Portal with RTX has the most atrocious profile I have ever seen. It's performance is not (!) limited by it's RT performance at all, in fact it is not doing much raytracing per time bin on my 6900 XT. 
This is one GIGANTIC Ubershader that takes 99ms of frametime by itself. The hardware utilization is horrifically bad, and it's using the absolute maximum amount of VGPRs possible on RDNA2.
To show how bad it is, I have taken a capture from the start area in Portal with RTX. This has 323 VALU instructions per RT instruction, which is a BAD ratio.
Compare that to Control, which is generally regarded as a bad Game in terms of RT performance on RDNA2. Here we have a ratio of 82 VALU per RT instructions in the reflection shader, which is in the normal range (according to @b_nieuwen.)
So, if RT performance can't explain the poor performance, then what can? Well, one explanation would be vector Registers spilling to cache or VRAM. This shader is pegged at 256VGPRs, which is the maximum that's possible to use per wavefront.
I haven't tested it on RADV, but according to @b_nieuwen the situation is the same there, so it's likely not an issue with the Windows driver.
This is an interesting hypothesis on what could have gone wrong:
https://twitter.com/_plop_/status/1601163128363831296
An important clarification:
https://twitter.com/b_nieuwen/status/1601201811054284802
β’ β’ β’
Missing some Tweet in this thread? You can try to
force a refresh
γ
