There's a PC in the corner of the room that hasn't been switched off in a month. It's seeding, streaming Plex to the family, recording the security cameras, or just sitting there so you can RDP in from anywhere. It's 3 AM, nobody's at the desk, and it's still lit up and humming away.
You probably gave up on saving power on that one a long time ago — because it can't sleep. Every power-saving guide on the internet ends with "let it sleep," which makes all of them useless to you. But here's the twist: "can't sleep" and "can't save" are two different things. A PC that's never allowed to nap can still have a real chunk shaved off its power bill.
This is the part where you learn how — how to cut what that machine draws without ever letting it doze off. There's a hands-on way, and if you'd rather not babysit it, a way to automate the whole thing.
The waste profile of an always-on box
Look at what the machine actually does at 4 AM. Seeding is a trickle of I/O. A media server between streams is essentially idle. Camera recording is constant but light. Most always-on workloads need a few percent of the machine's capacity, most of the time — yet the default Balanced plan keeps the CPU boost-ready, the fans on active duty, and a discrete GPU awake at desktop clocks.
The most common homelab pattern makes this worse: the always-on box is yesterday's gaming desktop. Hardware picked for peak performance, now doing a job a NUC could handle, idling at 60–100 W around the clock. The overnight hours — say 01:00 to 08:00 — are the purest waste: nobody is streaming, nobody is at the desk, and the downloads genuinely do not care whether the CPU can boost to 5 GHz.
The knobs that cut power while staying awake
You can do all of this by hand today. The recipes, for the DIY-inclined:
- Never sleep, but kill the display fast.
powercfg /change standby-timeout-ac 0disables sleep;powercfg /change monitor-timeout-ac 5turns the screen off after 5 minutes. A monitor (or a GPU driving one) is power you don't need on a headless box. - Cap the CPU's maximum state.
powercfg /setacvalueindex scheme_current sub_processor PROCTHROTTLEMAX 50followed bypowercfg /setactive scheme_currentcaps the processor at 50% of its rated speed — turbo is gone entirely, and the chip tops out around half its base clock. Trickle workloads won't notice; power and heat under load drop. - Switch cooling policy to passive. The same
sub_processorgroup has a cooling policy setting (SYSCOOLPOL) — passive slows the CPU before spinning fans up. Caveat: this only applies where Windows actually controls the fans (mostly OEM boxes and laptops); on a self-built desktop, the motherboard's BIOS fan curve is in charge, and the CPU/GPU caps are what actually quiet things down. - Power-limit the GPU. On NVIDIA,
nvidia-smi -pl 120(admin shell) caps the card's power draw. Note it resets on reboot, so it needs to run again at startup.
One caution: if your server does real work at night — a transcode queue, backups — cap less aggressively, or schedule around those windows.
The problem: all of those settings are static
Here's where the DIY route gets annoying. At 8 PM you sit down at this same machine to actually use it — and it's still capped at 50% with passive cooling, because that's what you set. So you flip everything back. Later you forget to re-apply it. Two weeks later, the box is back to full power 24/7 and you've stopped bothering.
The honest DIY fix is two batch scripts and Task Scheduler: one job at 01:00 applying the caps, one at 08:00 removing them, plus a startup task to re-apply the GPU limit. This genuinely works, and if you enjoy maintaining it, you have everything you need. Its limits are the usual ones: it doesn't react to whether you happen to be using the machine, it silently breaks when a Windows update resets a setting, and you never find out what any of it actually saved.
Automating the night shift
This exact pattern is what PowerDoze was built around. You define power modes — bundles of CPU limit, cooling policy, screen timeout, and NVIDIA power limit — and schedule rules that switch between them by time of day: Balanced during the day, a capped low-power mode from 01:00 to 08:00. Your server apps go on a keep-awake whitelist, so no rule, mode, or future misconfiguration can ever put the box to sleep mid-seed. The GPU limit is re-applied automatically, updates and reboots included.
And because the obvious question about all of this is "did it actually save anything?", there's an analytics page (part of Pro) tracking estimated watts, kWh, and cost over time — the before/after for your specific machine, not a hypothetical.
The schedule rules, power modes, whitelist, and monitoring are free, no account required. Everything runs locally; nothing about your machine leaves it.
Who doesn't need this
If your always-on hardware is a Raspberry Pi, a NUC, or a purpose-built NAS drawing 10–20 W, you already solved this problem with hardware — an article about shaving desktop idle power has nothing for you. This is for the (very common) other case: a repurposed desktop running Windows because that's what the software needs, or that's what was in the closet. That machine has 20–40 W of overnight fat to trim, and it can be trimmed without ever letting it sleep.