Comparison methodology

Specifications

Specifications come from manufacturer pages and reference sources. Every record has a source link, a check date and a verification status. If the manufacturer does not publish a value, we show “no data”: an empty value never turns into zero. If sources disagree, the model page shows both values.

For graphics cards, the reference chip (for example, “GeForce RTX 5060”) is stored separately from partner cards: the length, clocks and connectors of a specific card model may differ.

Verification statuses

Official data: a value from the manufacturer’s page.
Independent measurement: a measurement by a source with an open methodology.
Community data, filtered: user measurements that passed manual moderation or statistical filtering (the median of many submissions without outliers). Contributors’ test systems are not controlled.
Community data, small sample: reference measurements with no conclusion about which is better.
Estimate: not a measurement; not used in verdicts.
Pending review / rejected: not shown publicly.

When benchmarks are comparable

Two results are compared only if all of these match:

the benchmark name, its version and subtest, the unit and the direction (“higher is better” or “lower is better”);
for games: the game version, resolution, preset, API, ray tracing, upscaling and frame generation (all must be known);
for CPU gaming tests: the graphics card in the test system.

Other results are shown for reference only and are not used in verdicts.

Percentage difference

difference = |A − B| ÷ the smaller value × 100%. A difference below the significance threshold (by default 4%; some benchmarks have their own threshold) counts as parity.

Rankings

A ranking is built separately for each benchmark at one version, under the same conditions and on one render backend, from at least three released models. “Rank N of M” means the rank among the M models tested in that benchmark, not among the whole catalog. The “% of leader” column is the model’s result divided by the leader’s result (the other way round for “lower is better” benchmarks).

Blender Open Data

Cycles results come from Blender Open Data (CC0). For Blender 3.x–5.x we use the sum of samples per minute across the monster, junkshop and classroom scenes. Graphics cards render with their vendor’s standard backend: NVIDIA uses OptiX, AMD uses HIP, Intel uses oneAPI and Apple uses Metal (CUDA is a separate scale). Results from different backends depend on the software stack as well as the card, so we do not calculate a percentage advantage between them: comparisons show such values side by side for reference, and rankings are kept separately for each backend. For Blender 2.x we compare render seconds for one scene on one version with the same render method.

The site shows the median after filtering with the 1.5 × IQR rule. Results with at least 20 runs are used in comparisons. Reference measurements are marked as a small sample. Runs under a shared GPU name count toward the model’s shared page. Measurements of a specific card need a confirmed GPU variant.

Comparison verdict

We do not calculate an overall 0–100 score. The verdict is built per scenario: single-thread and multi-thread work, gaming and power efficiency for CPUs; rasterization, ray tracing, compute and efficiency for graphics cards. If one model leads in one scenario and the other leads in another, the verdict is “depends on the task”. If there is too little data, the verdict says so.

Prices

The launch price is the manufacturer’s suggested price in US dollars at release (from official pages and Wikipedia). The English version shows it in dollars, rounded to whole dollars. It is a reference point, not a store price.
A store price is an observation: store, region, currency, availability and check time.
Fresh means up to 6 h old; recent, up to 24 h; stale, up to 72 h (shown dimmed and not used for the “best price”); anything older appears only in the history.
Prices from different regions are not mixed; a price without a region is not used in calculations.
Affiliate links are marked “Ad” with an erid identifier; they do not affect the choice of the “best price”.

Price/performance

benchmark score ÷ price in rubles × 1000 (points per 1000 ₽). It is calculated only for a comparable “higher is better” benchmark with current prices from one region, collected no more than a day apart.

Compatibility

We check the socket and memory type against official data. A matching socket does not guarantee it works: a BIOS update may be needed, so such a case is marked “needs checking”. For graphics cards, the power supply is compared with the manufacturer’s recommendation, and the length only for a known card model.

AI explanations

When the feature is on, the AI receives only structured facts from the comparison page with source identifiers and explains them. The AI does not add numbers and is not a measurement.