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 — the button above runs them under the same measurement engine, scored in their own space and never submitted anywhere: pure nostalgia, honestly measured. And every scene, both tiers, can run unscored in the effects gallery — with the particle load under your control.
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 windows of 0.7–1.6 seconds, 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 in the 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, not discarded. Each scene rechecks a baseline load at the end; if the device slowed more than 15% mid-scene, that scene is flagged. A run whose scenes were all measured and whose only issue is that drift is marked throttled — it scores normally and ranks on the board wearing the flag, because a warm device is still a real measurement.
- Scores are versioned. This is benchmark 2.2.0. 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 197 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 in What 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.