How our FPS estimates work

The short version

GamerSpecs estimates FPS with a benchmark-anchored performance model, not live measurements. Every GPU and CPU in our database carries a relative performance score normalized to two anchors: the GeForce RTX 4060 = 100 for graphics cards and the Intel Core i5-13400 = 100 for processors. Every game carries a demand profile and the settings it is calibrated at. An estimate is the deterministic combination of the two — the same inputs always produce the same number, never a random one.

It is one model. Every page and every calculator calls the same function, so the same game, graphics card, resolution and settings give the same figure everywhere on the site, and the settings a figure is at are printed beside it. An automated test checks every curated game on every card, at every resolution and every preset it offers, before anything is published.

What goes into an estimate

• GPU score. Relative gaming performance for 252 desktop, laptop and integrated GPUs, built from published benchmark results. Which cards were measured and which were modelled is explained below.
• Ray-tracing score. A second score on the same scale for cards with ray-tracing hardware, used for games that cannot run without it.
• CPU score. Gaming-weighted (single-thread-heavy) scores for 260+ processors. It changes a figure in one way only: placing your processor between the processors a published test measured, for the games with a measured CPU limit.
• Game demand profile. Anchored to each game's official recommended GPU, taken as ~60 FPS at 1080p (matching it ≈ 60 FPS, double the power ≈ 120) — or better, where the publisher states a higher target: DOOM: The Dark Ages' recommended spec promises 1440p at 60 FPS, which means more than 60 at 1080p. A stated target is treated as a floor, not an estimate: it can raise a figure but never cut one, so a publisher that promises only 30 FPS on its recommended card does not halve our figure. A target that only holds with frame generation is never used. Where a game doesn't name a recommended GPU we fall back to its minimum-spec GPU, then its stated VRAM, then genre. For a handful of heavily played games where that rule measurably missed, demand is set from a published benchmark table instead, and the source is recorded with it.
• Measured figures, where a game's benchmark table has them. For the 29 games we hold a published benchmark table for, a card the table measured shows the measured number, and a card it did not measure is estimated from the table; explained below.
• Resolution scaling. 1440p and 4K apply per-resolution scaling factors — GPU load grows faster than CPU load as resolution rises.
• Preset and upscaling. Each game is calibrated at one setting, and every other preset or upscaling mode is reached by a measured ratio; explained below.
• Frame-rate caps and CPU limits. Where a game's engine caps the frame rate, or a published test shows frame rates flattening because the processor has become the limit, estimates stop at that cap rather than climbing past it.
• Below the minimum: no figure. A card gets no frame rate, and says why with the source, when it lacks the hardware ray tracing a game requires (can't run), when the publisher excludes its generation (CD PROJEKT RED's minimum for The Witcher 3 Remastered, because NVIDIA ended Game Ready drivers for GTX 10 cards), or when it has less video memory than the minimum requirements state (below minimum). Only stated amounts count ("8 GB VRAM" is checked, a card name alone is not), and only on cards whose video memory we hold from a cited specification; integrated graphics, which share system memory, are not checked. Where the minimum names a card that never had as much memory as it states, in any version, that card's biggest version sets the bar: Dragon's Dogma 2 asks for "GeForce GTX 1660 / Radeon RX 5500 XT with 8GB VRAM", and the 6 GB GTX 1660 meets it; Superliminal names the GTX 750 and asks for 3 GB, and the GTX 750 topped out at 2 GB. A figure in the requirements' Additional Notes counts when it is stated outright; one tied to a condition (Bandai Namco's "a 4GB VRAM GPU ... required for DirectX 12") is shown as a publisher note and never decides a verdict.
• Below-spec warnings, not penalties. Above those minimums, RAM, video memory or a processor below what the publisher lists produce a warning in words beside the figure. They do not change the number: no measurement tells us by how much, and a guessed penalty would be a figure we cannot back.

