How to compare graphics cards properly

Two cards can swap places depending on the resolution you judge them at. A card with a wide memory bus often gains on a rival as the resolution rises, and a card with fast cores can lead at 1080p and fall behind at 4K.

Pick the resolution you will actually play at and compare there. The comparison tool defaults to 1440p and lets you change it.

Core counts and clock speeds look like the natural way to compare cards, and they are the wrong one. A shader or core is a different piece of hardware in each company's design, and it is not even comparable across generations from the same company. Clock speeds are no better, because what each clock cycle accomplishes varies by architecture.

A performance index, built from measured results across many games, summarises what the card actually does. Compare indices first and treat specifications as the explanation of the result rather than the evidence for it. The indices on this site are estimates aggregated from public review coverage, and each page labels them that way.

A price is only comparable to another price of the same kind. A verified street price against a launch MSRP is two different kinds of number, and a gap between them tells you about the market, not about the cards.

Look at the basis shown next to each price. If both are verified, compare directly. If one is an MSRP, check the live listing before drawing any conclusion.

Memory capacity decides how long a card stays useful, and bandwidth decides how well it handles high resolutions. A card with more capacity but a narrower bus is not automatically better, and the reverse is also true.

Capacity is the safer thing to weigh when buying for the long term. Bandwidth explains why a card performs better at 4K than its 1080p result suggests.

Last, check what the card can do and whether it fits. Upscaling and frame generation support differ by company and change what a card can play. Board power determines the supply you need. Physical size decides whether it fits your case at all.

These do not change the ranking, but they decide whether a card is a sensible purchase for your build. The comparison table groups them so they are easy to scan.

A comparison that names a winner in every row is being dishonest. Real pairs are mixed: one card leads on speed, the other on memory or price. The useful output is which card wins on the priority that matters to you.

That is how the verdicts on the head-to-head pages are built. They name a winner per priority, state the number that decided it and say when two cards are too close to call.

The most common mistake is treating one number as the whole story: picking the card with more memory, or the higher clock, or the higher benchmark in a single game. Each of those is a real input, and none of them is the answer by itself.

The next is comparing cards from different tiers because they share a name. A smaller sibling of a product line can have the same family name, a different chip and a very different price, and listings that differ only in memory size can hide a large gap in speed as well.

Prices are the third trap. Two prices from different countries, currencies or stores are not directly comparable, because tax treatment, shipping and exchange rates differ. Compare like with like, or convert both to one currency and say so.

Finally, remember that board partners sell many versions of one card, with different coolers, clock speeds and sizes. The reference specification is a starting point, and the exact model you buy decides noise, temperature and whether it fits.

The comparison tool takes up to four cards and builds the table described above: price and value first, then performance, memory, power, features and physical size. Each area folds open on its own, so you can read one topic at a time.

Rows where one card leads by more than 3% mark a winner. Rows where the cards are within that margin are marked as level, and rows where a card has no published figure say so instead of guessing. The biggest differences are summarised at the top for the pairs that have their own page.