What it runs
The Standard run is five Signature scenes on WebGL2 — each stressing a different part of the pipeline — plus one deliberately plain Canvas 2D scene as a CPU-bound control:
| Scene | Stresses | Looks like |
|---|---|---|
| Nebula Drift | GPU particle update + fill rate | A million-point nebula advected through curl-noise flow. |
| Neon Metropolis | Draw calls + instanced geometry | An instanced emissive city flythrough. |
| Liquid Chrome | Fragment shader / raymarching | Raymarched metaball fluid with mirror shading. |
| Starfall Cascade | CPU physics + per-frame upload | Verlet debris streams colliding with moving geometry. |
| Hypernova | Mixed finale | Core collapse, shockwave, and a half-million-particle burst. |
| Canvas 2D Basic Circles | Canvas 2D fill + simple physics | Bouncing colored circles — the honest baseline. |
The eight original effects from Screen Test Pro v1 live on in Classic mode — same measurement engine, separate score space, pure nostalgia. And every scene can run unscored in the gallery.
How the score works
Each scene ramps its particle load up and down, searching for the heaviest load your device sustains within a fixed frame budget of 16.67 ms — checked against the 95th percentile of measured frame times over ~1.6-second windows, with several consecutive windows required to accept or reject a level. That "max stable load" is normalized against per-scene reference values and combined into a weighted geometric mean, giving a total score and a tier (Entry / Balanced / Fast / Extreme).
Some details worth knowing, all covered inthe full methodology:
- High-refresh screens aren't penalized. The budget is fixed at 16.67 ms regardless of whether your display runs 60 or 165 Hz.
- Backgrounding voids the window, not your run. Switch away and the current measurement is discarded and retried; the benchmark pauses rather than recording garbage.
- Thermal drift gets flagged. Each scene rechecks a baseline load at the end; if the device slowed more than 15% mid-scene, the result is marked degraded instead of silently blended in.
- Scores are versioned. This is benchmark 2.0.0. Any change to scene workloads bumps the version, and scores from different versions are never compared.
- Hard timeouts everywhere. A scene can't exceed 30 s, a run can't exceed 150 s, and every state has an exit. The benchmark can be cancelled, but it cannot hang.
What this can and can't tell you
This is a browser workload score. It reflects your GPU, CPU, browser, OS, thermals and settings acting together — which makes it a fair test of what your device actually delivers in a browser, and a poor substitute for a native benchmark when you're comparing silicon. We wrote up the difference honestly inWhat a browser benchmark can and can't measure. Two quick implications: close your other tabs before running, and treat small score differences (a few percent) as noise.
The anonymous sample
Finishing a Standard run submits one anonymous row — score, coarse hardware family, screen numbers — that feeds the public hardware benchmarks. Sharing is on by default; the start card shows the exact values your device would send, and unchecking the box there runs the benchmark without submitting anything — the choice is remembered. Details in the privacy policy.