Reproducible MangoHud benchmark protocol for Linux gaming

Scope

MangoHud can display and log performance metrics for Vulkan/OpenGL workloads. A benchmark is the whole controlled procedure, not the overlay screenshot or highest observed FPS.

Public log uploads can disclose filenames, hardware, paths, or other context. Store and analyze logs locally by default; sanitize before sharing.

Baseline

Freeze device/firmware, OS/kernel, GPU driver/Mesa, game build/branch, runtime/Proton, launch options, resolution, preset, cap/refresh, TDP/clocks, battery/AC state, brightness, radios, ambient temperature, and background tasks. Define a fixed scene, route, input script, warm-up, and logging duration.

Change one variable

Duplicate the baseline record and change one named field. Do not compare presets that also change resolution, cap, TDP, Proton, and display mode unless the whole profile is explicitly the variable.

Measurements

MetricReportGuardrail
FPS
Average and 1% low
Never publish peak alone
Frame time
Distribution/graph and spikes
Inspect pacing, not only summary
Power
Average over matching window
State sensor scope and AC/battery
Temperature
CPU/GPU range plus fan/noise
Warm up and watch thermal drift
Battery
Start/end energy or charge and duration
Same charge band and radios

Run warm-up separately, then at least three equal logged passes. Reject invalid passes with a written reason. Use the median and include run-to-run spread; do not average incomparable scenes.

Side effects

Overlay/logging has overhead; sensors can update at different rates; first-run shaders distort frame times; a battery run changes voltage/temperature over time; thermal soak can make later runs slower; a benchmark route driven by hand introduces input variance.

Rollback

Disable logging/overlay launch options, restore the original game and device profile, stop any test-only sensor or plotting process, and retain raw logs plus the manifest. Delete public uploads through the service's current process when a log was shared unintentionally.

Result manifest

Publication rule

Publish raw-log hashes or archived logs where privacy allows, the complete manifest, exclusion reasons, calculation method, and evidence date. “Feels smoother” can accompany measurements but cannot replace them.

Frequently asked questions

How many runs are enough for a comparison?
How many runs are enough for a comparison?
Use at least three equal, valid runs and report the median. Add runs when shader compilation, thermal drift, scene variance, or an outlier makes the result unstable.
Are MangoHud uploads private?
Are MangoHud uploads private?
Do not assume so. MangoHud's documentation says logs uploaded to FlightlessMango.com are public. Prefer local analysis and review filenames and log contents before sharing.

Sources

Version history

  • 2026-07-15: Phase 5 reproducible benchmark and privacy protocol published.