01 · Normalize parts to a 0–100 index
Each CPU and GPU is projected onto a 0–100 performance index per workload (gaming, productivity, streaming, AI) and per resolution (1080p, 1440p, 4K).
The index is a weighted blend of measured benchmarks: PassMark single- and multi-thread for CPUs; FP32 TFLOPS, VRAM size, and memory bandwidth for GPUs. Weights shift by workload — gaming favors single-thread; productivity favors cores and multi-thread; AI favors VRAM headroom.
02 · Compute a balance ratio
Once both parts are normalized, the engine computes the ratio of CPU index to GPU index and compares it against a target ratio for the chosen resolution and workload. 4K gaming has a GPU-heavier target; 1080p high-FPS gaming has a CPU-heavier target.
Pairings within a ±15% tolerance band are classified as balanced. Heuristic precision below that magnitude would be false confidence — we don't report it.
03 · Predict FPS with a three-tier fallback
For top games, the engine looks up the closest matching CPU + GPU + resolution + preset benchmark in our games-FPS database. The result carries a confidence indicator:
- HIGHExact match in measured benchmark data.
- MEDIUMSame parts, different resolution or preset — interpolated.
- LOWNo matching benchmark — estimated from performance indices alone.
Every FPS prediction in the result panel is tagged with its confidence so you can tell measured truth from estimated estimate.
04 · Recommend the smallest upgrade that helps
When a build is bottlenecked, the engine scans every part of the same kind in the catalog and ranks them by FPS-gain-per-dollar. We surface the single best swap, not a wall of options — most users want one decision, not a comparison spreadsheet.
Measured accuracy
We validate the engine against a held-out set of measured FPS benchmarks every commit. As of engine v0.3 (May 2026), running on the “estimate” path alone (i.e., predictions where we have no matching benchmark in the database), the engine achieves:
- MAE21.6% of mean measured FPS (44 FPS off a 205 FPS average)
- WITHIN ±15%36% of predictions
- WITHIN ±25%56% of predictions
- WITHIN ±35%72% of predictions
- WITHIN ±50%92% of predictions
These are the floors we promise. When the engine has a HIGH confidence match (the exact CPU + GPU + game + resolution + preset is in our benchmark database), accuracy is essentially 100% — we’re returning the source. The numbers above describe the worst case: predictions where the engine has to extrapolate from indices alone.
Validation runs on every CI build. The full test is at tests/lib/engine/accuracy.test.ts. If we ever ship a regression, the build fails.
Why this beats the competition
Most bottleneck calculators on the internet output one opaque percentage and nothing else. We output a verdict, FPS predictions per game, a performance grade, a PSU sizing, a confidence score, and a concrete upgrade — all anchored on real measurements you can verify on the data sources page.
We also publish our measured accuracy (above). Most competitors either claim “90% accuracy” with no proof or refuse to commit to a number at all.
Engine version: v0.3 (May 2026) — per-game CPU sensitivity weights, PSU transient-spike modeling, validation harness. The methodology is open: every coefficient, threshold, and the validation set lives in the project repository.