Each headline figure carries a 95% interval: the range it would be expected to fall in if the same run's windows were drawn again.
- It is the spread inside one run. It is not the spread between runs or between machines, and a second run can land outside it.
- The interval comes from a moving-block bootstrap: 1000 resamples of whole blocks of neighbouring windows, from a fixed seed.
- Figures with too few values for a bootstrap, such as memory latency and the storage passes, show the range of what was measured instead.
Why a single number is not enough
A stage measures in short windows, and no two windows read exactly alike. A single sustained figure hides how much they varied. Since measurement revision 2026.09.16 each headline figure carries an interval so you can see that spread: CPU sustained throughput, GPU sustained frame rate, path tracing samples per second, memory bandwidth on one thread and on all threads, the soft-shadow rates and the WebGPU compute median.
How the interval is computed
Neighbouring windows are not independent: a window that ran slow tends to be followed by another slow one. The textbook interval assumes they are independent and comes out too narrow on such a series. SystemCheck uses a moving-block bootstrap instead.
It draws blocks of consecutive windows, with replacement, and lays them end to end until the resample is as long as the run, so the short-range dependence inside each block is kept. The block length grows with the cube root of the number of windows, is lengthened when the series is strongly dependent, and is never under 2. This is repeated 1000 times, and the interval is the middle 95% of the results. The seed is fixed, so the same run gives the same interval on every machine.
With fewer than 8 windows there is too little to resample and no interval is given.
What the interval does not cover
The interval describes this run's windows. It does not describe what a second run would read, because a second run can start warmer, meet different background work or be on battery instead of mains. It does not describe other machines of the same model either. To compare two runs, match their conditions and repeat each one; two intervals that overlap widely are not evidence of a difference.
Memory latency, the soft-shadow cost and the storage figures have only a handful of values. For those the report shows the lowest and highest values measured, and labels it as a range rather than a confidence interval.
Terms used here
Questions
- My two runs' intervals do not overlap. Did something change?
- Possibly, but first check the conditions: power source, background apps, how warm the machine was and whether the tab stayed in front. The interval covers the spread inside a run, not the change between runs, so different conditions can move the figure outside it.
- Why does the interval sit off-centre around the figure?
- A bootstrap interval follows the shape of the data rather than assuming a symmetric spread, so it can extend further on one side.