Presets, upscaling, and the label beside every figure

Every FPS figure is calibrated at one setting per game, its basis, and the label beside the figure names it. The basis is, in order of preference:

• Measured (29 games): the preset of the published benchmark table the game's demand was derived from, under the game's own name for it. Black Myth: Wukong's figures say Cinematic · Quality upscaling because that is what ComputerBase tested.
• The publisher's stated target (31 games): the label reads, for example, High (publisher's target) · native.
• Assumed (1,149 games): where the publisher states no target settings, the figures assume High, the usual target of a recommended spec, and the label says so in words: High (assumed from the recommended spec) · native. Where space is tight it shortens to High (assumed), never to plain High.

Any other preset or upscaling mode is reached by a ratio measured on the same card, so a card's place on our GPU scale never enters it. Where the game's own review measured the step, that ratio is used (14 games' preset ladders and 10 games' upscaling gains, from TechSpot and ComputerBase). Everywhere else the step is the median of the games that measured it, at least two of them, and the label adds (modelled). Today the medians are +30% for Quality upscaling (10 games) and +63% for Performance (8 games), and +59% for Medium against Ultra (11 games). A setting nothing measured is not offered at all — which is why there is no "Custom" preset: no measurement covers a custom mix, so no figure for one could be honest.

The ratio is applied before any cap, so a lower preset can never lift a figure past a game's frame cap or a processor's measured limit. And frame generation is never counted: every figure is rendered frames. Where a publisher's own target counts generated frames — Monster Hunter Wilds' recommended spec promises 1080p at 60 FPS on Medium "with Frame Generation enabled" — the page says so, and that target calibrates nothing.

Games with a benchmark table: measured and fitted figures

For 29 games we hold a published benchmark table, mostly from ComputerBase and TechSpot, and the table is linked wherever its figures appear. For those games:

• A card the table measured shows the measured number, at the table's resolution and settings, printed exactly as the reviewer published it, without the ~ that marks our estimates, and labelled measured where the page has room for a label. It is the same number on every page, and it is printed even outside our 20–300 FPS bounds, because it was seen, not modelled.
• A card the table did not measure is fitted to it: placed between the two measured cards either side of it on our GPU score, so it never shows more than a faster measured card or less than a slower one. Past the fastest or slowest card measured, it scales from that card in proportion to its score. Radeon and Arc cards are placed among measured cards of their own make where the table has two or more, otherwise among GeForce cards with the game's Radeon or Arc adjustment. A fitted figure is an estimate, shown with the usual ~ (and marked fitted on the GPU-versus-GPU pages), and it too is the same on every page.

At other resolutions and settings both start from that figure and take the usual resolution and preset steps. A game's own frame cap still applies; a measured CPU limit overrides a measured figure only in a calculator, for your own processor. On the GPU-versus-GPU pages a game's gap between the two cards is shown only where its table measured both. The accuracy figures below judge the model alone, without this anchoring, so it does not flatter them.

Measured CPU limits

A few games stop scaling with the graphics card well before our GPU model would: the processor becomes the limit. Where a published test measured that limit across many processors — Tom's Hardware's 25-processor test of Gears of War: E-Day, its 14-processor Monster Hunter Wilds table, the plateau the fastest cards hit in ComputerBase's Cyberpunk 2077 test — figures stop at the measured limit.

A page that names only a graphics card prints the limit measured on a named mainstream processor, and says which. The calculators know your processor: an exact match uses its own measurement, anything else is placed between the measured processors by CPU score. A processor slower than any measured is extrapolated and labelled as such; one faster than the fastest measured is held at the fastest one's limit rather than guessed higher.

A limit measured in one kind of scene says so. Gamers Nexus measured Dragon's Dogma 2 in its main city, where almost every system is processor-limited, and our figures for the game stop at that limit, labelled in cities. Away from towns the graphics card is the limit and fast cards run higher: ComputerBase's open-world table measures an RTX 4090 at 147 FPS at 1080p.

The same goes for a game's own frame-rate setting. Dead by Daylight's options stop at 120 FPS and Roblox's at 240, so figures for them stop there too, labelled as the in-game limit with the source cited, rather than promising frame rates no player can select.

