RADV+ACO Outperforming AMDVLK, AMDGPU-PRO Vulkan Drivers For X-Plane 11.50 Flight Simulator

[Azultra]

It's sad that Valve needs to do this because AMD doensn't seem to be capable or willing. What does nvidia have to do with this?

What does ACO have to do with the abilities or will of AMD developers?

AMD put resources into the LLVM AMDGPU backend which was good enough, much later Valve came forward with an alternative implementation from scratch that improves on it, maybe after analyzing the hard-to-fix issues of AMD's initial approach. What is sad about it and why do you want to see it as an indicator that AMD is less competent than Valve ? On the contrary I think it's pretty incredible that it's possible today for a third-party to contribute entirely on its own initiative a large improvement to a graphics driver.

[spykes]

It's a good progress, but I wonder how a RTX 2060 manages to be faster than a RTX 2060 Super and a RX 5700 XT... There is something wrong with those numbers.

[duby229]

It's a good progress, but I wonder how a RTX 2060 manages to be faster than a RTX 2060 Super and a RX 5700 XT... There is something wrong with those numbers.

You're seeing a bottleneck at the CPU in the graphics driver or game logic. (It would take a series of synthetic tests to figure out what.)

EDIT: After looking at it more, it seems like a CPU bottleneck in the game logic somewhere. In most of those cases it shows the nVidia cards performing very near to each other, same with AMD on ACO, and also same with AMD on AMDGPU-PRO. So the drivers are each achieving what their maximum performance is, therefore the bottleneck must be somewhere in the game logic.

[microcode]

Wow, those peak frame latencies are great. Good job RADV/Mesa people.

[bug77]

Kind of sad when 3rd parties can write better drivers than the team with direct access to hardware people

It's a tradition from the ATI days.

[qarium]

Wow, those peak frame latencies are great. Good job RADV/Mesa people.

yes it is really amazing. my vega64 cards love it.

[M@yeulC]

I was not going to give into that discussion, but come on...

RADV is based on the awesome work AMD devs made on the kernel module, the LLVM compiler, the RadeonSI OpenGL driver (and Mesa components), and it takes some of the AMVLK code for inspiration. I don't want to diminish RADV developers' achievements here, but saying this is all of the work of the community is overstating things a bit. Everyone stands on the shoulder of giants.
The work on ACO was done and funded by Valve. Why? Because they want to, and they can. Period. I don't think they could do the same with a proprietary blob. They want to push the state of the art further. And it works. These frame times are impressive! [strike]I don't think I saw a single frame below 60 FPS in the red camp (worst~10ms IIRC -- 100FPS).[/strike] That was the first test. 20ms it is, but even the other versions are quite good. I'm tired, but I don't get why max frametime is 10ms on one graph, and average is only about 60FPS...

Also a reminder that FPS is not a good metric: the cards are much closer if you compare the frame times

[mphuZ]

Kind of sad when 3rd parties can write better drivers than the team with direct access to hardware people

Don't be absurd.

https://www.phoronix.com/vr.php?view=28951

Draw conclusions from one indie game, where GTX 1060 = RTX 2060 Super.....
You can already see that the game has poor optimization.

[finalzone]

Also a reminder that FPS is not a good metric: the cards are much closer if you compare the frame times

Also FPS failed to take account the quality of frame rate and animation.

[Linuxxx]

I wonder if it says anything positive that the radeon drivers all have a small deviation in frame times, while the Nvidias seem to have high extremes in very long frame times.
Coming from an audio corner, you'd rather want low jitter. I don't know if that's the same for video.

Oh it's definitely also true for video & even more so for videogames (the low jitter preference).

If Michael had used the lowlatency kernel flavor instead of the generic one, the frametimes on the nVidia side would have been more competitive.
(I mean really, on a desktop system it's the only sane choice, especially if one prefers an actually enjoyable gaming experience on Linux!)

Also, here are a few more things to consider:

• The kernel versions are different (5.4 on nVidia, 5.6 on AMD).
• The nVidia stable driver used here still only exposes Vulkan 1.1, whereas the AMD bleeding edge driver is already on Vulkan 1.2
• Do note that the nVidia Vulkan 1.2 beta driver improves the performance ~25%!

Another point of interest is the chosen performance CPU scaling governor:

From my own experience with AMD's Zen design, it hinders the CPU from hitting the maximum boost clock; and since this particular engine seems to be particularly CPU-bound, another governor would have been more preferable. (Guess which one?!)