Initial Radeon vs. GeForce Vulkan Ray-Tracing Performance On Linux

[dpanter]

Quick Quake2RTX test with 6900XT in 3440x1440... ~35fps with 100% resolution scaling and settings turned up as much as I understand them. For reference a RTX2070 with same resolution and settings saw ~25 fps.

[ThoreauHD]

The 6900 XT is now the highest clocked GPU in the world. Had to be unlocked first in software.

https://youtu.be/fhP46XWMkdY

[Stefem]

That correct AFAIK. Yeah AMD doesn't have hardware dedicated ONLY to RT. But its compute units can perform tasks to accelerate RT. So yeah, I'd guess there would be some sacrifices. What interest me really is that according to some sources NVIDIA hardware provide a way to speedup build of BHV and for games that might provide huge benefit for RT.

AMD only accelerate ray/triangle intersection, BVH traversal is done on the shader core while NVIDIA do it with dedicated hardware (APIs have good abstraction that allow it) but update and buid of the acceleration structure is done on the CPU by design.
There have been always big difference in ray tracing performance based on the scene so those results doesn't really surprise me, overall NVIDIA is still on a different league in real world usages

[blacknova]

AMD only accelerate ray/triangle intersection, BVH traversal is done on the shader core while NVIDIA do it with dedicated hardware (APIs have good abstraction that allow it) but update and buid of the acceleration structure is done on the CPU by design.
There have been always big difference in ray tracing performance based on the scene so those results doesn't really surprise me, overall NVIDIA is still on a different league in real world usages

According to RDNA2 ISA 8.2.10

Ray Tracing support includes the following instructions:
• IMAGE_BVH_INTERSECT_RAY
• IMAGE_BVH64_INTERSECT_RAY
These instructions receive ray data from the VGPRs and fetch BVH (Bounding Volume Hierarchy) from memory.
• Box BVH nodes perform 4x Ray/Box intersection, sorts the 4 children based on intersection distance and returns the child pointers and hit status.
• Triangle nodes perform 1 Ray/Triangle intersection test and returns the intersection point and triangle ID.

So It does accelerate BHV traversal and ray/triangle intersection.

[Stefem]

According to RDNA2 ISA 8.2.10


So It does accelerate and ray/triangle intersection.

Those instruction run on the shader block, AMD lack dedicated silicon for BHV traversal

[smitty3268]

Those instruction run on the shader block, AMD lack dedicated silicon for BHV traversal

Semantics. It's hardware accelerated.

[artivision]

Semantics. It's hardware accelerated.

BVH traversal on AMD is accelerated by shader cores, it cuts horse power from the gpu. But this is more versatile from the other hand and can make better scenes. On those complex scenes NV will lack more.

[pinguinpc]

Very interesting outcome, although like other commenters have said, it seems to mostly fall into two buckets for now.

I'd be interested in more benchmarks: trough wine/proton, and Quake II RTX.

Besides this, can this vulkan-based RT API be leveraged by blender? Would that be more efficient than OpenCL?

this maybe can help you but only in windows

https://www.game-debate.com/news/29996/rx-6000-series-runs-nearly-50-slower-at-ray-tracing-than-rtx-30-series-in-quake-2-rtx

[bridgman]

AMD lack dedicated silicon for BHV traversal

I don't think this is correct - we just put the dedicated silicon in the shader core rather than segregating it into a separate block.

At the risk of stating the obvious the RT instructions make use of the dedicated RT logic.

[smitty3268]

BVH traversal on AMD is accelerated by shader cores, it cuts horse power from the gpu. But this is more versatile from the other hand and can make better scenes. On those complex scenes NV will lack more.

Nobody will argue if you say that NVidia dedicated more die space to ray tracing hardware. But when someone says it's hardware accelerated and then someone else chimes in that it isn't, the 2nd person is unequivocally wrong. It is hardware accelerated. Every hardware design is going to have different tradeoffs about what they view as more or less important.