GreenCluster is three NVIDIA GB10 servers (128 GB of unified memory each) working as a trio, orchestrated by k3s + ArgoCD (GitOps). Storage lives on ZFS, replicated by Syncoid. Every piece of infrastructure is versioned in git : nothing is "hacked by hand", everything redeploys in one gesture.
Three servers, one cluster
- MSI — the control plane (k3s controller + ArgoCD) and primary ZFS storage. An accepted SPOF, but replaceable by design.
- Dell — the compute node (worker). Its lifecycle is automated : drain on shutdown, rejoin + cleanup on boot.
- Joy — the third one, moved in on 10 September. It changes the game: the brain of the house can now run in tensor parallelism 2 without tying up the control plane.
🏠 Home Assistant — the conductor of energy
A cluster running around the clock burns power for nothing. So the house also has home automation, and it is what watches over the machine's cycle: on when work arrives, off when there is none. We deliberately keep the details of those rules off this page — domotics lives in the house; it is not a shop window.
Learn more about the hero watching over us : home-assistant.io.
✨ The lifeline — the brain of the house
The heart that keeps the team going : a self-hosted open-source LLM, serving an OpenAI-compatible API (alias gpt-4o) across the cluster, powered by vLLM (the open-source inference server) and a community Docker image built for the NVIDIA GB10 — ready recipes, multi-node loading, RoCE acceleration.
July 2026: DeepSeek V4 moves in. August: Qwen3.8-Flash-Next — a 125-billion-parameter cherry tree (6 billion active at a time) — takes over, now with eyes of its own. DeepSeek sleeps next door, ready to serve again at a moment's notice. Brains evolve; the data never moves.
- solo — a light model on MSI, for everyday life without effort.
- cluster — the brain of the moment in tensor parallelism 2 (both GB10 linked by QSFP/RoCE), when real power is needed.
💪 Power mode
Three machines is not three more of the same: it is a separation of roles. Since 20 September the brain of the moment spreads across the two compute nodes (linked at 400 GbE), and the control plane gets room to breathe.
The first obstacle was not the network but the cellar: ZFS caches everything it reads, and on unified memory that cache eats the model's lunch. Capped, measured — and the brain moved in. Caching is a feature, when it knows its place.
📜 The logbook
Since July 25th: more than 50 missions, close to a thousand commits — the story fits in three dates — and one chapter per day on the journal page.
- 25/07 → 06/08 — the hull and the first brain: k3s, ArgoCD, GitOps, a self-hosted lifeline in tensor parallelism, and exactly one black day (05/08) from which the house took its golden rules on the network.
- 07 → 31/08 — the great sharing and the graft: the household's data onto native, replicated, snapshotted ZFS; then the house changes heads while still running, without moving a byte.
- 10 → 26/09 — the third machine (Joy), backups that can say where they are, power mode, and the workshop opening: a wall, a voice, an ear, a hand, and a face.
The founding idea : serious AI power only when necessary, while limiting consumption the rest of the time. Sober by design 🌿.
And this is only the beginning: the house is built to grow — more machines, other brains, the same memory.