How this site works — model v4, our own index, and a verdict we don't publish
This page documents model v4 (November 2026). Earlier versions of this site converted published benchmark hierarchies into frame-rate ceilings. Version 4 replaces that entirely: every index here is our own, derived from manufacturer specifications only — no licensed benchmark tables, no scraped frame-rate figures. We also built a bottleneck verdict model on top of the index. Before publishing a single verdict we froze its method and ran a pre-registered validation. It failed. So this site publishes the index, the rankings, and the failure — not the verdicts. Everything below is the full account.
The index
CPU gaming index. Each processor is scored from its manufacturer-published specifications: per-core performance (boost clock × an architecture-generation IPC factor), a multi-core contribution that scales with the logarithm of core count (games use a few cores hard and the rest lightly), and a bonus for very large L3 cache (the class of design AMD sells as 3D V-Cache). Scores are normalised so the Ryzen 7 9800X3D = 100.
GPU index, per resolution. Each graphics card is scored at 1080p, 1440p and 4K separately: shader throughput (shading units × boost clock) blended with memory bandwidth — the bandwidth share rises with resolution — adjusted by a VRAM-adequacy factor for the resolution, and normalised so the RTX 5090 = 100 at each resolution. Inputs are the card makers' published specifications: core counts, clocks, memory bus and speed, capacity, board power.
How soft are these numbers? We perturbed every constant in both formulas by ±10–20% (102 perturbation sets) and re-scored all 2,464 pairings at all three resolutions. The GPU index moved a median of 4.6 points; the CPU index moved a median of 31.2 points — the CPU index is the softer number, and small CPU differences (a few points) should be read as ties. Within each GPU family, the ordering of adjacent cards agreed with the manufacturer's own product positioning in 100% of checks after three documented exception rules. The full constant list, marked claim-derived vs. our judgment, is kept with the index's working files.
What the comparison means
The calculator and the pairing pages divide one index by the other. A ratio of 1.79 at 1440p means exactly this: the CPU's gaming index is 1.79× the GPU's 1440p index. It is a comparison of two spec-derived estimates — not a frame rate, not a bottleneck percentage, and not a measurement of your PC. Which part actually limits a given game depends on the game, the scene, your settings, your driver and your RAM; no index can see those. What the index can do honestly is rank parts and flag pairings where one side is far out of the other's league.
The verdict model — and its failed validation
The verdict layer we built works from a different kind of evidence: when a game publisher publishes system requirements, each tier asserts a CPU and a GPU that are matched for a stated resolution, preset and frame-rate target. We transcribed every such tier from publishers' official charts, matched the named parts to this database exactly, and froze the corpus before scoring anything: 89 pairings across 13 games (the corpus target was 80+ pairings across 25+ games; the pairing count was met, the game count was not — most publishers' charts omit per-tier resolution, preset or frame-rate targets, and incomplete charts were excluded rather than completed by guesswork). A robust (Theil–Sen) balance curve, Cbal(G), was fitted per resolution; a pairing's distance from the curve, in the fitted units, would have been its verdict.
The validation criteria were written down before the fit was scored. Run blind against the frozen corpus on 2 November 2026:
| Criterion | Required | Result |
|---|---|---|
| Flagship anchor: 9800X3D + RTX 5090 balanced at all three resolutions | Balanced ×3 | FAIL — mild GPU-side at 1080p, balanced at 1440p, mild CPU-side at 4K |
| Direction on 16 known pairings | ≥90% | FAIL — 12 of 16 (75%) |
| Inverted verdicts (wrong direction entirely) | 0 | PASS — 0 |
| Distribution over all 7,392 pair-resolutions | No class >70%, balanced ≥15% | PASS — 33.4% balanced, 57.3% GPU-side, 9.3% CPU-side |
| Stability under bootstrap resampling | Anchors stay balanced | FAIL — follows from the anchor result |
| Corpus back-test: publisher pairings balanced under the fitted curve | ≥80% | FAIL — 79.8% (71 of 89) |
Verdict on the verdict model: not shipped. The diagnosis is in the data: the per-resolution corpora don't span the GPU range — 1080p charts pair modest cards, 4K charts pair flagships — so each resolution's curve is well supported only locally and extrapolates badly to the opposite end. A single pooled curve passes the flagship anchor but washes out the resolution structure (its back-test falls to 67%). What could change this: a corpus that spans the full GPU range at each resolution (high-frame-rate 1080p charts are the biggest gap) and a partial-pooling fit instead of the pooled-or-split switch — validated from scratch, as a new model version, against criteria fixed in advance again. Until a version passes, this site shows indices and rankings only.
Update — model v2 (November 2026). We ran that second version: the corpus grew to 157 pairings across 31 games and the fit changed to partial pooling (each resolution's curve shrunk toward the pooled curve). The flagship anchor now passes at 1440p and 4K but still fails at 1080p; the direction score improved to 13 of 16; and the back-test fell to 61.8%. The per-game results show why: under one curve, some publishers' charts score 100% and others 0% — publishers do not spec their tiers to a shared balance point, so no single curve at a ±15% tolerance describes all of them. Verdicts remain unpublished. A third version would need to model per-publisher differences explicitly, against criteria fixed in advance again.
The two upgrade answers
Every result separates two questions other calculators blur into one. The sensible step is the weakest part that closes at least half the index gap between your two parts at your resolution. The full index match is the weakest part whose index meets or exceeds the other part's — the mathematical answer, labelled a paper target because it is often a flagship part and usually overkill. CPU steps are searched on your current socket first. Every recommendation is verified by re-checking the index of the recommended part before it is printed, and the share of the gap it closes is printed next to it. When no part in the database fully matches, we say so and name the strongest option. Both steps are chosen on performance only: we hold no price data, so we never claim a recommendation is the lowest-priced way to buy — only the smallest step that does what it claims.
The PSU checker
The wattage checker adds the GPU's rated board power, the CPU's peak boost power (Intel maximum turbo power / AMD PPT — not the smaller base TDP on the box), and 100W for the rest of the system, then returns the smallest standard PSU size that keeps at least ~25% reserve over that estimated load.
Sources
Index inputs: manufacturer specification pages — AMD, Intel and NVIDIA product and ark/datasheet pages for the parts in the database (accessed October–November 2026). Validation corpus: publishers' official system-requirements charts for Cyberpunk 2077, Kingdom Come: Deliverance II, Battlefield 6, Assassin's Creed Shadows, Indiana Jones and the Great Circle, DOOM: The Dark Ages, Gears of War: E-Day, Dune: Awakening, Clair Obscur: Expedition 33, Borderlands 4, S.T.A.L.K.E.R. 2, God of War Ragnarök and Metal Gear Solid Delta: Snake Eater. We are not affiliated with any of them, and no figure on this site is copied from a review site's benchmark tables.
What this cannot tell you
It cannot predict your frame rate in a specific game, account for your RAM kit, cooling, or overclock, or replace measuring your own system (tools like CapFrameX and PresentMon can show you which part is really limiting your frame times). Upscaled rendering (DLSS/FSR) and frame generation are different regimes these native-resolution indices do not claim to cover. If an index comparison surprises you, trust measurements of your own PC over any calculator — including this one.