Read the metric before naming the limit
Average FPS describes throughput across a sample. A 1% low summarizes the slower end of delivery and often tracks perceived consistency better, but it is sensitive to capture length and scene selection. Use both metrics with the underlying frame-time trace rather than turning either into a universal score.
A metric is useful only when the capture window, scene and warm-up state are documented. First hypothesis to test: Short captures make percentile values unstable.
Match the trace to the controlled change
Read each observation with its matching control
| Observation | Controlled change | Likely interpretation |
|---|---|---|
| High average, weak lows | Inspect spike locations | Stutter or intermittent work controls the experience |
| Both values low | Reduce graphics load | Sustained throughput may be insufficient |
| Large run variance | Standardize the route | The benchmark is not yet reliable |
When the observation is “High average, weak lows,” apply “inspect spike locations.” The evidence table reads the response as follows: Stutter or intermittent work controls the experience.
Capture conditions that can distort the reading
Compare two early explanations: (1) Short captures make percentile values unstable; and (2) Menus and loading screens distort the slow tail. They can resemble one another in a headline result but predict different responses to the checks below.
Variables that change the comparison
- Short captures make percentile values unstable.
- Menus and loading screens distort the slow tail.
- Shader compilation can punish the first run only.
- Different routes create different CPU and GPU demand.
Keep the sampling tool and route unchanged so a difference in the trace reflects the tested variable rather than a new capture method.
Create a repeatable performance capture
Use the first two controls in order: Warm the game before the measured route; then Capture at least three comparable runs. This establishes whether the first observation survives a controlled comparison.
Matched-run protocol
- Warm the game before the measured route.
- Capture at least three comparable runs.
- Report average, 1% low and sample duration.
- Inspect the slow frames instead of trusting the percentile alone.
Repeat until the normal run-to-run range is visible; a claimed change must be larger than that range.
Set a measurement-based action rule
Upgrade decisions need a stable target. If the average meets the display goal but the lows remain poor, investigate memory, background work, compilation and temperature before buying a faster GPU. If both move strongly with resolution, graphics throughput is a better lead.
Comparison record
Record sample duration, warm-up pass, route, cap and every reported unit beside the result.
- sample population
- percentile cutoff
- arithmetic mean
- slow tail
- outlier frame
- benchmark duration
- median run
- distribution shape
Checks that can overturn the decision
- Quoting a 1% low without sample length.
- Mixing built-in and manual benchmark routes.
- Assuming every slow frame comes from hardware.
Decision boundary: The decision above should also survive the most likely shortcut: quoting a 1% low without sample length.
Related measurement guides
Build the cluster one verified step at a time
Compare this result with PC Benchmark Run-to-Run Variance Explained, CPU Time vs GPU Time: Find the Slower Frame Stage, and How to Read a Frame-Time Graph; keep the same workload and baseline while testing the competing explanations.
Technical sources
References for the measurement method
