UNDER THE HOOD
Good numbers. Clear ground rules.
A benchmark is useful when you know what it measures. Here is where our figures come from, how the maths works and which questions need more than one FPS reading.
One continuous 1080p workload
GPUvsGPU uses the benchmark figures supplied by GPUsed from its own GPU testing records. The recorded workload is a continuous FurMark run at 1920×1080. GPUsed reports that these runs typically last 30 to 60 minutes, alongside its checks for heat and instability.
Each included variant has one supplied benchmark FPS average. The comparison uses that figure consistently across the bench, rankings, profiles and detailed match-ups. We do not create individual test runs, temperatures or test dates to fill in the history.
The retained source does not include a complete per-run record of test rigs, drivers, FurMark versions or anti-aliasing settings. That limits cross-lab reproducibility. Use the figures to compare this recorded workload, and use game-specific testing for the games and settings you actually play.
Why swapping changes the percentage
The bench always describes GPU X relative to GPU Y. GPU Y supplies the reference:
(GPU X FPS ÷ GPU Y FPS − 1) × 100
A positive result means X leads; a negative result means Y leads. Only identical original FPS values produce 0% Level. We calculate with those original figures and round the displayed percentage to one decimal place.
For a real example, RTX 3060 12GB records 142 FPS and RTX 3060 8GB records 120 FPS. Put the 12GB card in X and the 8GB card in Y: the result is +18.3%. Reverse them and it becomes −15.5%. The cards have not changed; the reference has.
Our rating out of 100
The rating makes the benchmark scale easier to read across several GPU generations. The reference result is 784 FPS, and the formula is:
100 × (GPU FPS ÷ 784)0.35, rounded to one decimal place.
This compresses the spread between older and newer cards. It is a consistent transformation of benchmark FPS, rather than an arbitrary mark for how good a GPU feels. A rating of 50 does not mean half the FPS of a card rated 100. Percentage comparisons always use raw FPS.
What “we’ve tested” tells you
These are GPUsed’s supplied historical totals. An exact-model total belongs to the recorded model. A shared-family total is labelled explicitly and may cover several variants in that family; it is not a separate quantity for every memory version.
The historical count is distinct from the retained sample size of the displayed FPS average. It does not establish which individual readings contributed to that average. The site shows the available average and historical total without inventing a dropdown full of unit-level results.
Names, releases and GPU personality
Choose the full variant, including memory capacity, Ti, Super, XT, XTX and any documented memory or BIOS distinction. Similar names can hide different configurations. The current catalogue contains 257 variants with supplied FPS; a missing result is not filled in from another lab.
Release dates use the documented month and event shown in the profile, with their source and scope. A first documented retail listing for one board or market is not automatically the worldwide launch date of every card sharing its GPU.
The GPUsed Legend Rating, archetype and occasional nickname are editorial character and legacy judgements. They are separate from benchmark performance and are not an audience vote. Profile references support the identity and context; GPUsed supplies the benchmark records.
An upgrade is more than a bigger number
Use the performance difference as one part of the decision. Your games, resolution, processor, memory requirements, software features and the price of the actual card all matter. Check the exact board’s dimensions, PSU requirements and connectors before buying it.
Spotted a result or identity that needs another look? Send the exact variant, page link and the evidence through our corrections route. If you want to discuss your own upgrade or sell your current card, contact GPUsed.