How we test and rank GPUs
Where the numbers come from, and how much to trust each one.
The scale
Every GPU sits on one scale: its average frame rate across a suite of 19 current games (8 of them with ray tracing), at ultra settings and native resolution, on a fast gaming CPU so the graphics card is the limit. Averages are geometric means, so no single game dominates. Upscaling (DLSS, FSR, XeSS) and frame generation are off, because they change image quality and make cards hard to compare.
The overall score blends rasterized performance (75%) with ray tracing (25%) and is scaled so the GeForce RTX 5090 is 100. GPUs without hardware ray tracing are scored on rasterized performance alone.
Where each figure comes from
Tested directly (48 cards). Recent desktop cards measured in the current game suite, cross-checked against several independent reviewers, which agree within a few percent.
Rescaled (49 cards). Older desktop cards measured in a previous game suite, converted to today's scale using cards measured in both suites, preferring siblings of the same architecture.
Estimated, marked “est.” (153 cards). Laptop GPUs (45) and integrated graphics (27) are placed from their typical 3DMark Time Spy scores across many reviewed laptops, using the relationship between Time Spy and real game performance on desktop cards (typically within ±7%). Laptop results depend heavily on the power limit the manufacturer sets. Desktop cards from before about 2016 are estimated from compute power and how efficient their architecture has proven against measured cards; many of them can't start games released since 2023 and are labelled accordingly.
Corrections
When a single result disagrees with every sibling card and with other reviewers, we smooth it rather than publish an obvious outlier. If you spot a number that looks wrong, tell us.
How we make money
Amazon buttons are affiliate links: if you buy through them we may earn a commission, at no extra cost to you. This never affects the numbers or the rankings.