Question regarding the latency of the intersection unit

[leonhsi]

Amazing work but I have a few questions for this project.

How do you estimate the latency of the intersection unit in the ray predictor in src/gpgpu-sim/ray-predictor.cc? That means how did you benchmark the real intersection unit in NVIDIA RT core? Did you use some micro benchmark to achieve this estimation?

Thanks, looking forwards to your reply.

[lucylufei]

Hi Leon,

Thanks for your interest in our work!

Since NVIDIA does not reveal details of their implementation, we estimate the intersection unit latency based on the T&I engine. As I also mentioned in my email, we did not microbenchmark the NVIDIA RT Core for the precise number of cycles / pipeline stages.

The latencies in ray-predictor.cc are modelling access latency to the predictor table and not the intersection computation itself. The ray-box and ray-triangle intersection test latencies are modelled in the rt_unit in shader.cc. These latencies can all be modified in the configuration file.

Please let me know if you have any further questions 😊