Average frame rate is not what smoothness feels like. The brain does not perceive an average; it perceives the worst moments. A game that averages ninety frames per second but drops to twelve for a quarter of a second feels broken — and the average will not tell you that.
What the numbers represent
FROM THIS ENTRY| average frame rate | mean of all frame durations; hides variance |
|---|---|
| one per cent low | mean frame time of the slowest one per cent of frames in a captured run |
| zero point one per cent low | mean of the slowest 0.1%; isolates extreme outliers |
| per-frame capture | recording every individual frame duration, not sampled averages |
The one per cent low is the answer to that. Take the full frame-time distribution for a captured run, sort every frame from fastest to slowest, and find the slowest one per cent. The average frame time of that bottom one per cent, expressed in milliseconds or converted back to a frame rate, is the one per cent low. It is not the single worst frame; it is the sustained character of the worst region. That distinction matters, because a solitary outlier spike often registers as a visual artefact rather than a smoothness problem, while a cluster of slow frames is exactly what the visual system notices as stutter.
The metric became widespread alongside tools that record per-frame data rather than sampling at intervals. Interval sampling — averaging over a fixed window — can straddle a hitch and report it away. Per-frame capture cannot. Once you have every frame's duration, the one per cent low emerges from a sort and a mean; nothing exotic is required.
Where the number is meaningful is in separating where the milliseconds go into their real causes. A healthy average with a collapsed one per cent low points to something episodic: shader compilation stutter, a pipeline stall, a garbage collection pause, a translation-layer hiccup at an unexpected code path. A low average with a one per cent low that tracks proportionally suggests steady overhead, not spikes — a different problem entirely.
The limitation is what any single statistic carries: it compresses shape. Two configurations with an identical one per cent low can feel different if one's worst frames are clustered and the other's are scattered. The zero point one per cent low — the bottom tenth of the bottom tenth — goes further, isolating the most extreme events. For most practical diagnosis, however, the one per cent low is the first number worth reading after the average, precisely because it describes the experience the average was designed to conceal.
