Methodology
Every number on this site falls into one of three categories: a manufacturer specification, a cited third-party measurement, or a CompareYourGPU estimate. The third category is always labelled. This page explains where each kind comes from, how the derived scores are calculated, and what CompareYourGPU refuses to publish.
- Graphics cards
- 35
- Games profiled
- 12
- Measured benchmark runs
- 0
- Verified current prices
- 0
Three kinds of number, never mixed
The most important distinction on this site. Each kind is labelled at the point of use, not in a footnote.
From the manufacturer
Core counts, clocks, memory, board power, connectors, outputs and launch MSRP, taken from the vendor’s own product page. 18 of 35 cards have had their table read directly off that page and carry the date it was checked; the rest are pending re-verification.
Measured by a named third party
Frame rates published by review outlets. CompareYourGPU holds 73 suite-level measured ratings covering 25 of 35 cards from ComputerBase, plus 0 individual per-game runs. Suite ratings are shown on each card’s page next to the estimate, never merged with it.
Produced by our model
The relative performance index and everything derived from it, including frame-rate ranges, value and efficiency. 35 of 35 cards carry one. These are aggregated from public review coverage, not measured here, and are labelled wherever they appear.
Where both a verified benchmark and an estimate exist for the same card, game and resolution, the verified figure is used and the estimate is shown separately for context. They are never averaged together, and a measured run taken with upscaling enabled is never compared against a native estimate.
The three rules
These are constraints on the codebase, not editorial aspirations.
Nothing is invented
No frame rate, price, temperature, noise level, rating or review appears anywhere on this site unless it comes from a manufacturer, a cited measurement, or a documented model that says so. There is no code path that generates a benchmark figure.
Missing data is shown as missing
Every field that can be unknown is typed as such, and every component renders “Data unavailable” for it. A blank cell, a zero, or a quietly omitted row would all misrepresent the state of the data.
Estimates are labelled at the point of use
Provenance is attached to the number, not buried in a footnote. If you see a performance index or a frame-rate range, the label saying it is a CompareYourGPU estimate is next to it.
The performance index
The single most important number on the site, and an estimate.
Each card carries a relative performance index at each of the three resolutions, where 100 is the fastest card in the database at that resolution. A card at 50 is roughly half as fast. The indices are aggregated from public review coverage across many outlets and titles; they are not CompareYourGPU measurements, and they are tagged as estimates in the data model itself so that the interface cannot display one without saying so.
Ray-traced performance is stored as a separate index, because it does not track rasterisation performance: two cards that are level in a normal game can be far apart once ray tracing is enabled. Every page that compares ray tracing uses the separate index rather than applying a blanket penalty to the rasterisation figure.
The index exists to rank cards against each other and to size a card against a resolution. It is not precise enough to predict a specific frame rate in a specific game, and nothing on the site presents it as if it were.
The frame-rate model
How an index becomes a frame-rate range, stated in full so you can judge it.
A frame-rate range is produced by multiplying the performance index by an anchor value for the resolution, then by a multiplier for how demanding the game is. The anchor is the approximate average frame rate a card with an index of 100 reaches in a typical modern AAA game at the High preset, at native resolution, with no upscaling and no frame generation, paired with a current high-end processor.
The result is reported as a range of plus or minus 18%, rounded to the nearest 5 FPS. Game-to-game variance is far wider than that band; the range communicates “roughly here”, which is the honest precision of a model like this. Where a specific title is involved, the demand multiplier for that title is applied on top.
Two things the model deliberately does not attempt: it does not predict 1% low frame rates, because those depend on memory pressure and engine behaviour that an index cannot capture; and it does not model upscaling gains, because those vary by implementation, by quality mode and by title.
Anchor frame rate at index 100
- 1080p
- 290 FPS
- 1440p
- 220 FPS
- 4K
- 125 FPS
Game demand multipliers
- light
- 2.60x
- moderate
- 1.50x
- heavy
- 1.00x
- extreme
- 0.68x
Value, efficiency and VRAM scores
Three derived measures, each with a stated definition.
Value
Performance index points per $100, calculated at the relevant resolution. The raw figure is shown directly on GPU pages. For ranking, it is normalised to a 0-1 scale against every other card in the database, so a value score answers “where does this sit relative to everything else” rather than being an abstract number.
