Accuracy and limitations

For popular modern titles with at least 200 community FPS reports in our calibration set, estimates are within ~15% of observed performance at 1080p on a mid-range preset. The wider the 95% confidence interval shown on the result card, the less data we had for that game — high-confidence results (~5% interval) are common; low-confidence results (~30% interval) are surfaced for less-tested games.

We compare every estimate against a corpus of community-reported FPS, then fit a per-game residual model. The model is re-fit weekly as more reports come in. The calibration method is:

1. Bootstrap the calibration corpus (10,000 iterations, 95% CI).
2. Fit a per-game residual (observed − predicted) on the calibration set.
3. Re-derive the FPS estimate for each game using the residual as a prior.
4. Cross-validate on a held-out 20% of the corpus.

The full math is in the methodology.

Five known failure modes, in order of how often they bite:

• Heavily CPU-single-thread-bound games (some strategy and simulation titles) are overestimated on rigs with weak single-thread performance. The estimator weights multi-core throughput more than it should.
• Poorly-optimized console ports are underestimated — the model assumes reasonable scaling, and ports often scale worse than native PC builds.
• New drivers (within 14 days of a major GPU driver release) can shift estimates by 10-20%. We do not currently ingest driver-version data into the estimator.
• Console-to-PC ports sometimes do not reflect actual PC scaling. Estimates for these games carry a wider confidence interval.
• Modded or unofficial builds are out of scope. Estimates are for the stock, unmodified game as sold on Steam.

The estimator is for frame rate, not frame quality. We do not currently measure:

• Frame-time variance — a 60 fps game with 16 ms frames feels different from a 60 fps game with 8-24 ms frames. We return the average.
• 1% lows — the worst 1% of frames over a play session. Critical for competitive play; not currently reported.
• Stutter analysis — shader compilation hitches, asset loading, and streaming artifacts. Not in the model.
• VRAM headroom — the estimator checks VRAM at the configured texture preset but does not model texture-streaming behavior under load.
• Network latency — online play performance depends on the connection, not the rig. We do not model it.
• Ray tracing / DLSS / FSR impact— these are exposed in the per-game advanced options but the impact of each upscaler on a given rig is outside the estimator's scope.

Every game page has a "Report FPS" button. Submitting a report feeds the next calibration run. Include your CPU, GPU, RAM, the resolution and preset, and the observed frame rate — that's all we need.

For a methodological issue, email [email protected] with the subject line "[Accuracy]".