Investigating the workload before the hardware arrives
2022 · Ray-tracing workload characterisation. A mobile GPU vendor commissioned a release report from us on mobile ray-tracing behaviour, covering sun-shadow workloads, ray coherence and TLAS/BLAS configurations.
The hardware the report was written for did not exist yet in a form we could run on. That is the normal situation in this kind of work, and it changes what the measurements are for: the question is not how fast a particular device is, but how the workload behaves and which of its properties will dominate once silicon arrives.
What the experiments needed to separate
A ray-tracing workload is shaped by more than its ray count. Ray coherence, scene organisation, the relationship between top-level and bottom-level structures, and shading work all affect the cost of a frame. The report explored those variables rather than treating a single benchmark score as the answer.
Measuring on the hardware that existed
We ran the workloads on a Galaxy S22 and on desktop RDNA2 hardware. Neither is the part the report was about, and that is the point: the phone supplies a real mobile memory and power envelope, and the desktop part supplies a hardware ray-tracing implementation whose behaviour can be read in detail. Between them, a variable's effect can be separated from any one device's quirks.
Every number carried that distinction with it. A measurement taken on a stand-in is evidence about the workload, not a prediction of the target's performance, and a report that blurs the two is worse than no report: it invites a hardware decision to be made against a number that was never about that hardware.