Jawed
Legend
We can see that Ampere versus Turing shows a strong dependence upon triangle intersection rate.I think you’re saying that Ampere pulls ahead with triangle based geometry because of its advantage there and not primarily due to being faster at traversal.
Scene 3, which mixes triangles and procedural, appears to have the most complex BVH. It's 61% faster on Ampere, which appears to demonstrate that it's triangle acceleration that's the win.
Scene 4, which admittedly has nearly no triangles and presumably has the smallest BVH of all five of these test scenes, shows a 35% gain. This may be fully procedural?
Scene 5 shows an 87% gain for Ampere in something that looks to be entirely triangle based (though the Cornell box may be procedural).
So two scenes that are dominated by triangle-based geometry are scaling strongly with Ampere's triangle acceleration.
Scene 3 with seemingly the most complex BVH is the same speed on 6900XT and 2080Ti. There's obviously some procedural geometry to "slow down 2080Ti", but the scene is dominated by rough materials (which creates a lot of ray divergence) along with caustics that suck up rays like a sponge. The depth of field in this scene adds to ray divergence, too.
There's a lot of reasons for ray divergence to be seen as a problem in this scene, but it doesn't seem to be dominant.
I agree we need better experiments.There’s probably truth to that but the other way to interpret the test results is that RDNA 2 is faster at intersecting procedural geometry which negates Ampere’s traversal advantage. We would need more controlled experiments to be sure.
It looks like RDNA 2 is faster at intersecting procedural geometry, but why? It's shader code isn't it? Shouldn't that be "FLOPS-bound"? Maybe it's not FLOPS-bound because it has to do with scheduling and latency-hiding.