HOW EVERY NUMBER HERE IS MADE
Every figure on this site comes from a file somebody uploaded, through code that is published. Nothing is modelled, estimated or filled in.
The population is the hardware attached to Steam reviews: 227,663 observations across 105 games. It says what people own. It does not say how their machines perform, and the two are never treated as the same claim.
We measure nothing ourselves. There is no synthetic test, no score of our own devising, and no ranking of parts.
free parameters on this page
These numbers were chosen rather than measured, and every one of them changes what this site tells you. Each is listed with what it does and why it is what it is, because a constant nobody wrote down becomes folklore. Where a row ends with a decision record number, that is an entry in DECISIONS.md in the published code, which gives the full argument and what was rejected.
| what counts as a stutter | twice the typical frame | A frame is called a stutter when it takes more than twice as long as this session's typical frame. Measured against this session rather than a fixed number of milliseconds, because a slow frame at 30 FPS and a slow frame at 200 FPS are not the same length. Chosen on two recordings and checked against a third that was made at a different power setting. Decision record 002. |
|---|---|---|
| how the session is cut up | about a tenth of the session | A stutter rate over a whole session hides when the rough parts happened, so the session is cut into blocks and each block gets its own rate. The block length is chosen to land near ten blocks and rounded to a whole number of seconds. Decision record 002. |
| when neither part is ahead | 5% apart | For each frame we compare how long the graphics card was busy against how long the processor was. If the two are within this much of each other, the frame is counted as neither one leading. At zero almost every frame picks a side on noise alone; at twenty percent almost every frame ties. Decision record 019. |
| how much a year of progress is worth | 3.0 places a year | Cards are placed on one ordering so a 2016 card and a 2023 card can be compared. A year of progress is worth this many places on it, which is what lets an older high-end card sit near a newer mid-range one. Fitted to four pairs of cards whose relative standing is not in dispute. Decision record 007. |
| where that ordering starts | 2012 | The year the card ordering counts from. It sets no card's position on its own; moving it would slide every card equally and change nothing. Decision record 007. |
| the laptop version of a card | -6.0 places | A laptop card with the same name as a desktop card is not the same part: it has less power and less cooling. It is placed this many places below its desktop namesake, which is one step of the ordering. Decision record 007. |
| when a driver counts as old | 365 days | A display driver older than this is reported as old. It is a line we chose, not something measured: nothing here knows whether a newer driver would have helped your session. Decision record 018. |
| when graphics memory counts as nearly full | 90% of dedicated | Memory use above this share of what the card carries is reported as close to full. The share is a line we chose; the figure printed beside it on your page is the measurement. |