How we score

Scoring model version v2-absolute.

Absolute performance scores

We score each metric against a fixed reference value, the raw figure worth 100 points. We pick that figure once and never rebase it. A component's composite is a weighted sum of its per-metric points, uncapped. This buys you two things.

• An existing score never moves when we add faster hardware.
• A newer, faster part scores above 100, so the scale holds its meaning across generations.

Weights per use-case

Each use-case weights the metrics differently. Gaming leans on raster compute and clocks; AI leans on tensor throughput and VRAM; VR rewards bandwidth and frametime consistency. VR headsets score on their own axes, where pixel density and refresh rate carry the weight. Weights sum to 1, so reference-class hardware lands near 100.

Value score

Value score is performance per dollar: (composite / price) × 500. So 100 points of performance at $500 reads as 100. When we have a live retail price for a part we use that, so the value score tracks what you actually pay. When we do not, we fall back to launch RRP.

Data sources

We take specifications from manufacturer sheets and Wikidata, synthetic performance from open datasets (OpenBenchmarking, MLPerf), and usage context from the Steam Hardware Survey. We never scrape terms-protected sites. Raw specs and benchmark numbers are facts; we compute the composite scores ourselves.

Known limitations

• No free bulk source publishes per-GPU gaming FPS, so gaming scores lean on synthetic proxies until user-submitted results build up.
• We normalize GPU AI throughput to a common FP8 basis so different generations compare fairly.
• AMD AI figures stay provisional until real MLPerf imports land.
• Varifocal headsets (Varjo XR-3, VR-3) score on average pixel density, not their sharper focus-area peak. Their detail pages flag this.