Proxmox power usage: measure it, then cut it

Proxmox has no power meter of its own, and a VM can't see real watts. Measure the host at the wall (smart plug or meter), through the UPS (NUT), or through the server's management controller (IPMI). Then check the CPU scaling driver and governor, enable power-saving states, spin down idle disks, and remove idle hardware, measuring before and after each change.

The short answer

Proxmox doesn’t measure watts, and a VM can’t see real power at all. Measure the host from outside, then work through the usual idle-power settings, measuring after each change. Most of a homelab host’s energy goes into idling, so idle watts are the number to bring down.

Step 1: measure

Method	What it measures	How
Plug-in meter / smart plug	Whole host at the wall (what you pay for)	Read it at idle and under typical load; a smart plug can log it over days
UPS through NUT	Everything on that UPS	upsc <ups>@<host>: use ups.realpower if present, otherwise ups.load (%) × ups.realpower.nominal ÷ 100
IPMI / BMC (rack servers)	Whole server, from its own sensors	ipmitool dcmi power reading
RAPL (Intel and recent AMD)	CPU package (and DRAM) only	turbostat --Summary --show PkgWatt --interval 5 on the host, not in a VM

RAPL is good for comparing settings but leaves out disks, fans, NICs and power-supply losses. For cost, use the wall, the UPS or IPMI.

Step 2: check the CPU’s power settings

On the Proxmox host shell:

cat /sys/devices/system/cpu/cpu0/cpufreq/scaling_driver     # intel_pstate, amd-pstate-epp, acpi-cpufreq …
cat /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor   # performance, powersave, ondemand, schedutil …
cat /sys/devices/system/cpu/amd_pstate/status 2>/dev/null   # AMD only: active, guided or passive
cat /proc/cmdline                                           # any governor / pstate overrides

• With intel_pstate or amd-pstate in active mode, only performance and powersave are offered. powersave still scales up under load; the energy preference (EPP) decides how eagerly.
• With amd-pstate in guided mode, the classic governors (schedutil, ondemand …) return.
• Try powersave (or schedutil) for a day: echo powersave | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor (lost at reboot). Measure idle watts before and after. To keep it, add cpufreq.default_governor=powersave to the kernel command line and run proxmox-boot-tool refresh (systemd-boot) or update-grub (GRUB).

Don’t trust a default; check what your host actually runs.

Step 3: cut idle watts (measure after each)

1. BIOS: allow deep C-states and PCIe ASPM. Many servers and some desktop boards ship with them off or in a “performance” profile. This is often the biggest single change.
2. powertop --auto-tune (apt install powertop): turns on power management for USB, SATA and PCIe devices. Test for a day before making it permanent; some NICs and USB devices dislike it.
3. Spin down idle disks. hdparm -S 120 /dev/sdX sets a 10-minute spindown (values 1–240 are multiples of 5 seconds). Don’t do this on disks a ZFS pool or a VM touches every few minutes: constant spin-ups cost more than they save.
4. Remove what idles. Unused HBAs, 10G NICs, GPUs and extra drives each draw several watts or more around the clock.
5. Stop or consolidate idle VMs. Busy-polling guests and many small VMs keep cores out of deep idle. Containers (LXC) usually cost less than a full VM for the same service.
6. Fans and BMC profiles. On rack servers, a “minimum power” or “optimal” profile in the management controller can lower fan speed and idle draw.

Step 4: is new hardware worth it?

Two small tools:

Savings from a change. Enter idle watts before and after, plus your price per kWh, to get watts saved and money saved per month and per year. Example: 150 W → 30 W at €0.30/kWh saves 120 W, ≈ €26 a month, ≈ €315 a year.

Payback of a replacement. Enter the old and new average watts, the new hardware’s price and what you’d get for the old box. The output is months to pay back (or “never” if the new box doesn’t draw less). Example: 150 W → 30 W, €400 new, €100 resale → ≈ 11 months.

See it on your rack

Measured numbers are worth keeping next to the rack they describe. Add your Proxmox hosts to the Devices sheet (rack, U, model, name) and your wall or UPS readings to the Readings sheet. Homelab Twin then shows cost per host on a 3D model of your rack, the power chain through the UPS, and how long the UPS lasts at today’s load.

Sources

• Linux kernel documentation: CPU performance scaling, intel_pstate, amd-pstate
• Network UPS Tools: variable names (ups.realpower, ups.load, ups.realpower.nominal)
• hdparm manual: -S standby timeout encoding
• Proxmox VE forum: governor and amd-pstate threads (community reports)

Questions

Can I see power usage in the Proxmox web UI?

Not as watts. Proxmox shows CPU, memory, disk and network per node and VM, but not electrical power. Use a wall meter or smart plug, your UPS through NUT, or the server's IPMI/BMC, and graph it alongside Proxmox if you like.

Why does my Proxmox host use more power than the same machine on Windows?

Usually power-saving settings: the CPU governor or energy preference, deep C-states disabled in the BIOS, PCIe power management (ASPM) off, or disks that never spin down. Running VMs also keep the CPU from staying in deep idle. Check each one and measure after every change.

Does powertop --auto-tune help on Proxmox?

Often, by enabling power management on USB, SATA and PCIe devices. It can also make some network cards or USB devices misbehave, so run it once, test for a day, and only then make it permanent (for example with a systemd unit).

Is it worth replacing an old server with a mini PC to save power?

Work out the payback: (new hardware cost minus the old server's resale value) ÷ monthly electricity saving. Going from 150 W to 30 W average saves about €26 a month at €0.30/kWh, so a €300 net upgrade pays back in roughly a year.