OpenGL vs. Vulkan On The AMD Ryzen 3

[bridgman]

I have some other OpenGL game benchmarks coming out this weekend on Ryzen 3 and so far it's looking like there may be a regression mostly affecting the Dota2/SS3/MadMax/Dow3 tests as at least for my other (games without Vulkan renders), the performance appears more normal in relation to GTX 1050/1060.

Hi Michael,

User frosth noted in the post above yours that this set of tests was run with ondemand governor rather than the performance governor used in most of your tests. I suspect the NVidia driver is getting around this somehow (fglrx used to do something similar) but the open drivers do not, and there is a good chance that is where the perf difference is coming from.

Is it possible you went back to performance governor for the later tests, which would explain why only this set of tests seemed wrong ?

[artivision]

Hi Michael,

User frosth noted in the post above yours that this set of tests was run with ondemand governor rather than the performance governor used in most of your tests. I suspect the NVidia driver is getting around this somehow (fglrx used to do something similar) but the open drivers do not, and there is a good chance that is where the perf difference is coming from.

Is it possible you went back to performance governor for the later tests, which would explain why only this set of tests seemed wrong ?

Power governors are still a problem in 2017?

[bridgman]

Power governors are still a problem in 2017?

Last I heard, yes. That said, my recollection was that it only took some fairly small tweaks to the on-demand governor's parameters to fix most of the performance gaps. I raised this at last year's LPC (urk, maybe it was 2015's LPC) but there was some more ambitious work happening that took precedence, but if it is still an issue then we could try to get some on-demand tweaks upstream to better handle GPU workloads.

[bridgman]

This is on Windows. On HD7850 and RX480 (the other person) Vulkan just crash. With other games from ID, too many problems.

OK, sounds like you are talking about individual problems then, right ?

I have seen enough published benchmarks on both cards (particularly RX480) to know that support is not "generally" broken.

[mphuZ]

..sounds like you are talking about individual problems then, right ?

Unfortunately, no. I have not been able to find out the reasons why the game crashes (I'm the curator of the topics of AMD drivers on overclockers.ru).
The situation is simple: someone starts, and some don't. Just yesterday people from this forum have complained about the game crashing (HD7950).
On steam forum you can also find a lot of complaints from owners of AMD when the game does not start or crashes with the Vulkan.
Even on the AMD forum have been complaints. 1, 2.

[bridgman]

Unfortunately, no. I have not been able to find out the reasons why the game crashes (I'm the curator of the topics of AMD drivers on overclockers.ru). The situation is simple: someone starts, and some don't. Just yesterday people from this forum have complained about the game crashing (HD7950). On steam forum you can also find a lot of complaints from owners of AMD when the game does not start or crashes with the Vulkan.

Even on the AMD forum have been complaints. 1, 2.

The first couple of posts I looked at in the overclockers.ru thread seemed to be people having problems downloading the SDK and trying to run without it. That's never going to work.

Will try to go through some more over the weekend though. Thanks for the links.

[mphuZ]

The first couple of posts I looked at in the overclockers.ru thread seemed to be people having problems downloading the SDK and trying to run without it. That's never going to work.
Will try to go through some more over the weekend though. Thanks for the links.

I once on the drivers 17.5.2 tried the demo. And I still have the logs of the error. According to user reports, the Vulkan in the Doom, then mend, then break.

[nadro]

My previous results were wrong (I though that results came from OpenGL, but it was from Vulkan, sry for that). After a fix to madmax script for phoronix-test-suite (I have steam games installed in custom directory) I was able to run benchmark for both OpenGL and Vulkan modes without problems. Fixed results are here:

http://openbenchmarking.org/result/1708199-NADR-NADROMA41

As you can see I use ondemand governor, so I still think that those low results from article are related to some bug in Mesa. Tomorrow I'll try to test Mad Max again, but with just 2 cores + SMT enabled (4 threads). With that settings performance of my Ryzen 5 1600X should be more similar to Ryzen 3 1200.

[frosth]

But my Polaris10 on Intel have no regression. So probably mesa is fine. Maybe kenrel vs cpu problem or any kind of shit happened. I don't have any Ryzen so i duno.

[timofonic]

If you look at other recent tests it seems to show a regression in Mesa, although that has not been confirmed. It is possible that performance is actually wildly lower on a 4-core without HT than a 4-core with HT but that seems less likely.



I don't understand. You were talking about Mesa earlier and using it on Windows, and that was the topic of all the questions you claim I did not answer (despite having answered a bunch of times).

The OpenCL driver has already been open sourced and work is proceeding on open sourcing the Vulkan driver.


Which cards ? AFAIK all the cards supported by AMDGPU-PRO are working with Vulkan.

bridgman Sorry, but I feel skeptic. I'll believe about the promised open-source Vulkan driver from proprietary AMD drivers when it becomes a reality, the same thing about DC/DAL...