One game is limited mainly by the processor and measured by no GPU review: Minecraft: Java Edition. Mojang's recommended spec targets 1080p at 60 FPS on Fancy, so on that card or anything faster the figure prints as 60+, labelled CPU-bound, rather than a precise number nobody measured. Below the recommended card the graphics card does limit it, and the usual estimate is shown.

Where the GPU scores come from: measured or modelled

Each GPU score is one fit to published game benchmarks, on the RTX 4060 = 100 scale. Two kinds of measurement go into it:

• Single-game benchmark tables (24 games). The same published tests, mostly from ComputerBase and TechSpot, that our frame rates are checked against.
• Tom's Hardware's GPU hierarchy, 2 test suites. The 2026 suite (11 games on a Ryzen 7 9800X3D) and the 2022–2024 suite (8 games on a Core i9-12900K), which is the one that still measures the GTX 10, GTX 16 and RTX 20 generations. Each suite is an average across its games, so each counts as much as 4 single-game tables.

Fitted (91 cards). One least-squares fit finds the score of every card that appears in any of those measurements, so a card's score reflects all of them at once rather than any single table. Modelled (115 cards). Cards no measurement covers, mostly laptop GPUs and integrated graphics, keep their place within their generation and take that generation's adjustment from the fit. Placed from PassMark (5 cards). Older cards that old games' requirements name and no measurement covers, such as the Radeon HD 5850 and Intel HD Graphics 3000, are placed from PassMark's G3D Mark next to the cards of their generation that PassMark and our table both score. PassMark is a synthetic test, so treat these as rough. Rated from reviews (8 cards). PassMark cannot rate cards from before DirectX 11 checkably, so the GeForce 9800 GT and 8800 GT (one chip), the Radeon HD 4870 and HD 4850 and the GeForce GTX 260 are rated from 2009 reviews of the Radeon HD 5770 (Tom's Hardware, bit-tech, AnandTech and ComputerBase) that tested them in the same charts, and the GeForce GTS 450 from its September 2010 launch reviews (Tom's Hardware, AnandTech, ComputerBase and Hardware Canucks), which tested it beside the HD 5770: each card's frame rate over the HD 5770's, game by game at 1920×1200 or below, in at least five games per review, with the reviews agreeing within 15%. A card is only ever compared with the HD 5770 in the same chart, never with another card rated this way. Other cards of that era are not rated yet. Laptop and integrated graphics are held to the same rule against desktop cards we measure: so far Intel's UHD Graphics 750 meets it, from ComputerBase's and TechSpot's 2021 reviews, and AMD's Radeon Vega 8, from TechPowerUp's and TechSpot's 2018 reviews of the Ryzen 3 2200G. Placed from Tom's Hardware (3 cards). Cards that only Tom's Hardware's 2026 suite measures, such as the Radeon RX 9060 XT's 16 GB and 8 GB versions and the GeForce RTX 5050, are placed against two neighbouring cards that both Tom's and our fit measure, from their frame rates at 1080p and 1440p. Scored as another card (1). The Radeon RX 580 2048SP has the RX 570's 2048 stream processors, so until a measurement covers it, it takes the RX 570's score, and its page says so. Laptop GeForce RTX 20 to 50 (21 cards) are placed from PassMark's G3D Mark, which averages every laptop each chip was tested in, next to the desktop cards of the same generation that we measure, so their scores are an average across laptops and models with a low power limit run slower. Placed from Notebookcheck (8 cards). Some laptop and integrated GPUs, such as the GeForce RTX 5050 Laptop and Intel's UHD Graphics 620, are placed from Notebookcheck's averages of two 3DMark graphics tests, Fire Strike and Time Spy, over the laptops it tested with each GPU. On the desktop cards we measure, both tests run in proportion to our scores. A card is placed only from averages of five or more results, and where both tests count they must agree within 15%. The laptop cards are set to the level of the laptop GeForce cards placed from PassMark, so all laptop scores sit on one level. The other laptop and integrated scores are modelled, not measured, and assume no particular power limit: laptops ship the same graphics chip at very different wattages, and few reviews test one beside a desktop card.