Uses launch MSRP unless a verified current price exists.
Efficiency
Performance index points per 100 W of board power at 1440p, normalised the same way. It uses the manufacturer’s published board power figure, which is a design target rather than a measured draw, so treat it as comparative rather than absolute.
VRAM headroom
Scored against editorial guidance per resolution rather than against the largest card on the market. Below the practical floor scores poorly; at or above the comfortable figure scores well, with a small bonus for capacity beyond it.
- 1080p
- 8 GB floor, 12 GB comfortable
- 1440p
- 10 GB floor, 16 GB comfortable
- 4K
- 12 GB floor, 16 GB comfortable
Value is three numbers, not one
A card can be excellent value for frame rate and poor value for memory. Collapsing that into a single score hides the thing a buyer needs to know.
Performance value
Performance index points per $100 at the relevant resolution, normalised across the database. Weighted 60% in the blend.
VRAM value
Gigabytes of memory per $100, normalised the same way. Weighted 20%, because capacity is usually what ends a card’s useful life before raw speed does.
Efficiency value
Performance index points per 100 W, which is money the card saves over its life and heat it does not put into your room. Weighted 20%.
The blend drops any component it cannot compute and renormalises over the rest, so a missing figure never drags a card’s value down. A card with no price at all is excluded from value ranking entirely rather than scored zero.
Good, fair or expensive
Where the price thresholds on each GPU page come from.
CompareYourGPU fits a straight line through price against 1440p performance index across every card with both figures, then reads off what this card’s performance would cost if it sat exactly on that line. That is the fair price. It describes the market as it currently is; it is not a recommendation, and it moves whenever the underlying prices move.
A card at or below 92% of its fair price is called a good buy. At or above 112% it is called expensive. Between the two it is fair. Those two multipliers are the only arbitrary numbers in the calculation and they live in one config file.
Because CompareYourGPU currently holds no verified street prices, the line is fitted through launch MSRPs. That makes it a fair comparison of what manufacturers intended to charge, not a claim about today’s market, and every page that shows a price verdict says which basis it used.
How recommendations are calculated
Deterministic, weighted, and shown in full on every result.
Every card is scored on six criteria. Each criterion produces a 0-1 score, and each has a weight that depends on your answers. The match percentage is the sum of score multiplied by weight across all six, expressed out of 100. The results page prints that table, so the number is never presented without the arithmetic behind it.
Weights start from a neutral base and are then adjusted four times: by your stated priority, by your resolution, by how much ray tracing matters to you, and by how ambitious your frame-rate target is. A 4K user gets more weight on performance and VRAM; a budget-focused user gets more on value; someone who says ray tracing is unimportant sees it drop to near zero.
Two hard filters apply before scoring. Cards out of production are excluded, because recommending something you cannot reliably buy new is not useful, and a manufacturer preference is honoured exactly. If nothing fits your budget, the search widens by 15% and the result says so rather than quietly ignoring the constraint.
Frame-rate targets are matched by running the frame-rate model against the most demanding game you selected, not the average. That makes the recommendation hold up in your worst case rather than your typical one.
Neutral starting weights
- Gaming performance
- 30%
- Performance per dollar
- 25%
- VRAM headroom
- 15%
- Ray tracing performance
- 10%
- Power efficiency
- 10%
- Features and connectivity
- 10%
Your answers adjust these, then they are normalised back to 100%.
Sources, and what happens when they disagree
Which source wins depends on what is being claimed.
Sources carry a tier. On specifications the manufacturer’s own page wins: it is the authority on what it built. On measurements an independent outlet that publishes its methodology wins, because a vendor has an interest in the result. Specification aggregators are used to fill gaps, never to overrule either.
When two credible sources disagree, CompareYourGPU records both rather than quietly picking one, and the GPU page shows the disagreement along with which figure the site uses and why. There is currently 1 recorded discrepancy in the database.
Measurements are only accepted from outlets that publish a test methodology. The approved list is: TechPowerUp, Tom's Hardware, Hardware Unboxed, Gamers Nexus, ComputerBase, TechSpot, Digital Foundry, Igor's Lab, PC Games Hardware. A benchmark whose publisher is not on that list, or which has no specific review URL and publication date, fails the data-integrity check and breaks the build. That is enforced in code, not by convention.
