We use cookies for analytics and for the ads that fund this site. Limited cookieless measurement may run before you choose; full analytics and personalized ads only with your consent. Read the Cookie Policy or the Privacy Policy.
VERSION_LEDGER
CHANGELOG
Every change to the model or the text, with its version and date. Corrections say what was wrong, not only what is new.
One upgrade decision per page, and corrections to published claims
Correction: some comparison pages recommended "a stronger graphics card" in the analysis and said "no card swap pays off" in the buying block. Both now read one decision. A part is only recommended if it adds at least 25% more frames (graphics card) or 10% (processor) in the same pairing; otherwise the page says no part in the catalogue clears that bar and names the closest one with its gain.
Correction: the "GPU Interface" line showed the memory bus (for example 256-bit) and every page claimed PCIe Gen 4.0/5.0 support for the processor. The memory bus and bandwidth are now labelled as video memory, and the PCIe link is marked as not recorded, because the catalogue does not hold it.
Correction: the memory advice said 16 GB covered the recommended 4K resolution and, in the same sentence, to move to 32 GB for 4K. It now states the model's own capacity target at the recommended resolution and at 4K. Tier words such as "mid-range" were removed from the analysis — they were read off the model's 0–100 scale and called an RTX 5080 mid-range — and the intro no longer attributes that score to PassMark or G3DMark.
Correction: an over-specified graphics card was described as performance you could get "for free" at a higher resolution, and the 4K question was answered "yes" from balance rather than frame rate. More pixels cost frames, and the copy now says so; the 4K answer uses the same 60 and 30 FPS bars as the recommended resolution.
Correction: the accuracy figure is now described as internal cross-validation of average frame rates, probably optimistic. The claims that the 0.2-point in-sample gap was "the size of the overfitting" and that the figure was a "lower bound" were removed, and the sensitivity figure is read from the report (5.4%) instead of a stale 5.6% typed into the page.
Correction: sections 1 to 4 of the methodology described formulas from an earlier engine (fixed PassMark and G3DMark baselines, a cache bonus, VRAM bandwidth choking, a desync penalty) that the current engine does not run. They now describe the gaming index, the two ceilings per game and how the verdict and buying advice are decided. The About page no longer carries an unverifiable job title.
2026.08 · 2026-09-02
Upscaling, frame generation, refresh targets and a rebuilt pairing rank
DLSS, FSR and XeSS are modeled as a lower internal render resolution, so they raise the GPU ceiling and leave the CPU ceiling untouched. A CPU-limited build now correctly gains nothing from them, and the calculator says so instead of showing a larger number.
Frame generation multiplies the delivered rate by a measured factor rather than its nominal 2x, 3x or 4x, and always discloses that latency does not improve. It is excluded from the target analysis entirely, because a purchase decision made on generated frames would be wrong.
A target frame rate can now be set. The result reports how many titles reach it and, for those that do not, which of the two ceilings is the one below — including titles that no hardware can take there because the game caps itself.
The best-CPU-for and best-GPU-for rankings no longer sort by system efficiency. That measured how well two parts matched rather than what they delivered, and it ranked a Core i5-14400 above a Ryzen 7 9800X3D for an RTX 5090 — giving up 31% of the frames. Parts are now ranked by delivered frames, and each page states which processor is the entry point, which keeps headroom, and how much of the card each one unlocks.
Frames above 240 Hz no longer count when judging whether a processor fully feeds a graphics card. Without that cap the average was decided by esports titles running at 400-700 FPS, where the difference is invisible, and no processor in the catalogue counted as sufficient for an RX 7900 XTX.
New head-to-head pages compare processors against processors and graphics cards against graphics cards, reporting the resolution past which the difference stops being noticeable.
Sharing a calculation now produces a card showing the actual verdict, computed on the server so it cannot disagree with the page it links to.
2026.08-ceilings · 2026-08-12
Ceilings model and presentation consistency
Each game is now modeled as an independent CPU ceiling and GPU ceiling, so the headline verdict and the per-game FPS chart are derived from the same two numbers.
Component power now reads a measured gaming index instead of a synthetic score, so gaming order is respected rather than PassMark or G3DMark order.
Memory is modeled on two axes: capacity decides whether a game fits in RAM, while transfer rate and channel count decide how fast the CPU can feed the GPU.
Comparison pages no longer re-derive the verdict from a synergy threshold. The label, the tooltip and the explanation now always agree, and the recommended PSU is quoted once from the engine.
Removed the unnamed engineer byline from generated pages. Those pages are produced by the model, and now say so.
2026.07-baseline · 2026-07-23
Validation and delivery baseline
Expanded the physical resolution and RAM catalogs.
Added deterministic engine regression coverage.
Documented native raster assumptions and external-validation limits.
Separated measured, modeled, and pending accuracy claims.