Before every deploy, an automated check compares the scores with Tom's Hardware's measured results: the median card must be within 6% of them, every card within 20%, and every GPU generation with three or more measured cards within 8%. No two cards may be ordered against a measured gap of more than 10%, and on our GPU-versus-GPU pages the modelled game rows may never favour a different card than the measured verdict. A score changes only when a measurement changes or one is added.

Radeon and Arc in benchmarked games

Some games favour one graphics vendor. For the 27 games with a published benchmark table, we measure how that game treats Radeon and Intel Arc cards compared with GeForce, from the table itself, and adjust their figures by that much (in ComputerBase’s Counter-Strike 2 test, for example, Radeon cards run about 18% below their usual standing). Games without a table get no adjustment: for them, a card's average standing across many games is the best estimate.

Ray tracing: a second score

Ray tracing loads a graphics card differently from ordinary rendering: two cards that are level in regular games can be far apart with ray tracing on. So every card with ray-tracing hardware carries a separate ray-tracing score on the same scale. 30 of those are measured, from ComputerBase's ray-tracing ratings (its 2026 game suite; for cards that suite doesn't test, its GPU ranking's 2025 suite or its 2024 suite, each put on the same scale through the cards both measure), and 2 are placed from Tom's Hardware's ray-tracing results the same way as their regular score, against the RTX 4060 and RTX 5060 Ti. The other 82 are modelled: the card's regular score multiplied by how much of its standing its architecture keeps with ray tracing on, as measured on that architecture's cards. The RTX 20 series, which neither suite tests, is anchored on a 2021 ray-tracing test. Cards with 8 GB of video memory or less lose more to ray tracing than larger cards of the same generation; that extra loss is applied only where it was measured (RTX 40 and RX 7000 series).

The ray-tracing score is used for games that cannot run without hardware ray tracing, such as DOOM: The Dark Ages. The 138 cards without ray-tracing hardware (every GeForce GTX and MX chip, Radeon RX 5000 series and older, and most integrated graphics) have no ray-tracing score at all, and on those games they are shown as not supported rather than given a frame rate.

Measured, here, means the card's standing was measured, not your frame rate. A frame rate is a measurement only where it says measured (games with a benchmark table); every other figure is an estimate from the model above, which is why it carries the estimate status described next.

How accurate is the model?

We test it on cards it has never seen. Every card with a published benchmark goes into one of 5 groups. Each group in turn is left out of the fit completely, the scores are refitted without it, and its cards' results in the 24 single-game benchmark tables are predicted and compared with what the reviewers measured.

93.3% of 387 results came within 20% of the measured frame rate for the 82 held-out cards; the median error was 6.7%. Measured 12 October 2026; the weekly data refresh measures it again.

What that covers: how well the model places a graphics card, in games that have a published benchmark table, at the settings those tables used. What it does not cover: games without such a table, whose demand is worked out from the publisher's requirements and is less certain (see where the model is weakest); settings away from a game's basis; and your own CPU, memory, drivers and cooling.

The three FPS data statuses — and which one you are actually looking at

Every major FPS figure on GamerSpecs carries a status badge saying where it came from. The badge is chosen automatically by the data pipeline, never by hand. The pipeline supports three:

• 🟢 Verified Benchmark — a measured result. Shown as exact values (average and 1% lows) because it was measured, not modeled.
• 🟡 Community Data — calculated from multiple structured user FPS submissions, shown as a range with the number of underlying reports.
• ⚪ GamerSpecs Estimate — generated by the performance model described on this page. Shown as a range, never as a single false-precision figure.

Why estimates are shown as ranges

A model can be honest about its own precision. Instead of a false-precision "156 FPS", the headline estimate displays a range like ~150–165 FPS, built deterministically from the estimate's confidence: high confidence ±5%, medium ±10%, low ±15–20%, with the ends snapped to clean values. PC estimates all carry the medium margin today; the mobile model varies its confidence (below). That margin is how the range is drawn, not a measured error: how close the model comes is measured above. A figure sitting on a frame cap shows the cap with no range, because the game holds that exact number. The same hardware, game, resolution and settings always produce the same range, and the calculators print its low and high ends beside the central figure — the estimate's uncertainty, not in-game lows and highs.

