Every so often I get curious about something that nobody else would ever care about. Today it was “how loud is my homelab when the GPU node is training something.”

I put a phone on a stool about a meter from the rack, let it run a synthetic load that pins the GPU at 90%, and captured five seconds. Then I fed the audio through a calibrated SPL measurement. The answer was 58 dBA at one meter, which is — for context — about the level of a quiet office with someone making coffee, or a running dishwasher on the other side of a wall.

It’s not quiet. But it’s not the jet-engine experience homelab horror stories suggest, either. The tolerance for “noisy rack” is a learned thing, and mine has gone up in the six years I’ve lived with one.

Here’s the clip. It is five seconds long, 2.8 MB, and the audio is the whole point. Turn on your volume.

A few notes on methodology, in case you want to do this yourself:

  • The phone was on a stool so that the mic wasn’t in the air path of the fans. Putting it directly in front of the intake will give you a nonsense number that’s mostly wind noise.
  • I ran the load for five minutes before capturing, so the thermal controller had reached steady state. Spikes are a different measurement.
  • I used a calibrated SPL app rather than the phone’s default audio visualizer. The difference is often 5-8 dB.

The GPU node itself (a Supermicro chassis with four Noctua fans swapped in) is the noisiest part. The two 1U servers above it and the switch below don’t register on the load-vs-idle comparison. If I wanted to cut another 5 dB I’d have to either buy a quieter GPU case, move the rack, or accept it. I’ve picked “accept it.”

Bonus: if you’re wondering why there’s a clip at all when dB numbers would have done the job — because I wanted you to hear it. Numbers are easier to measure; clips are easier to feel. Both are useful.