About & methodology

webgpu.report collects WebGPU capability reports from visitors' browsers and publishes aggregate statistics: which optional features are available in the wild, what the limits look like, and how both trend over time. The landing page shows you your own machine compared against those aggregates; Explore and the public API expose the full dataset.

What is collected

When you load the landing page, a small script reads what the WebGPU API itself reports: adapters (requested at high-performance and low-power), their features, limits, and adapter info (vendor, architecture, device), WGSL language features, and the preferred canvas format. Alongside it we record browser and OS family (from the user agent and client hints), screen size, and hardware concurrency. Machines without WebGPU submit exactly that fact - availability failures are data. The full schema is versioned; press Download Report on the landing page to see precisely what your machine submits.

Metric definitions

• Unit: one accepted report per machine per day (reloads deduplicate; nothing is weighted by reload count). Headline stats include only first-party, non-quarantined submissions.
• WebGPU availability = reports with a working adapter ÷ all reports. Software-fallback adapters count as available but are also published as a separate split.
• Feature availability = reports with the feature on the primary hardware adapter ÷ reports with WebGPU on a hardware (non-fallback) adapter.
• Trends compare the current window against the same-length window one year earlier with identical filters.
• Limit percentiles are computed per day and combined across a window as an n-weighted median of daily percentiles - an approximation that keeps queries instant.

Sample caveats

Visitors of a WebGPU statistics site skew developer and enthusiast. Numbers here describe this sample, not the global installed base - treat them as an upper bound on capability availability. Fleet machines behind one network egress deduplicate into fewer reports. Every stat page shows n so you can judge significance yourself.

Data integrity

The submission API is open, so we bound and audit rather than pretend to prevent abuse: schema and plausibility validation, per-network daily caps, deduplication, quarantine of reports carrying unknown tokens, an origin split that keeps third-party embeds out of headline stats, and a daily distribution-shift monitor. Raw reports are retained so aggregates can be rebuilt after any cleanup.

Who runs this

webgpu.report is built and operated by Company Named. For questions about the data, a correction to a misclassified adapter, or an API key for sustained heavy use, reach us through the contact details on that site.