Powering a GPU Node On-Demand with iDRAC and Redfish

There’s a Dell PowerEdge R630 in my homelab whose only job is GPU work — mainly Ollama for self-hosted LLM inference. Rack servers are power-hungry, and paying to idle one 24/7 for something I use a few hours a week is absurd. So the machine stays off by default and boots on demand.

Redfish: the API your server already has

Every server with a BMC (iDRAC on Dell, iLO on HP) exposes the Redfish API — a REST interface to the hardware itself. The BMC draws a couple of watts and is always reachable, even when the host is powered off. Power control is a single POST:

curl -sk -K ~/.bmc-creds \
  -H 'Content-Type: application/json' \
  -X POST "https://$BMC/redfish/v1/Systems/System.Embedded.1/Actions/ComputerSystem.Reset" \
  -d '{"ResetType": "On"}'

Two details worth stealing:

1. Keep credentials out of argv. Curl’s -K flag reads options from a file, so the BMC password never shows up in ps or shell history. The creds file is chmod 600, owned by root.
2. GracefulShutdown first, ForceOff as the escape hatch. Redfish supports both; the graceful path lets the OS unmount cleanly.

I wrapped this in a ~40-line script (esxi up | down | status) that also queries PowerState, so any machine on my VPN can check or change the server’s state in one command.

Making it a button

CLI is fine at a desk, but the whole point is “I want to ask a model something now.” So the script is wired into a small dashboard on my hypervisor — one page showing every service’s health, with Power on / Shut down buttons for the GPU node. Boot to Ollama-ready takes about four minutes; the dashboard’s status dot flips green when the tunnel to it comes back up.

The pattern generalizes to anything with a BMC:

• State lives in the BMC, not the OS — you can always trust it.
• Actions are idempotent Redfish POSTs — safe to put behind a button.
• Health is best checked end-to-end — I probe the actual Ollama endpoint, not just a ping.

The economics

An R630 idles around 90–120 W. At my rates, leaving it on costs roughly $10–15/month to do nothing; on-demand it’s pennies. The four-minute boot is a fair trade for inference that runs on my hardware, on my network, with no tokens metered by anyone.