Search games

How we estimate FPS

Every number on FPSRange has to come from somewhere. A measured figure names its source. An estimate is labelled. Low confidence shows no number.

Updated 16 Nov 2026

What a game page can say

A game page compares the PC saved in this browser with the official minimum and recommended requirements, then shows an expected preset. That verdict does not depend on a frame rate.

Where a published test states one average for a card, resolution, and preset, the page shows that number with the source name, a link, and the test date. Other cards and resolutions can show a wide range labelled Estimate. Low confidence shows no number. A developer’s stated frame rate stays a quote and is never turned into a result.

Where the requirements come from

Draft. Draft — not legally reviewed. This section explains the approach. It is not legal advice and it is not a finished licence statement.

System requirements are the publisher text from the Steam store page (store.steampowered.com, app details), stored with the date we retrieved it. The original wording is on the game page. A field we could not read stays empty.

We do not copy PCGamingWiki, SteamDB, or TechPowerUp databases. Hardware specs used for the comparison index come from manufacturer data we recorded ourselves. The index is a spec-based estimate with low confidence. It is not a benchmark ranking.

Developer targets

Some store pages state a goal, such as 1080p on High. We quote that sentence with the source link and the retrieval date. It is the studio’s target, not a measurement by FPSRange.

A figure that includes frame generation is labelled frame generation. It is not treated as a native frame rate, and it is not used to draw an FPS number.

Spec comparison

Recommended met means every compared part meets the official recommended bar. Minimum met means the compared parts meet the official minimum, and the page is not a full recommended match: a part is still short of recommended, a graphics card could not be checked, or no recommended spec is stated. Below minimum means a required part is under the official minimum.

Graphics, processor, and memory are the parts we compare. Storage and the operating system stay on the page, and the browser cannot see them. A resolution the publisher did not name is marked “no official target · coarse read”. The expected preset there is the same spec comparison, not a measurement at that resolution.

When a requirement leaves a part out, or names a part we cannot match, the other parts are still compared. The missing part is marked as not checked, and the confidence label says the comparison is partial. An unchecked graphics card keeps the page at Minimum met.

AMD cards use the same generation ladder as NVIDIA, by product era: HD 7000 and R9 200 with Kepler, R9 300 with Maxwell, Polaris with Pascal, RDNA 2 with Ampere. Those older AMD names sit 20 points above the class they match, so an RX 580 is the same rung as a GTX 1060 and clears an R9 290X. An R9 390 is the same rung as a GTX 970. An HD 7970 is the same rung as a GTX 680. An RX 6500 XT stays under a GTX 1650. That is a naming heuristic, not a frame rate. If the publisher states no minimum VRAM, we use the smallest VRAM of the named minimum graphics cards. Falling short of recommended VRAM alone is not below the official minimum.

spec-v1 has no generation term, and a card that publishes memory bandwidth is not on the same scale as a card scored from shader rate alone. CUDA cores and stream processors are different units. A requirement check does not use that score across generations or across vendors. It uses the generation and the product tier already in the catalog name: one tier step (60, 70, 80) per generation. That ladder is not a benchmark and not a frame rate.

Confidence levels

Measured means the number is the published average for that card, resolution, and preset. Estimate means the range was scaled from a measured anchor. High or medium confidence can show that range. Low confidence shows no number, only the spec verdict.

A test from before 2023 is still used, with its date on the page. As a basis for an estimate it is down-weighted: the range gets wider, and confidence drops one step. A drop from medium to low removes the number.

User reports will later replace these estimates. A report is not collected on this preview. Until then an estimate stays labelled Estimate and names the anchor it came from.

LabelWhenWhat you see
MeasuredA published average for this card, resolution, and preset, with the source link and test date.That single number. Not a range, and not a rating.
EstimateScaled from a measured anchor by spec-v1, for the same vendor and the same architecture only. High or medium confidence.A wide range labelled Estimate, plus the anchor’s source, link, and date.
No numberNo usable measurement, a different architecture or vendor, no spec-v1 score, a stale basis that drops confidence, or only a threshold sentence.No FPS number. The spec verdict remains.

How an estimate is computed

high: scaled value ±12% · medium: scaled value ±20%

An estimate starts from one measured average. Frame generation is never used. Upscaled figures are not used as a basis for other cards. A GameGPU sentence of the form “at least N frames from card X” is only a lower bound: it is never shown as its own frame rate, and it can only raise the bottom of an estimate that already came from a measurement. If the whole estimate sits under that bound, the number is withheld.

An estimate scales only when both cards are the same vendor and the same architecture, and both have a spec-v1 score: FPS(target) ≈ FPS(anchor) × (spec target ÷ spec anchor). The generation ladder is a naming heuristic for the requirement check. It is not a frame-rate scale. A different architecture, a different vendor, a missing architecture, or no spec-v1 score produces no number.

The estimate is only for the anchor’s resolution and preset. A different resolution or a different preset is not folded into that range. High confidence shows the scaled value ±12%. Medium confidence shows ±20%. Those widths are the whole uncertainty band. They live in the estimate config. If the comparison is too far for that band, including a stale basis that drops medium confidence to low, the page shows no number. The anchor’s card, resolution, and preset are written next to the range, with the source link and the test date.

A figure taken at 2560×1600 and filed next to 1440p is not relabelled as a 1440p average. 2560×1600 has about 11% more pixels, so the result is a lower bound: at least that many frames at 1440p, measured at 2560×1600, with the source and the date. An estimate for another card from that bound is one confidence step lower.

A pair rated below the official minimum does not get an estimate or a 2560×1600 floor. A measurement on that card is still shown. The verdict and the frame rate do not contradict each other.

Nothing here claims the bands have been calibrated against held-out tests. User reports will later replace an estimate for the same game, card, resolution, and preset.

What the browser can see

The browser can see a graphics string, a count of logical processors, a rounded memory figure, a screen size, and sometimes a WebGPU adapter. It cannot see the processor model, video memory, a laptop power limit, free storage, the DirectX version, or the driver. Recognition is a suggestion. Graphics, processor, and memory have to be confirmed before a PC is saved, and that save stays in local storage on this device. Nothing is uploaded.

What your browser tells us

  • Graphics card model — usually, in Chrome and Edge (WebGL)
  • Number of CPU threads (not the model)
  • Rough memory size — capped at 32 GB, 8 GB on phones
  • Screen size (not a refresh rate)

What it can't tell us

  • Processor model
  • Graphics memory (VRAM)
  • Laptop power limit (TGP)
  • Free disk space
  • Windows and DirectX version, drivers

That's on purpose (privacy protection), not a bug. Set these yourself when they change the result.

Benchmark articles

Draft. Draft — not legally reviewed. This section explains the approach. It is not legal advice and it is not a finished licence statement.

The list below is every article considered for this preview, including ones we skipped. A used row contributed one or more single figures, each stored with the source name, the URL, and the test date. Skipped rows were read and not copied. Candidate sites whose terms forbid reuse are not in this list.

This is a preview. A lawyer still has to review EU database right before any launch. We do not copy whole tables or charts.

What we will not do

Draft. Draft — not legally reviewed. This section explains the approach. It is not legal advice and it is not a finished licence statement.

No invented FPS. No number derived from a developer quote. No tracking script. Hardware from the check stays in this browser.

Results are not for sale: a shop cannot buy a better verdict, a higher place in a list, or a different upgrade order. Affiliate links are not on game pages yet.