The soft-shadow stage renders one scene two ways, with soft shadows and with hard ones, and reports the frame rate of each and the cost of the soft version.
- The segments run in ABBA order, soft, hard, hard, soft, soft, hard, so a GPU that warms or slows over the stage affects both kinds alike.
- Since measurement revision 2026.09.12 the stage is a 6 second warm-up and six 9 second segments, 60 seconds in all.
- A segment that loses more than half its windows to a hidden tab or lost focus drops its whole soft and hard pair, so the balance holds.
The question
Soft shadows look better than hard ones and cost more to draw. How much more depends on the GPU, so SystemCheck measures it directly: the same scene, the same camera path and the same resolution, drawn once with soft shadows and once with hard shadows. The difference between the two frame rates is the cost of the soft shadows on that machine. It runs as stage 6 of a Deep run and on its own at /soft-shadows.
Why the order matters
A GPU under load does not hold one speed. If all the soft segments ran first and all the hard ones after, any slowdown over the stage would land on the hard segments and make soft shadows look cheaper than they are. Alternating soft and hard in ABBA order, soft, hard, hard, soft, soft, hard, puts each kind early, in the middle and late in equal measure, so a steady drift cancels instead of favouring one side.
Since revision 2026.09.12 the stage runs a 6 second warm-up and then six segments of 9 seconds, 60 seconds in all. The camera repeats its path once per segment, so each segment draws the same views.
Keeping the pairs whole
A segment measured while the tab was hidden or the window was out of focus describes the browser rather than the GPU. One-second windows that overlap such a span are left out. A segment that loses more than half its windows drops its whole soft and hard pair from the cost, so the remaining pairs stay balanced. With no complete pair, the report says the cost was not measured.
What the method showed about itself
When the stage moved to the current schedule, the old and new schedules were run interleaved on the validation laptop. The new schedule alone read about 10% lower on the soft-shadow rate and 7% lower on the hard one, which is why reports from before and after that revision are not compared. That result came from one laptop, and SystemCheck states it as such.