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

[phoronix]

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

With today's Radeon Software for Linux 21.10 packaged driver release is the first time Vulkan ray-tracing is being exposed on Linux for AMD Radeon graphics cards with any of the multiple driver options. Here are some initial benchmarks looking at how the Radeon RX 6000 series Vulkan ray-tracing performance is on Linux compared to NVIDIA's Vulkan ray-tracing support with the existing RTX 20/30 series hardware.

https://www.phoronix.com/vr.php?view=30135

[tildearrow]

Oh wow, AMD outperforms NVIDIA even in ray-tracing in some cases?
And this is only their first try on ray-tracing. Amazing!

[schmidtbag]

Wow I wasn't expecting results like this. I assumed AMD was going to get crushed in all tests the first time around. Makes me wonder why the results vary so much though.

In either case, these are some promising first results.

[gotzl]

How about benchmarking Quake II RTX? It's running for me ...

[blacknova]

Interesting results. I wonder which exactly parts of tests were considerably slower on AMD. AFAIK AMD's hardware can accelerate BHV/triangle ray intersection only. So what exactly is accelerated on NVIDIA to make it faster in some cases?

[schmidtbag]

So what exactly is accelerated on NVIDIA to make it faster in some cases?

Having dedicated hardware to handle RT. The real question is why AMD was so much faster in some cases.

[blacknova]

Having dedicated hardware to handle RT. The real question is why AMD was so much faster in some cases.

Both have dedicated hardware. And both can accelerate BHV/triangle intersection. The question is what else is accelerated on NVIDIA? Or is it just a number of execution blocks dedicated to ray tracing? If so then why AMD perform faster in some cases? Ray Tracing should pretty much scale linear with number of dedicated blocks.

[Jumbotron]

All valid questions. Leads me to believe that one should look at GPU memory speeds and latencies and data interconnect differences along with cache types and how they are wired to the cores. Also bandwidth differences and width of data paths. All these things can make huge difference between certain benchmarks revealing strength and weaknesses of each cards' architectural differences and how that causes each card to react to different workloads. It seems to me that where AMD shines on certain benchmarks one could find similarities in those particular benchmarks' workloads that are amenable to AMD's GPU architecture above and beyond Nvidia. And likewise when Nvidia's GPUs across the board take a decisive lead over AMD.

[Jumbotron]

Wow I wasn't expecting results like this. I assumed AMD was going to get crushed in all tests the first time around. Makes me wonder why the results vary so much though.

In either case, these are some promising first results.

AMD DID get crushed. Look at the Geometric Mean results again. Out of 13 cards tested only 3 were AMD. The very best AMD card came in only 5th Place followed by a 7th Place finish and an 11th Place finish out of 13 cards. Those are not stellar results by any means.

[cb88]

AMD DID get crushed. Look at the Geometric Mean results again. Out of 13 cards tested only 3 were AMD. The very best AMD card came in only 5th Place followed by a 7th Place finish and an 11th Place finish out of 13 cards. Those are not stellar results by any means.

Geometric mean means little when you only ran two tests effectively... and the one Nvidia excelled at was still slow.

Anyway... more realistic benchmarks are need to draw real conclusions preferably in an open source game so we can be sure that it is fair to both vendors.