Evidence
Measured, not claimed.
Estimating accurately is the whole product, so here are both of our benchmark runs on rented hardware — including the one that went least well.
The benchmark runs
8 renders on 2 rented RTX 4090s. 0 over their estimate.
Four scenes — classroom and 3 generated test scenes (flat, ramping, volumetric) — each on two machines. “Actual ÷ estimate” below 1.00 means the render took less time than we estimated. You are billed for the minutes used either way; above 1.00 you would pay the metered actual, up to 1.25× the amount held when you approved, and a render that would cost more is stopped and parked rather than billed onwards.
2026-08-14
RTX 4090 · instance 47728910 · gate-reports/2026-08-14/
| Scene | Estimate | Held | Actual | Actual ÷ estimate | Actual ÷ held | Peak VRAM |
|---|---|---|---|---|---|---|
| classroom | 0.804 min | 1.028 min | 0.754 min | 0.94 | 0.73 | 3436 MB |
| flat | 0.512 min | 0.565 min | 0.484 min | 0.95 | 0.86 | 2754 MB |
| ramping | 0.490 min | 0.539 min | 0.468 min | 0.96 | 0.87 | 2754 MB |
| volumetric | 0.347 min | 0.433 min | 0.328 min | 0.95 | 0.76 | 2942 MB |
2026-08-15
RTX 4090 · instance 47823102 · gate-reports/2026-08-15/
| Scene | Estimate | Held | Actual | Actual ÷ estimate | Actual ÷ held | Peak VRAM |
|---|---|---|---|---|---|---|
| classroom | 0.962 min | 1.231 min | 0.808 min | 0.84 | 0.66 | 3436 MB |
| flat | 0.461 min | 0.577 min | 0.366 min | 0.79 | 0.63 | 2754 MB |
| ramping | 0.418 min | 0.460 min | 0.340 min | 0.81 | 0.74 | 2754 MB |
| volumetric | 0.334 min | 0.418 min | 0.260 min | 0.78 | 0.62 | 2942 MB |
Why those two runs disagree
Same scenes, same settings, same model of card — and the second run over-estimated noticeably more. The only variable was which physical machine the marketplace gave us. Rented hardware varies, and our estimates vary with it.
We could show you only the better run. It is more useful to know the spread: across both, the renders finished between 4% and 23% under their estimates. The amount held is the estimate times a buffer for how uncertain the scene is, and it ran between 15% and 61% above what the render took. Your scene is not one of these, and can land either side of its estimate.
The scenes we calibrate against
Published Blender scenes, rendered on hardware we rent to measure how fast each card really is. A scene is named and credited here only where we have read its licence off the page it was published on; one we could not is described instead.
The anchor. Real geometry, real textures, real light paths.
Heavy geometry and glass.
Fur — the only scene exercising the dense-hair surcharge.
The most-published benchmark scene, so our ratios can be checked against it.
Measured across three RTX 4090 hosts, first frame excluded. Every scene is rendered at our own normalised settings — 1920×1080, 300 samples — rather than the settings it ships with, so these figures are not comparable to published benchmark numbers for the same files. “CV” is how much the timing varies between machines: the lower it is, the more the scene is measuring the card rather than the host it landed on. No frame of any of these scenes is published: a licence from a scene's author does not clear what the scene depicts.