WebGL performance testing on real devices

Mobile GPU-land is wildly fragmented: Adreno, Mali, PowerVR and Apple GPUs each have different fill-rate limits, shader compilers, texture bandwidth and driver quirks. A scene bound by fragment-shader cost on a Mali-G52 can be completely fine on an Adreno 640 from the same year. CPU throttling in Lighthouse or DevTools tells you nothing about any of this — the only honest measurement is your actual frames on actual silicon.

Then there's heat. Phones throttle hard: a demo that opens at 55 fps can settle at 30 fps after two minutes of sustained GPU load. A 30-second manual check on a cold device misses it entirely; measured runs over time catch it.

You give us a URL and, optionally, a Playwright-style script that drives the app into its heaviest state — load the scene, start the effect, zoom the map. We execute it on a curated spread of real devices, from current iPhones and Pixels down to the 2019–2021 mid-rangers that dominate real-world traffic, and send back per device: median & p95 FPS, frame-time distribution, jank events, and thermal behavior across the run.

The plan for launch includes a CLI and GitHub Action with FPS budgets per device — “fail the build if the gallery scene drops below 30 fps on a Galaxy A52.” Performance stops being a vibes-based launch-week discovery and becomes a number in your CI, like test coverage.