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

[illwieckz]

AMD DID get crushed. Look at the Geometric Mean results again.

There is not enough data to make geometric mean relevant. Also we don't know how much those benchmarks are significative, and if their significativeness is equal. If there was 5 more GeeXLab-like benchmarks tested the geometric mean would be very different and Nvidia would looks like a joke, though in reality there would be nothing tested more than what we see.

[cb88]

I suspect that cb88's point was that the tests Michael ran all seemed to fall into one of two buckets, where all of the tests in a bucket had similar performance characteristics.

My understanding is that the two buckets are (in DXR-speak) triangle mesh geometry vs procedural primitive geometry. At a 10,000 foot level, NVidia seems to be faster with triangle mesh geometry while we seem to be faster with procedural primitive geometry, which makes sense when you consider the implementations.

https://microsoft.github.io/DirectX-...-mesh-geometry

That is vaguely what I meant yes, since we are in fact looking at RT microbenchmarks effectively, though you added more knowledge to it... real games aren't going to usually be entirely one or the other.

[cb88]

I have seen the argument in the past that the reason AMD Ray Tracing showed such awful results against Nvidia was due to the test environment. The premise starts with the initial tests being games in Windows, and that all game developers using ray tracing had developed specifically for Nvidia's solution (intentionally or not). AMD did not recreate RTX, having their own priorities and a different design around ray tracing, therefore they underperformed on what were effectively RTX tests. The reasoning went that if the test software had been developed as much for AMD, AMD's ray tracing results would look fine.

I had no way to know how much truth there was to that idea (other than that games were developed for RTX, which is true enough). I hope these results are a sign that there was something substantial to the argument and that AMD will quickly establish a strong position in the new market. Tyranny of a single vendor has never been good for anything.

It probably still remains to be seen if procedural geometry or triangle mesh geometry dominates... honestly I'd bet on triangle mesh at this point, AMD certainly can't afford to be slower at either. Basically it's tesselation all over again...

[Teggs]

I have seen the argument in the past that the reason AMD Ray Tracing showed such awful results against Nvidia was due to the test environment. The premise starts with the initial tests being games in Windows, and that all game developers using ray tracing had developed specifically for Nvidia's solution (intentionally or not). AMD did not recreate RTX, having their own priorities and a different design around ray tracing, therefore they underperformed on what were effectively RTX tests. The reasoning went that if the test software had been developed as much for AMD, AMD's ray tracing results would look fine.

I had no way to know how much truth there was to that idea (other than that games were developed for RTX, which is true enough). I hope these results are a sign that there was something substantial to the argument and that AMD will quickly establish a strong position in the new market. Tyranny of a single vendor has never been good for anything.

[bridgman]

I suspect that cb88's point was that the tests Michael ran all seemed to fall into one of two buckets, where all of the tests in a bucket had similar performance characteristics.

My understanding is that the two buckets are (in DXR-speak) triangle mesh geometry vs procedural primitive geometry. At a 10,000 foot level, NVidia seems to be faster with triangle mesh geometry while we seem to be faster with procedural primitive geometry, which makes sense when you consider the implementations.

https://microsoft.github.io/DirectX-...-mesh-geometry

[Jumbotron]

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.

#1. Silly and uninformed to say that Geometric Mean results mean nothing. If I'm looking at "balanced performance" for any particular part, in this case a GPU, then Geometric Mean results are what I'm going to begin with in my search for the right product at the right price. Of course if my needs are very narrow I will look at individual benchmarks that replicate my use case as closely as any benchmark can, which of course is approximate to begin with.

#2. I don't know which 2 tests you are referring to when you stated without proof they were run ineffectively but seeing as how you know so much about Michael's benchmarks in general and the code in question of the two "ineffectively run" benchmarks you mysteriously mentioned, then I'm sure either the Benchmark creators and / or Michael would appreciate any optimizations or corrections to said benchemarks.

#3. Which begs the questions...have you submitted a bug report AND are you going to request a pull of your corrective code to said two "ineffectively run" benchemarks?

[onlyLinuxLuvUBack]

Dwarf fortress vulkan raytraced... no love for thee.

[ATFx]

Spoiler: Ray Tracing is More Demanding Software. AMD outperforms in some cases because it's (cough) driver is better optimized in these workloads Nvidia's is not vice versa.
It's a shame AMD has low market shares across platforms still(According to Steam hardware Survey), and that many continue to underestimate them.

[ernstp]

I think one thing this shows is that these RT accelerators are extremely sensitive to how you use them exactly.

https://www.youtube.com/watch?v=cC-DDAq3PCM

Here AMD talks a little about their "optimal threadgroup size" and what memory you should use etc...
Probably hard to do a one-size-fits-all implementation in a game..

[blacknova]

Last I checked, AMD effectively doesn't have discrete processing cores for RT. RDNA2 does have compute units specific for ray tracing, but it comes with sacrifices. This is a good thread to read:
https://www.reddit.com/r/Amd/comment..._accelerators/

Part of me wonders if AMD surges ahead in benchmarks where textures don't need much processing.

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.