NVIDIA 387.34 vs. Linux 4.15 + Mesa 17.4-dev Radeon OpenGL/Vulkan Performance

[M@GOid]

There must be some performance regression on Kernel 4.15. My RX 470, paired with a i7 3770k @ 3,9GHz, only 8GB of RAM (DE:MD build ~1GB of cache on the SSD) on the stock kernel 4.13 from Ubuntu 17.10, manages 36 FPS on Deux Ex Ultra benchmark, how can the RX 580 only manage 34 FPS, with a much higher clocked (and 4 more threads) i7 and double (and faster) RAM?

[mphuZ]

RADV is a community driver and...

And I'm happy for you. But I'm talking about performance in general. About a specific model.

By the way, about the "impressive".. Open source Vulkan driver enter at least to the short list of AMD plans? Who knows?

[Leopard]

actually it is same driver, linux games perform worse than windows conterparts

No it isn't. Try Doom 2016 on Wine and then Windows with Amd. You will see Wine performs worse.

On Nvidia side , it is performing literally same on Windows and on Linux via Wine.

[theriddick]

It is a bit sad to see AMD struggle so much with Vulkan, given that they practically pushed it into existence. Nvidia has really turned the table on them on that one.

[azdaha]

It doesn't help things when Valve, AMD, others focus only on Ubuntu. Although, I guess we're always thankful enough that they're at least dedicated any resources focusing on Linux.

It might be worth noting, or perhaps requesting, that Phoronix does tests and results on ArchLinux or Fedora as well. I've posted a bit of my own experience in another thread about AMD GPUs and an older Nvidia GPU. As much as I hate to admit it, nvidia's GPU drivers are superior to AMDs. Although, trying to play a game on Steam in Linux is always more problematic than on Windows. While a lot of this is attributed to Valve essentially reverse engineering some of the stuff to port it to run on Linux within Steam, games like their own CounterStrike:GO shouldn't have this limitation. Nevertheless, the gameplay is smoother in Windows even with a lower-level GPU than in Linux. This has to be due to the drivers from the two GPU makers.

In terms of nvidia vs. AMD:
As an example, I have one of the AMD GPUs used in these tests, the Radeon RX-580. I find that gameplay experience and performance is better with an NVIDIA GTX-750Ti than with the newer AMD GPU. While, using a Radeon RX-550 before that, I was not even able to play once without serious performance issues even on Ubuntu.

[azdaha]

There must be some performance regression on Kernel 4.15. My RX 470, paired with a i7 3770k @ 3,9GHz, only 8GB of RAM (DE:MD build ~1GB of cache on the SSD) on the stock kernel 4.13 from Ubuntu 17.10, manages 36 FPS on Deux Ex Ultra benchmark, how can the RX 580 only manage 34 FPS, with a much higher clocked (and 4 more threads) i7 and double (and faster) RAM?

Are both cases using the same driver?
As far as I know, the RX-580 is not that much of an upgrade to the RX-480, or perhaps the RX-470.

[humbug]

It is a bit sad to see AMD struggle so much with Vulkan, given that they practically pushed it into existence. Nvidia has really turned the table on them on that one.

Indeed it is unfortunate. When the khronos group started working on this in 2014 one of the key selling points was the thinner driver layers which would be a lot easier to maintain. The fact that AMD's Vulkan linux ecosystem would be in this state going into 2018 is something nobody could have predicted. AMD themselves marketed it as bringing high performance gaming to Linux for the first time.

However we cannot make a blanket statement that Nvidia has turned the tables on AMD with regard to Vulkan. That they have done that only on Linux, on windows AMD's Vulkan implementation surpasses Nvidia.

some people are conflating this issue with that of poor quality vulkan ports. Yes that is an issue that needs to be addressed, but what we are a talking about here is driver performance. I.e. for a given Vulkan application we expect to see the rx580 beat the gtx 1060, the vega56 beat the gtx 1070 etc... We are talking about performance relative to the competitor on a given app, not absolute performance which is what comes up when talking about badly optimized apps.

[azdaha]

RX580 is faster than 750ti in Linux gaming benchmarks, even RX560 is.

I am not sure how to explain that, except to share my own experience using CounterStrike:GO in Steam using ArchLinux and Ubuntu. I may see a larger number of FPS, but that doesn't translate into usability, quality, or performance overall. Also, I keep mentioning CS:GO, as I said before, because I'm assuming that Valve, as the (one of the?) games makers has more control over its code than of third-party games intended for Windows. That way, I'm assuming that the effects of the view that "games in Linux are worse than games in Windows" can be minimized, in order to focus on the GPU drivers as the bottle-neck; especially when hardware differences are more obvious to gauge and easier to use in comparisons.

[oooverclocker]

There must be some performance regression on Kernel 4.15.

I have the same impression for some months now. The benchmarks for AMD GPUs seem to having regressed. There should be some investigation to figure out if that's true.

On the other hand I'm still not completely satisfied with the performance of the Vulkan driver. It is amazing how far it got without official support from AMD but no matter which one will be the preferred one, it would help a lot to have the source code of the proprietary one.
In the current state it's possible to play well with these and as you can see in the Superposition and Mad Max results it just needs the smallest bit of optimization in the actual game or benchmark to get quite reasonable results with open source. Though you have to buy a quite strong GPU in comparison to other scenarios that include completely closed sources.

Surely, I can't tell whether you get the same image quality with Nvidia drivers but in many cases the FPS count indicates that there is much space for improvements in the open source Radeon drivers.
So even when Nvidia probably compresses data massively to achieve more FPS we should find out whether this is the cause.

Thanks Michael, for letting these benchmarks run regularly - that's very helpful.

[M@GOid]

Are both cases using the same driver?
As far as I know, the RX-580 is not that much of an upgrade to the RX-480, or perhaps the RX-470.

It's the Padoka's Git from 4 days ago, I assume the same Michael used. The RX 580 is faster, lets say 10 % at last. But Michel is using a top of the line Core i7, overclocked to 4.7 GHz, while mine is stuck at 3.9, with half and slower RAM. And of course 3770k have 2 cores and 4 thread less than the 8700k.

This is not the first benchmark with the 4.15 I noticed this. Lets hope the AMD boys can figure what is wrong before is too late.