Notes

Six of them, newest first. I would rather post one careful measurement than a weekly opinion.

Every note here was taken on one of two machines, and both are described in the same words every time so you can tell whether the numbers are comparable:

• The desk machine - a 2019 four-core desktop part, 65 W rated, 16 GB of memory, one SATA solid-state disk, one gigabit link, display powered off at the monitor.
• The shelf server - a 2022 eight-core server board, 64 GB, two NVMe disks, two gigabit links, no graphics card.

Both run the same long-term 6.6 kernel. Mains draw is read by an inline meter at the wall with 0.1 W resolution, logged once a second. A measurement is the median of five ten-minute windows, and nobody is logged in while a window runs - the one reading that breaks that rule has a note of its own.

1. 2026-10-06 What an idle login costs

   A terminal left open on the machine under test adds 1.2 W to the desk machine. That is why every other note was taken with nobody logged in.

2. 2026-09-18 Where 118 wakeups a second came from

   An idle server that was not idle. I accounted for every wakeup, removed 80 of them, and saved 1.2 W - which was not the point.

3. 2026-05-19 Letting the disks sleep

   Allowing two NVMe disks their deepest link state saves 0.8 W on an empty bench and 0.3 W under real duty, in exchange for a few milliseconds on the first cold read.

4. 2026-01-28 Taking four cores offline buys almost nothing

   Offlining half the cores of an idle eight-core server saved 0.3 W, inside the meter’s own error, and cost more operational trouble than it was worth.

5. 2025-09-30 The idle loop is not free

   What deep idle states actually buy you, what polling costs instead, and the two ways a residency script lies to you.

6. 2025-02-11 What four governors cost at idle

   Six watts separate the loudest governor from the quietest one on the desk machine. The cheapest watt is not the one you want.

Things I measured and did not write up

Because the honest answer was «no change beyond the noise», and that takes a paragraph, not a page:

• Disabling the second gigabit link on the shelf server: 0.2 W, which is inside the meter’s own error.
• Pinning a single-threaded nightly job to the first core instead of letting it wander: no measurable power or runtime difference across eleven nights.
• Dropping the memory clock one step on the desk machine: 0.6 W saved, and a 4 % slower compile. Not a trade I would take.

The whole set is also an Atom feed, and the index behind the filter is a small JSON file you are welcome to read.