We don't quote performance.
We measure it.

Frames per second with a fleet of eight on screen, every effect running.

A 2017 phone holds its panel's locked 60 with eight fully-dressed ships; the desktop runs the same scene at triple-digit headroom. Single ships are never the question.

The OnePlus 5T panel refreshes at 60 Hz - the dashed line is its ceiling, not ours.

* pinned at the panel refresh - real headroom is higher than shown. On the cold 2026-07-18 pass all three sit on the cap: the Mobile tier never drops a frame, and the desktop-grade Powerhouse tier holds 60 too. The only run where a single Powerhouse ship measured below the panel (56 fps) was the thermally-degraded 2026-07-11 handset - so treat 56 as the hot-device floor, not the number this phone gives you cold.

Milliseconds added per frame over the Mobile-tier baseline, on the OnePlus 5T. These come from the 2026-07-11 run, on a handset that had been benchmarked all evening and was thermally throttled. That is deliberate: when we re-ran it cold on 2026-07-18 the phone pinned at its 60 Hz panel in every single-ship scenario, so each feature subtracted to +0.00 ms - which measures the panel, not the effect. The hot figures are the honest upper bound, and the cold result is shown beside each row.

Distortion is the only feature with a real mobile price tag - which is exactly why it ships desktop-gated. Everything else, trails included, is effectively free.

• 17 scenarios, ~6 minutes. Empty floor, hull-only isolation, one ship and an 8-ship fleet at both tiers, one run per premium effect, subtractive isolation of the always-on systems, and a 30-second soak with live state cycling and FX rebuilds.
• Real flight, real pixels. Ships fly nose-first at full throttle so trails bend and particles emit like in a game; the camera keeps hulls and effects filling the frame so mobile fill-rate is genuinely stressed.
• Recorded: average fps, p50/p95/p99/worst frame times, decay slope, GC, managed and native memory slopes, peak particles, and every logged error. Zero errors, both devices.
• Three passes, median reported, first pass discarded. The opening pass is always low while shaders warm up, so a single run understates the machine. Passes two and three agree within about 1%.
• Every panel names its pass. Two benchmark runs feed this page: 2026-07-18 on a cold handset, which is where the headline fleet and single-ship figures come from, and 2026-07-11/12, which by its own report ran hot. We do not silently blend them. Where the older pass is the only one that could measure something - the per-feature costs, because a cold phone hides them all under its 60 Hz panel - the panel says so and shows both results.
• A throttled device is a feature here, not a flaw. Reporting the hot handset's per-feature costs rather than the cold run's zeroes overstates what the effects cost you. That is the direction we would rather be wrong in.
• Measured on a working desktop, not a clean-room bench. The dev machine runs its normal load while benchmarking. These are a floor, not a ceiling: a quiet machine measures higher. We would rather publish the number you would actually hit on a busy workstation.
• Honesty rules. Uncapped everywhere; refresh-capped rows flagged, never extrapolated; the test hull is our own asset (not part of the pack). Older Adreno drivers run OpenGL ES 3 - Unity 6's Vulkan path is unstable on that generation; modern devices run Vulkan.