Where a source qualifies a figure (“approximately”, “up to”), the qualification is stored with it and the number is not presented as exact.
How benchmarks are handled
The strictest rule on the site.
The measured benchmark database is entirely separate from the performance index. A run may only be stored if it comes from a named third party with a link, and it records the resolution, graphics preset, whether ray tracing was on, which upscaling mode was used, whether frame generation was enabled, the average frame rate, the 1% low where published, the test system, and the publication date.
Runs are grouped by test conditions before display, so a table never mixes two methodologies. A ray-traced result and a rasterised result never appear in the same column, and a result with upscaling enabled is never shown alongside a native one without both being labelled.
The database currently holds 0 runs. Where it is empty, the interface says so explicitly and points at the estimated index instead. CompareYourGPU will not fill a benchmark table with plausible numbers to make a page look complete.
How prices are handled
Two fields, and only one of them is currently populated.
A price cannot exist in the database as a bare number. It is stored as an observation: an amount, a currency, a type, a named source, a link where one exists, and the date it was seen. There is no shape that allows an MSRP to be recorded as today’s price, which is what makes the two impossible to confuse.
CompareYourGPU does not scrape retailers, so right now 0 of 35 cards have a verified current price and 35 have launch MSRP only. Every price on the site is labelled with which of the two it is, and value figures state the basis they were calculated from.
A card with no price at all is excluded from value ranking rather than scored zero. Scoring it zero would punish the product for a gap in CompareYourGPU’s data. The same rule applies throughout: any criterion that cannot be scored is dropped and its weight is shared across the rest.
A licensed price feed is built and waiting for credentials. The Amazon Product Advertising adapter handles request signing and response mapping, and reports itself unconfigured when no keys are present, so the site degrades to “Price unavailable” rather than failing. It will not return a figure without a currency, an amount and an observation date, and it never treats an affiliate link as evidence of a price.
Verified prices carry the date they were observed and age visibly. Within 7 days it is presented plainly; up to 30 days it is recent; up to 90 days the age is emphasised; beyond that the interface warns the figure may be outdated. The same four bands apply to benchmark transcriptions and specification checks, so “how old is this number” is answered the same way everywhere.
This matters because street prices frequently diverge from MSRP, sometimes by a great deal. A value ranking based on MSRP is a fair comparison of what manufacturers intended to charge, not a claim about what you will pay today, and the site says so wherever such a ranking appears.
How the upgrade advisor decides
A newer card winning every row of a table does not make it worth buying.
The advisor compares the two cards on the estimated index at your resolution, cross-checks against third-party suite ratings where both cards were tested, and divides the price of the new card by the percentage gained.
- Below 25% the upgrade is called not worth doing, because most people will not perceive the difference in normal play.
- Above 50% the difference is treated as obvious rather than measurable.
- Under 12 per percentage point is good value; over 25 is poor, however large the gain.
- If another card delivers at least 85% of the same improvement for at least 20% less money, the advisor recommends that instead.
- A card that clears a VRAM shortfall at your resolution can be worth buying on a smaller gain, because running short of memory shows up as stutter rather than as a lower average.
- The cost of the upgrade is the price of the new card. CompareYourGPU does not model what your current card might sell for, because it cannot know it.
What is editorial judgement
Some things on this site are opinions. They are marked as such here.
- Each card's summary, strengths, weaknesses, and who it suits or does not suit. These are written by CompareYourGPU, grounded in the specifications on the page.
- The VRAM guidance figures: what counts as a practical floor and what counts as comfortable at each resolution.
- Each game's demand classification, based on engine behaviour and published requirements rather than measurement.
- The weights each best-of category applies. The rule is printed at the top of every category page in the same words as the calculation.
Independence
Retailer links on this site will become affiliate links, which means CompareYourGPU may earn a commission when someone buys through one. This is disclosed wherever such a link appears and on the affiliate disclosure page.
Affiliate relationships are not an input to any calculation on this site. There is no field on a graphics card record that a commission rate could influence, and every ranking is produced by the rules on this page from the specification and price data alone.