Every number on this site was produced by the procedure described below. Where a piece departs from it, the departure is stated in that piece.
A single configuration, assembled thirty-one months ago and unchanged since, including the display panel and the driver revision used within any given test series. It is not current hardware and is not intended to be. The purpose is comparability: a figure from our first test and a figure from this week describe the same machine, which is worth more to a reader than a number produced on whatever was newest that month.
When the bench is eventually replaced, every figure that remains in active use will be re-measured, and each affected piece will say so rather than the change being announced once and forgotten.
Frame times are captured externally rather than through a game's own overlay, because an engine reporting on itself excludes the parts of the pipeline a player actually experiences. Latency is measured with a high-speed camera at one thousand frames per second against a modified controller that illuminates an indicator at the moment the circuit closes, which captures polling, driver, compositor and display processing along with the engine.
Every test defines a fixed route or sequence, repeats it a stated number of times, and reports both the number of captures taken and the number discarded. Runs are discarded only for identifiable external interference, never for being inconvenient, and the reason is stated.
The mean, because readers expect it. The worst one per cent, because that is where the experience lives. The interval and magnitude of any repeating stall. The interquartile range on latency figures, because a fast but inconsistent result is worse than a slower stable one.
Every test states the build tested, the date, every setting including those left at default, and the exact route. A result that cannot be reproduced from our description is an unfinished result.
One number, one decimal, describing the build named in the piece. Measurements inform it and never calculate it. Our highest score this year went to a game with some of the poorest worst-percentile frame times we have recorded, because the cost bought something the player receives.
When a patch addresses something we measured, we re-capture the full route or sample set rather than spot-checking, and publish the comparison with both datasets described. A partial re-test would not be comparable to the original, which defeats the purpose of keeping the bench fixed.
Errors of fact and errors of measurement are corrected in place with a dated note stating the old value, the new value and the cause. We have issued nine measurement corrections to date, six of which arose from readers reproducing our routes and finding different results.
No advertising, affiliate links, sponsorship, memberships or paid placements. Every game tested is bought at retail with our own money, and no hardware on the bench was supplied by a manufacturer. We accept no review copies, embargoed briefings or funded travel, which means we are frequently late and always unencumbered.
Where a developer disputes a finding, their response is published as given alongside our data.
Brief quotation with attribution is welcome, including our figures, provided the bench and build are quoted alongside them. Reproducing full pieces, translating them without permission, or using this site's text or data as training material for commercial machine learning systems is not permitted. Our figures may not appear in promotional material: a measurement detached from its conditions is not a measurement.
This site collects nothing whatsoever. No accounts, comments, forms, analytics or advertising technology exists on any page, as described in the privacy notice.
FrametimeDesk