Most bottleneck calculators return a single percentage produced by an undisclosed formula. That number is not measurable, cannot be checked against reality, and does not tell you what to do. This page explains exactly how our estimates are produced so you can judge them yourself.
The core idea
A frame has to pass through both the processor and the graphics card. Each can sustain some maximum rate on its own. The machine runs at the slower of the two. So instead of inventing a bottleneck percentage, we estimate two ceilings separately and compare them.
The component with the lower ceiling is the one limiting you. The gap between the two ceilings is the performance you have paid for and are not using.
Why resolution changes the answer
Rendering more pixels gives the graphics card more work. It does not give the processor more work. This is why the same pair of parts can be processor-limited at 1080p and graphics-limited at 4K, and it is the single biggest thing other calculators get wrong. Our processor ceiling barely moves with resolution; our graphics ceiling scales with pixel count, very slightly sublinearly because some per-frame work is fixed.
Where the numbers come from
We do not republish third-party benchmark scores. Every hardware figure in our database is a published manufacturer specification: shader counts, clock speeds, memory bandwidth, core counts, cache sizes, board power.
From those specifications we compute throughput. On the graphics side that is shader count multiplied by clock, multiplied by an architectural work-per-shader coefficient, then adjusted for two real effects: memory bandwidth starvation on narrow-bus cards, and the sublinear scaling of very wide dies whose front end cannot keep the shader array fed.
On the processor side it is gaming instructions per clock multiplied by boost frequency, with a saturating bonus for the cores a game engine can actually use, and a term for L3 cache size. That cache term is why 3D V-Cache processors rank where they do here without needing a special case written for them.
Calibration
Architectural coefficients are not guesses. They are derived by anchoring to well-established equivalences between real products — for example, that an RTX 4070, an RTX 3080, an RX 6800 XT and an RX 7800 XT all land within a few percent of each other, and that a GTX 1080 Ti performs like an RTX 3060 Ti. Solving for the coefficients that reproduce those relationships gives us values that then predict the rest of the range.
What this cannot tell you
These are model estimates, not measurements. They will not capture memory speed and timings, driver version differences, thermal or power limits in a specific case, background load, a particular game patch, or the difference between an open field and a dense city centre in the same title. Treat the output as a well-reasoned expectation, not a promise.
If a result looks wrong to you and you have measured data that says otherwise, we want to hear about it — that is how the calibration improves.
Changes to the model
A model that nobody can see changing is a model you have to take on faith. Every substantive correction since launch is listed here, including the ones that were our own mistakes, because a changelog that only records improvements is marketing rather than a record.
- 8 October 2026 — the site was cut from 515 indexed pages to 186
- Not a model change, but the most important entry here. Google removed most of this site from its index in late September, and it was right to. 515 pages had been published of which 34 had ever earned a click. The per-pairing pages and the generation comparison pages have been deleted outright; part pages with no measured search demand still render and are still linked, but no longer ask to be indexed. What remains is the calculator and its tools, the written guides, and the part pages people actually look for. PRUNE.md in the repository records exactly what went and on what rule.
- 3 October 2026 — resolution scaling corrected against measured results
- The first coefficient on this site to be checked against measured benchmarks rather than against product equivalences, and it was wrong. Frame cost was assumed to scale with pixel count to the power 0.92, making 4K 3.6× the cost of 1080p. An 11-game geomean across seven cards puts the real ratio at about 2.7×. The exponent is now 0.72. Figures at 1440p rise by roughly 12% and at 4K by roughly 30%; 1080p is unchanged, because the correction is entirely in how cost grows with pixels. One source — review sites publish these figures as chart images and automated reads of them proved unreliable enough to be unusable, so this replaces an unchecked assumption with a measured figure rather than settling the question.
- 3 October 2026 — known and unresolved: the ultra preset
- With resolution scaling corrected, the model remains about 30% pessimistic across the board against that same data. Two candidates, which this data cannot separate: the ultra preset multiplier of 1.38 is too harsh, or the measured suite (which includes four esports titles) is lighter than the demanding AAA baseline the model uses. Recorded here rather than fixed, because guessing between two explanations is how the resolution coefficient came to be wrong in the first place. The site's published figures use the high preset, not ultra.
- 3 October 2026 — video memory requirements corrected
- The memory requirement scaled with resolution far too steeply, putting a typical AAA title at 14.1 GB at 4K on the high preset when the real figure is nearer 9 to 10. That marked almost every card under 12 GB as short of memory at 4K, and a shortfall penalty then compounded on top without any floor. The effect was severe: an RTX 3080 was being quoted at 16 fps at 4K and an RTX 2080 Ti at 12. The coefficient is now 0.20 and the penalty is floored at half, which is closer to what happens in practice — a card short on memory stutters badly, it does not stop rendering. The 1080p and 1440p figures were not affected.
- 23 September 2026 — generation comparison pages added
- Each graphics and processor generation now has a page comparing every part in it against the same partner component, which is the one comparison an individual part page cannot make.
- 17 September 2026 — pairing selection widened
- The rule that chose which pairings earn a page excluded every budget graphics card, because it used an absolute threshold that a weak card cannot meet by definition — a weak card is the limiting component against every processor ever made. It now measures against what that card can actually achieve, and recommends the cheapest processor that gets there rather than the fastest.
- 5 September 2026 — launch
- Model calibrated against known product equivalences, 126 graphics cards and 201 processors from published specifications.
The largest outstanding weakness is that none of this has been validated against measured benchmark results. The model is anchored to product equivalences, which constrains it but does not verify it, and the 4K figures sit at the end of the longest extrapolation and are the least reliable numbers on the site.