A benchmark in the hardware team’s hands
We work with a mobile silicon vendor on graphics evaluation, supplying Evolve benchmark builds and neural-radiance-caching workloads from our Breda renderer. Their engineers run these workloads on their own hardware and bring the results back to our team.
That arrangement connects the renderer to the questions a GPU team actually needs answered. Where does a frame spend its time? How does a neural workload interact with the rest of the renderer? Which behaviour comes from the technique, and which from the hardware or the driver?
Working from the implementation
Because we build the rendering systems behind the workloads, we can investigate those questions in the code rather than around it. Pass-level analysis and direct engineering discussion connect a measured result back to the work producing it.