Two people measure the same server and come back with two different numbers. Usually neither of them is wrong. They just measured different things.
The short answer
Before you compare power readings, check that they were taken at the same point, which is called the measurement boundary. Then check that the meter suits the load it’s measuring and say how uncertain its reading is. More digits on the display don’t automatically mean a more accurate number.
How it works
- Measurement boundary: it defines what a value includes. That might be one component, the server’s own sensors, one of several power supplies, or the whole system at the wall socket. Values from different boundaries can’t be compared directly.
- Measure at the wall the way test labs do: in its maximum-power test, the ENERGY STAR test method for computers plugs the device straight into the meter’s outlet and allows no power strip or UPS (uninterruptible power supply) in between. Otherwise you also measure whatever else hangs off that strip.
- Real power, not just current: the meter should show real power in watts. Voltage times current gives apparent power in volt-amperes (VA), which is a different quantity.
- Range and low load: a meter’s uncertainty depends on how well its voltage and current ranges fit the value. SPEC, which runs the server energy benchmarks, notes that analyzer accuracy depends on the range setting and the load being measured, and doesn’t accept auto-ranging for compliant runs because samples can be lost while the range switches.
- Resolution vs. accuracy: resolution is the smallest step the display shows. Accuracy is how close the value is to the truth. ENERGY STAR, for example, requires 0.1 W resolution between 10 and 100 W, but an uncertainty of up to 2 % (at 95 % confidence) is still allowed.
- Repeats and specifications: repeating a measurement shows you the spread, which NIST calls a Type A evaluation. Uncertainty from the meter’s datasheet or calibration certificate is a Type B evaluation. A solid report states the value, the observed spread and the uncertainty from the specifications separately.
In practice
Suppose your plug-in meter shows 47.83 W (an assumed reading for illustration) and its manual promises ±2 %. That’s about ±0.96 W, so the honest result is “47.8 W ± 1.0 W”. The second decimal place is display, not information. If your server has two power supplies, both are part of the boundary “whole system at the wall”. Measuring one cord only gives you half the picture.
Three places to read power, three different answers
In a homelab you can usually get a number from three sources, and they rarely agree:
- Internal sensors (BMC/IPMI, `ipmitool sensor`, vendor tools): the server reports what its own power supplies or voltage regulators see. Cheap and continuous, but the boundary is inside the machine: PSU losses, fans on a separate rail and anything the board does not meter are missing. Vendors also smooth or sample these values differently.
- A metered PDU or smart plug: sits between the wall and the device. Good for long series and remote logging; check what it reports (W or VA, average or instantaneous) and what its datasheet promises at your load, because many are specified for several hundred watts, not for 12.
- A dedicated plug-in power meter: the reference for one device at one time. It is the only one of the three where you can read the accuracy class from the manual and plug the device in directly, as the lab methods require.
Use the sensors for trends, the PDU for the log and the meter to calibrate both. And write down which one each number came from.
Instantaneous value vs. average
A display refreshes once or twice a second and shows a snapshot. Server load jumps: a fan ramp, a disk spin-up or one busy core moves the needle by tens of watts for a few seconds. SPEC’s measurement setup therefore reads the analyzer at regular intervals and averages over the whole run instead of trusting a single reading. For your own log the rule is: note the sampling interval and the averaging window with the value. “48 W” means little; “48 W, mean of 1-second samples over 10 minutes at idle” can be reproduced.
Where cheap meters stop
Below a few watts, many plug-in meters either show zero or a number with no stated accuracy at all. Standby of a switched-off desktop, a sleeping NAS or a wall-wart is exactly where that happens. If a device reads 0.0 W, do not write “0 W” into your sheet. Write “below the meter’s floor (spec: X W)” or measure several identical devices on one strip and divide, which pushes the total into the range the meter is specified for.
Temperature, time of day, and the state you measure in
Fans spin faster in a warm room and the PSU runs a little less efficiently, so the same server reads a few watts higher on a summer afternoon than on a winter night. “Idle” also needs a definition: ten minutes after boot the system is still indexing, scrubbing or updating. Wait until the load average and disk activity settle, note the room temperature and the uptime, and compare only readings taken under the same conditions.
How to state a result
<device>, <boundary>, <meter model + accuracy>, <load state>, <mean ± uncertainty>, <sampling/averaging>, <date> – for example: “DL360 Gen9, whole system at wall, Meter X (±2 %), idle after 15 min, 112 W ± 2.5 W, 1-s samples averaged over 10 min, 18 Oct 2026”. Anyone can repeat that. “The server draws about 110 W” nobody can.
Checklist
- Decide on the measurement boundary before you start, and write it next to every value.
- Plug the device directly into the meter, without a power strip or UPS in between.
- Note the meter model, firmware and range setting.
- Check whether the meter shows real power in W or only current and VA.
- Look up the meter’s stated accuracy for your load range and round your result to match.
- Repeat the measurement and record duration, number of runs and room conditions.
- Never write “0 W” for a reading below the meter’s specified floor; note the floor instead.