Skip to content
CPU VERSUS — home
methodology methodology transparency scoring

How We Rank CPUs: Our Methodology Explained

A walkthrough of the four-dimension weighted scoring system we use across the site, including the data sources, the math, and where the model breaks down.

By CPUVersus Editors 5 min read

Most CPU review sites either give you a single composite “winner” with no explanation, or they hand you twenty individual benchmark charts and let you do the work. Neither is great. The first hides the trade-offs; the second buries them.

CPUVersus tries to split the difference: a transparent weighted scoring system, four dimensions, sliders you can move, and the formula visible right next to the result. This guide explains exactly what’s happening under the hood.

The four dimensions

Every CPU on the site is scored along four orthogonal dimensions:

  1. Gaming. Built from gaming-relevant benchmarks — single-core scores from Geekbench 6 and Cinebench 2024, weighted toward latency-sensitive metrics.
  2. Productivity. Built from multi-threaded benchmarks — Cinebench 2024 nT, Geekbench 6 multi-core, and similar parallel-throughput tests.
  3. Value. A normalized inverse of MSRP. Cheaper chips score higher, by design.
  4. Power Efficiency. A normalized inverse of TDP. Lower-draw chips score higher.

Each dimension is normalized to a 0–1 scale against the highest-scoring chip in our catalog for that metric. So “1.00 in Gaming” doesn’t mean “perfect gaming chip” — it means “the best gaming chip we’ve measured.” Every other score is relative to that.

Where the data comes from

We are strict about provenance. Every number in the database has a documented source:

  • Geekbench 6 scores are pulled from the public Geekbench Browser API. Aggregates use the median of recent submissions for each chip. The sync runs nightly.
  • Cinebench 2024 scores are entered manually from named reviewer sources (Gamers Nexus, Hardware Unboxed, etc.) with each row carrying a source_url field.
  • MSRP and specifications are entered from manufacturer-published data with a citation URL stored alongside.
  • Retailer prices are placeholder for now and will be populated in Phase 8 of the site build.

We do not scrape commercial benchmark databases. We do not accept payment from CPU manufacturers. We do not adjust scores to favor any particular brand. The full data policy is in the methodology page.

The weighting formula

The weighted score for a CPU, given a set of user-selected weights, is:

score = (gaming × wG + productivity × wP + value × wV + efficiency × wE)
        / (wG + wP + wV + wE)

Where each sub-score is the chip’s normalized 0–1 value on that dimension, and the weights are integers from the sliders (default: gaming 50, productivity 50, value 25, efficiency 25).

A few properties this formula has on purpose:

  • Doubling all weights produces the same result. The sliders express relative importance, not absolute. If you set every slider to its max, you get the same ranking as if you set every slider to its midpoint.
  • Missing data doesn’t penalize. If a chip has no Cinebench score in our database, that metric is excluded from the divisor rather than counted as zero. A chip can’t lose points for benchmarks we haven’t measured yet.
  • Each contribution is visible. The comparison tool breaks down the final score into its four contributing pieces so you can see where the differences are coming from. We don’t hide that detail.

Where the model breaks down

No scoring system survives contact with reality cleanly. Some honest caveats:

Gaming is not one workload. Our gaming sub-score is built from synthetic CPU benchmarks that correlate with gaming performance, but they don’t capture the cache-sensitivity that makes 3D V-Cache parts dominate specific titles. We surface this explicitly in head-to-head comparisons involving X3D chips, but the single-number score under-credits them.

Value is sticker-price-only. We use launch MSRP because street prices fluctuate. That’s a defensible choice, but it means the score doesn’t reflect today’s actual deals. A chip whose MSRP was high at launch but which sells at a steep discount today scores worse than it should.

Power efficiency is TDP-only. Real-world average power draw varies wildly with workload, cooling, and motherboard power limits. TDP is a proxy, not a measurement. A chip with a 105 W TDP might actually pull 130 W under sustained load on a generous board.

The catalog is finite. Our normalization is against the chips we have in the database (currently 30). When a new flagship is added that out-scores everyone on, say, multi-core, every other chip’s productivity score adjusts downward. That’s mathematically correct — these are relative scores — but it’s worth knowing.

How to use it well

Three suggestions:

  1. Set the weights to match your actual workload, not your aspirational one. A weekend gamer who builds for 8 hours of work and 2 hours of play should weight productivity higher than gaming, even if gaming is the more interesting use case.
  2. Look at the per-dimension breakdown, not just the total. A 6%-higher composite score might come entirely from one dimension you don’t care about. The contribution chart in the comparison tool surfaces this.
  3. Treat the score as a starting point, not a verdict. The point of the system is to help you ask the right questions, not to substitute for thinking. The verdicts on our /vs pages explicitly call out cases where one chip wins on the score but loses on the workload that probably matters to you.

Try it

The methodology lives in code and the code is straightforward. If you want to play with the weights:

We’d rather have you understand and disagree with our numbers than accept them on faith.