Quick answer

Run-to-run variance is the normal movement between repeated benchmark results. It must be smaller than the improvement being claimed. Warm-up state, background activity, boost behavior and random scene events can create a difference that looks like an optimization even when nothing meaningful changed.

Use the result as a hypothesis

No online score observes your clocks, temperatures, software version, memory pressure or exact scene. Confirm the suspected stage with a repeatable capture before changing hardware.

What can cause this performance pattern?

A useful diagnosis begins with several competing explanations. The following causes can produce similar headline utilization while requiring different fixes.

Signals worth recording

  • A cold cache changes the first pass.
  • Background updates create intermittent CPU and storage work.
  • Dynamic weather or opponents alter scene load.
  • Temperature changes sustained boost behavior.

Record the event in the workload where it matters. A lightweight menu, loading screen or empty benchmark scene cannot represent the busiest part of a game or application.

Run a controlled test

Keep the application, scene, software version and warm-up state fixed. Change one input, record the response and restore the baseline before the next test.

Step-by-step test plan

  1. Restart from the same checkpoint or benchmark preset.
  2. Run enough passes to estimate the normal range.
  3. Use the median and keep individual results.
  4. Reject captures with a documented interruption.

Repeat the capture at least three times when results vary. The measured change should be larger than normal run-to-run movement before it is used as buying evidence.

How to interpret the evidence

Observation, test and meaning

ObservationControlled changeLikely interpretation
Tight repeated rangeApply one setting changeA larger shift is more credible
First run is always slowerAdd a warm-up passCache or compilation affects the sample
Results drift downwardLog temperature and clocksSustained operating conditions are changing

The interpretation is directional. More than one limit can appear in a single workload, and the slowest stage can move after a setting, cap or hardware change.

Turn the result into a useful decision

A five-percent uplift is not persuasive when untreated results move by six percent. Stabilize the procedure first. Once the normal range is known, compare medians and frame-time distributions rather than choosing the best run from one configuration and the worst from another.

Article-specific measurement worksheet

Label these fields in the capture notes so another run can reproduce the same question:

  • confidence range
  • median result
  • warm-up pass
  • cache state
  • ambient condition
  • background interruption
  • result spread
  • repeat count

Common mistakes to avoid

  • Publishing only the fastest run.
  • Changing drivers and settings together.
  • Rounding away a small but unstable difference.

Short summary: define the target, capture the exact problem, change one variable and buy only when the expected stage responds consistently.

Continue the evidence path

The following resources connect this test to the broader bottleneck topic and the relevant comparison tool.

Check the PC performance magazine for the latest field guides, use the diagnosis FAQ when signals disagree, and compare reviewed CPU profiles with GPU profiles before changing hardware.

Run the CPU and GPU calculator

Related measurement guides

Build the cluster one verified step at a time

Technical sources

References for the measurement method