5%
11.1 FPS on average across 395 predictions
LOADING_
MODEL_VALIDATION / 2026.08
Internal cross-validation against 69 builds with frame rates published by review outlets: the engine is off by 5% on average in average FPS. Here is what that number measures, what it does not, and why it is probably optimistic.
5%
11.1 FPS on average across 395 predictions
3.6%
Half of all predictions fall below this
83%
98.7% within ±20%
-0.9%
Sign of the mean error: below zero the model underestimates
Five-fold cross-validation. The 69 reference builds are split into five groups; for each group the engine is refitted on the other four only, then predicts the group left out. None of the 395 predictions comes from a refit that saw its own build — although the starting values of that refit were fitted on all of them (see the limits).
The split is by whole build, not by game. Separating Cyberpunk from Counter-Strike within the same machine would let that hardware train its own evaluation and would flatter the number.
Reproducible: the split is deterministic and the calculation runs with npm run accuracy in the project repository.
It is an internal validation, and probably optimistic. Each game's starting values were fitted on the complete reference set, so information from the held-out group reaches its own evaluation through them. It is not an unbiased estimate of the error on new builds, and not a proven bound either: nothing here bounds it mathematically.
A sensitivity check repeats the calculation without the pull toward those starting values: the error moves from 5% to 5.4%. That difference measures one leakage path, not the whole bias, because the starting values still come from all the data.
Nor is the gap to the in-sample error (4.8%) the size of the overfitting: both figures share the same bias.
It measures average FPS only. It does not validate the bottleneck percentage, 1% lows, frame pacing, the buying advice, or games outside the eleven profiled ones.
Conditions of the references: native resolution, AAA titles at high/ultra with ray tracing off, esports and competitive titles at their competitive-low preset, no upscaling and no frame generation. The figure says nothing about error with DLSS, FSR or frame generation.
Provenance: the 69 builds were compiled from published reviews (TechPowerUp, TechSpot / Hardware Unboxed, Tom's Hardware, Gamers Nexus, Digital Foundry; 2023–2026). The source of each individual observation — article, date, driver, game patch — is not yet recorded, so an outsider cannot audit every number. None was measured on a test bench of our own.
69 builds is a small sample. Games with fewer than 15 observations (the N column) should be read with caution; no confidence intervals are published because none were computed.
| GAME | N | ERROR | ±20% |
|---|---|---|---|
| League of Legends | 10 | 7.8% | 100% |
| Counter-Strike 2 | 36 | 7.5% | 97.2% |
| Valorant | 22 | 6.7% | 100% |
| Call of Duty: Warzone | 12 | 5.7% | 91.7% |
| Apex Legends | 15 | 5.7% | 93.3% |
| Hogwarts Legacy | 60 | 5.3% | 98.3% |
| Red Dead Redemption 2 | 61 | 5.1% | 100% |
| Alan Wake 2 | 56 | 5.1% | 100% |
| Fortnite (Perf Mode) | 9 | 4.3% | 100% |
| Cyberpunk 2077 | 61 | 4.3% | 100% |
| Elden Ring | 53 | 2.2% | 98.1% |
| RESOLUTION | N | ERROR | ±20% |
|---|---|---|---|
| 5K2K Ultrawide | 5 | 8.4% | 100% |
| 8K | 4 | 7.3% | 100% |
| 720p | 6 | 5.9% | 100% |
| Super Ultrawide 32:9 | 4 | 5.7% | 100% |
| 1080p | 152 | 5.6% | 97.4% |
| 4K | 105 | 4.7% | 99% |
| 1440p | 100 | 4.5% | 100% |
| 5K | 4 | 4% | 100% |
| Ultrawide 1600p | 5 | 3.2% | 100% |
| Ultrawide 1080p | 5 | 2.6% | 100% |
| Ultrawide 1440p | 5 | 2.5% | 100% |
Deterministic tests for bounds, resolutions, RAM, and FPS.
[ACTIVE]Whole-catalogue sweep of the numbers and of the prose of all 7,566 comparisons in three languages.
[ACTIVE]Five-fold cross-validation against published frame rates.
[PUBLISHED]Builds frozen before any fitting, with a source recorded for every observation.
[PENDING]Before the next refit of the coefficients, a set of builds will be frozen with a source for every observation (article, date, driver, preset, resolution, ray tracing, upscaling and memory) and never used for fitting. Its error will be published here separately from the internal validation. Until then, this page reports internal validation only.
A bottleneck percentage describes estimated system balance, not a universal FPS loss, and this FPS error does not carry over to it. If your measurements disagree with a page, send us the build and what you measured: confirmed errors are fixed and recorded in the changelog.
Measured on 2026-09-11 · this page last revised: 2026-09-11. Read methodology · Read changelog · Report an error