Laptop GPUs and power limits (TGP)

The same laptop GPU ships at very different power limits — an RTX 4060 Laptop spans roughly 35W to 140W, and watts change real frame rates. When you set up My Rig or use the FPS Calculator and PC Game Finder, selecting a laptop GPU reveals a GPU Power / TGP field. The model applies a generation-aware scaling curve (efficiency-optimized RTX 40/50 chips flatten out above their sweet spot; RTX 30 chips keep scaling with watts) rather than a naive linear multiplier. If you pick "Not sure", the estimate uses that GPU's reference configuration and says so next to the result.

Mobile estimates

Phone estimates follow the same rules with mobile-specific honesty: results respect each game's FPS cap (an estimate is never shown above a confirmed cap — hitting the cap is flagged as "game FPS cap detected"), and confidence reflects thermal throttling risk on demanding titles. Mobile performance additionally varies with device temperature, manufacturer performance profiles and game versions — which is why every mobile figure is a range with a confidence indicator.

What the numbers mean — and don't

An estimate answers "will this be playable, and roughly how smooth?" — it is not a lab benchmark of your exact PC. Real-world results vary with drivers, cooling, power limits, background apps, game patches and scene complexity. The displayed range is the model's own margin around its central figure, not a guarantee; how far real results can fall from the model is measured under how accurate is the model?

Requirement verdicts (pass/fail) compare your hardware's scores against each game's official minimum and recommended specs as published by the developer, resolved onto the same benchmark scale.

Where the model is weakest

Five limits are worth knowing before you rely on a figure:

• Cards no review measures. A modelled GPU score keeps the card's place within its generation and takes the generation's adjustment from the fit, so any error in that place carries through. Estimates for laptop GPUs and integrated graphics are rougher than for the 91 cards fitted to measurements.
• Old games. The estimate scales your GPU against the game's recommended card. When that card is from 2004, the scaling runs off the end of anything it was calibrated on, and the number can be wrong by a large multiple in either direction. Anything from roughly the last decade is where this model is on firm ground.
• The catalog model has hard bounds, and it says so rather than hiding them. The estimator behind the Can-I-Run, GPU and upcoming-game pages reports between 20 and 300 FPS. A result below the floor displays as "under 20" — written "<20" where a table cell is too narrow for the word; one above the ceiling displays as "300+", which means the graphics card has stopped being the limiting factor and your CPU, the game's own frame cap or your display will decide what you actually see. Neither is printed as a plain number, because at that point the number would be the bound rather than an estimate. A measurement is the one exception: it is printed as the reviewer published it. The calculators run the same engine, with the same bounds.
• Light games print a band, not a figure. The model never treats a game as lighter than a set floor, so every game at that floor gets the same figure on the same card: Intel Iris Xe integrated graphics read about 85 FPS on hundreds of pages. A light game may well be lighter still, so for these games each card shows the band it clears, "300+", "120+", "60+" or "30+", labelled Light game, and a card too slow to clear 30 shows no estimate. A game's own frame cap and a measured benchmark still print as numbers.
• Settings away from a game's basis are usually modelled. Only 14 games have their own measured preset ladder; elsewhere a step is a median of other games, labelled (modelled). Checked against a table it was not fitted to — DOOM: The Dark Ages at Ultra Nightmare with Quality upscaling, calibrated only to id Software's stated 1440p/60 High target — the model alone lands within 20% on 23 of ComputerBase's 24 cards (DOOM's figures now start from that table instead), but a median cannot know a game whose presets scale unusually.
• Demand is one value per game. Two different games whose published requirements are similar will often return the same estimate on the same card. That is the resolution of the model, not a claim that the two games perform identically.

The full list of what this data set cannot tell you yet is on About & Editorial Standards.

Corrections

If a verdict or estimate looks wrong for hardware you actually own, tell us on the feedback page — real-rig reports are the fastest way the data improves. Changes to published data and to this model are recorded with dates on the changelog, so you can see when what you are reading was last revised.