Low-power node
The reference node is deliberately modest: a secondhand Optiplex, one or two Quadro K1200 cards, a small SSD, and a power meter. The point is not to build the fastest local AI machine. The point is to find the smallest machine that is still useful.
Why this build
Section titled “Why this build”- The K1200 is cheap, low-power, and widely available secondhand.
- Maxwell has a current driver branch, so the software stack is less fragile than Kepler.
- A 4 GB card is enough for a ternary 2B model and, in pairs, a 27B model at a usable speed.
- The Optiplex form factor is quiet, small, and easy to power.
| Part | Role | Status |
|---|---|---|
| Optiplex host | CPU, RAM, storage, case | Model to confirm |
| Quadro K1200 4 GB | Inference | 2 available |
| Power meter | Wall power measurement | Needed |
| SSD | OS and model files | TBD |
| Linux | Host OS | Ubuntu 22.04 or 24.04 |
No prices are listed until the build is complete and the secondhand costs are recorded.
Software
Section titled “Software”- Install Linux and the current Nvidia driver for Maxwell.
- Install Kindi. It targets low-cost Nvidia cards and serves an OpenAI-compatible endpoint.
- Download a Ternary-Bonsai model. Start with the 2B, then test the 9B and the 27B on two cards.
- Point a client at the endpoint and run the task suite.
The expected endpoint is http://localhost:8093/v1 with Kindi’s default model id. That detail is confirmed when the node is built.
What we measure
Section titled “What we measure”- Idle power, load power, and power during prompt processing
- Tokens per second at short and long context
- Tokens per joule
- Thermal behaviour in the Optiplex case
- Stability over a 24-hour run
- Task success on the lightweight harness
Status
Section titled “Status”Planned. The first milestone is a single K1200 running the 2B model at the wall, with a measured tokens per joule figure. The second milestone adds the second card and the 27B model.
Dated build notes will live at LocoNotas.