine: daily development, local AI inference, and the binary cache that speeds up the other hosts.


Profile	`workstation`
GPU	AMD Radeon RX 7900 — `amdgpu` driver, ROCm acceleration
Displays	DP-1 and DP-2, both 2560×1440@120
NFS	exports `/extdisk` to `192.168.1.*`
Boot	standard
Source	`hosts/p620/`

Why this host is shaped the way it is¶

p620 has the only GPU in the fleet suitable for ML work (Radeon RX 7900 + ROCm), so it is the local AI host: it runs Ollama against the GPU and fronts it with a LiteLLM router. It is also powerful and always-on, which makes it the natural binary cache server — the other hosts pull builds from it instead of rebuilding.

What runs here¶

AI stack¶

Ollama server with ROCm acceleration — local model inference.
LiteLLM router — a unified OpenAI-compatible endpoint in front of local and cloud models.
AI providers — OpenAI, Anthropic, and Gemini (keys via agenix), with Ollama as the local provider.
claude-router CLI, claude-hooks, and the PARR protocol tooling.

Development & virtualisation¶

Full development environment (editors, languages, git tooling).
Docker virtualisation enabled.
Syncthing for file sync.
Email with mbsync and AI-assisted features + notifications.

Desktop & extras¶

Desktop environment (COSMIC disabled; Hyprland/GNOME stack).
glance dashboard, scrcpy over Wi-Fi, temperature dashboard.
Media tooling including the yt-x terminal YouTube browser.

Networking¶

Tailscale mesh; the local firewall is left to Tailscale's trust model.
NFS export of /extdisk to the LAN.

Hardware notes¶

p620 carries a few hardware-specific tweaks under hosts/p620/nixos/:

AMD/ROCm setup (amd.nix), CPU and memory tuning, power.
An AQC107 NIC link pin and a USB power fix.
VFIO setup and DNS fallback for resilience.

For the exact files, see the p620 manifest.

Deploy¶

just test-host p620      # build without switching
just quick-deploy p620   # deploy only if the build changed
just p620                # optimised deploy