Measure home-server power: idle, inference and storage

Try in your browser: watts, watt-hours and logging interval

The Lab above integrates a synthetic power log in your browser. “Saved: 40 W over 600 s” reproduces the original CSV: 601 samples yield 6.666666… Wh. Save a result as A, change power or logging interval, and compare. No Python installation is required.

The observation window is fixed at 0–600 seconds. Set synthetic baseline power from 0 to 200 W and additional power at the peak from 0 to 400 W. The added load is a triangle centered at 305 s, with half-width h from 1 to 120 s and total width 2h. Select a logging interval of 1, 2, 5, 10, 30 or 60 s; logging always starts at zero. This is a known-waveform integration exercise, not a server model predicting power or workload duration.

Save “2-second load, 1-second logging” as A, then compare “10-second logging.” The 305 s peak falls between samples and the added load is missed completely. With “5-second logging,” a sample lands on the apex but the trapezoidal interpolation spreads it over too much time. Peak timing is fixed: coarser logging does not always change error in the same direction.

The top panel shows watts and samples; the bottom shows cumulative watt-hours. Blue represents recorded samples and their trapezoidal integral; brown is the synthetic reference. A is lighter/dashed and B darker; agreeing lines overlap. Exact reference energy follows the triangle's area: over the whole window, divide “baseline power × 600 + added peak power × h” by 3600. Cumulative reference energy uses the partial triangle area up to each time. The sign of sample integral minus reference distinguishes overestimation from underestimation. These references come from the known synthetic formula, not reconstruction of an unknown measured waveform.

The original integration function is reused unchanged. Mean W is time-weighted and peak W is the largest sample. Incremental energy subtracts baseline power × 600/3600 from integrated Wh; it does not isolate CPU/GPU component energy. Inputs are instantaneous point samples, not interval averages. No missing samples are generated; the analyzer's maximum permitted gap is explicitly set to twice the chosen logging interval, preserving the original 2 s limit for default 1 s logging. Original rejection rules for duplicate/reversed times, nonfinite values, negative power and excessive gaps remain intact.

The numerical table retains all samples and reference values; shared conditions and results JSON preserve the experiment. This Lab has no CSV upload or hardware measurement. Original CSV, source and expected results are preserved; local reproduction below is optional. Save the browser entrypoint, original analyzer and synthetic CSV together to run using only the Python standard library.

For a home server, consider long idle periods and energy per completed workload as well as peak watts. This article provides a plug-in-meter logging method. No real server was measured; the attached 40 W data is synthetic arithmetic validation.

Fix the measurement boundary

Specify whether the meter covers only the server, UPS losses, external storage or networking equipment. Use a suitably rated meter; no internal wiring modification is required.

Read active power W and energy Wh. Voltage times current gives apparent power VA; power factor matters for AC loads. Do not treat V×A as consumed watts without that distinction. See Yokogawa’s power-measurement fundamentals for active/apparent power and power factor. Record meter resolution, accuracy, logging interval and low-load range from its manual.

Diagram 1 · Use the button to switch views

Define an idle state

Avoid startup updates and indexing. Record services, CPU settings, GPU model residency, disk rotation state, display, fans and room temperature. Choose a settled interval, not just a low CPU percentage. A ten-minute window can miss a task that runs every thirty minutes; begin with a longer log to identify cycles.

Spun-down and rotating disks are separate conditions. Energy savings, first-access latency and start/stop count are distinct evaluation axes.

Hold inference and storage work fixed

For AI, fix model, quantization, prompt, output limit, concurrency and cache state. Use local-LLM measurement to align request start/completion and generated counts with the power log. Separate loading-inclusive from resident-model tests.

For storage, fix file count, bytes, reads/writes and cache state using dedicated test data. Define completion as application return or synchronization completion. Cache-hit reads do not measure the disk alone.

Integrate watts over time

With timestamps t in seconds and power P in watts, a trapezoidal approximation gives:

Divide Wh by duration in hours for mean W. Constant 40 W for 600 s gives 6.667 Wh. Assuming 40 W continuously for 24 hours over 30 days gives 28.8 kWh. These are arithmetic examples, not this site's measured usage or a bill forecast.

Optional: reproduce CSV integration locally

Save power_log.py and the synthetic CSV together. Execution was checked with Python 3.12.3 using only its standard library:

python3 power_log.py server-power-synthetic.csv

Expected output: duration_s=600, mean_W=40, energy_Wh=6.666666… and samples=601. The validation result is available. Use time_s,power_W columns and one continuous interval per file. The tool rejects duplicate times, nonfinite values, negative power and gaps over 2 s. This default assumes 1 Hz logging; explicitly change --max-gap for another interval.

peak_sampled_W is the largest sample, not a guarantee about between-sample peaks. A meter reporting interval averages may require integration matched to that output convention.

Interpret comparisons

Report total Wh separately from incremental Wh after subtracting idle mean W times workload duration. Subtraction assumes the idle state would have persisted and does not precisely isolate CPU, GPU or disk energy. Fans and power-supply efficiency can change even when one workload setting changes.

Higher watts can still yield lower Wh if equal work finishes sooner. Long idle periods can also outweigh small savings on short jobs. Include boundary, environment, logging interval, repetitions, duration, mean W, Wh and completed work in each comparison.