The Ray Accelerator is a Fixed-Function Unit added to the Texture Filtering and Mapping Unit Pipeline., this creates what AMD is calling the "Texture Processing Unit" and is entirely new to the RDNA 2.0 Architecture.
As it stands all AMD GCN 2.x and after Hardware are capable of running DXR 1.0 in Fallback Layer., which switched to a DirectCompute ("Software") approach., however as of DXR 1.1 (DirectX 12 Ultimate, and it's "Official" Release), Microsoft have ceased development and support of the DXR Fallback Layer due to "The Growing Options for Hardware-Based Acceleration"
I however believe this is premature as neither AMD nor NVIDIA currently offer a Mainstream Graphics Card with Hardware DXR Support., and we aren't expecting to see one for at least 3-6 months.
Meaning the only "Budget" Option for DXR Support is the Xbox Series S.
•
Now as it stands information on AMD's Ray Acceleration performance has been leaked out.
At present this is the DXR Procedural Geometry Sample., on an "Unknown" RX 6000-Series Card.
Providing 34fps (Fallback) and 471fps (Hardware).
This places said Card roughly on-par with the RTX 3070 in terms of DXR Performance.
Based on my Benchmarks (RX 5700 XT - 14fps Fallback)., well this would strongly suggest that the showcased hardware is the RX 6800 XT.
In fact we can use this to est. the Performance we should see from all of the RX 6000 Cards.
RX 6800 - 28fps (Fallback) - 388fps (Hardware)
RX 6900 XT - 34fps (Fallback) - 471fps (Hardware)
RX 6900 XT - 38fps (Fallback) - 529fps (Hardware)
Meaning it is approx. 26% Slower than the RTX 3080 and 37% Slower than the RTX 3090 at Ray Tracing.
With this said., this isn't a "Real World" example.
And this is just assuming Game Clocks... it's almost certain that these figures will likely increase up to 10% when Overclocking (to similar limits as the RTX 30-Series Drivers already do)., plus we could see some additional improvements from RAGE Mode.
Remember this is just a Synthetic Benchmark., testing purely the Raw RayTris/Sec Performance.
Actual Real-World examples are going to more interesting performance benchmarks... as how AMD and NVIDIA have implemented Hardware Ray Mathematics is quite different.
For AMD, it is part of the Graphics Pipeline., where-as for NVIDIA it's a separate Processor Core.
As such., I think it isn't too surprising that AMD losses in a "Pure" Ray Tracing Scenario (but that it's by only 25%, when remember the RTX 30-Series gains over the RTX 20-Series is ~50%) but in "Hybrid" Ray Tracing Scenarios we might see very similar performance; as the AMD design doesn't have to wait on Barriers or handle Ray-Tracing separate from the Raster Pipeline.
I still think NVIDIA will have an advantage... which is to be expected from 2nd Gen Hardware and almost 2 years extra Software Development; but if AMD can see a rise in Market Share (not to mention their implementation being on Consoles).
Well this then does mean that we could see optimisations or at least more considerations for AMD Hardware., given the poor sales of RTX 20-Series and Low Volumes for RTX 30-Series; meaning we could see a reasonable parity in Market Share for DXR Hardware Adoption.
And really I'd argue more that the Mainstream (which has yet to be released) is an even bigger factor.
NVIDIA last gen omitted RTX Support from their Mainstream Hardware... likely due to the cost overhead.
They can't this generation as it is baked into AMD RDNA 2.0 Architecture (every RDNA 2.0 based GPU will have DXR Hardware Support., just maybe not great performance).
So this opens up the question as to who can provide the most / best performance at the Sub $300 Price Point.
