Radeon RX Vega On Linux: High-Performance GPUs & Open-Source No Longer An Oxymoron

[profoundWHALE]

But to say the AMD card is the same technical achievement as the 1000 series is absurd.

This is what I said:

Besides that, this card is already great at games, but is fantastic for machine learning, mining, and I'm assuming other OpenCL workloads. It's already a great value for the performance and its brand new. The gaming performance can only get better.

1) It is a great value as long as you don't buy one of the price-gouged ones. On a new architecture, there are always growing pains. We saw that with Ryzen and memory stability as well as some overclocking settings that didn't work right. It's not like this is an AMD only thing.

2) Linux drivers (while getting very good) are still in second class. Windows is where the first class performance is right now, and even they've got their share of growing pains. For day 1 performance, on Linux, that is this good and without having to install blobs is simply incredible.

3) The size of AMD is a dwarf compared to the money behind Nvidia and Intel. I like competition, but I also like cheering for the underdog.

[Michael]

Thanks to Michael for the fantastic setup guide! My Vega 64 works out-of-the-box with the open stack and performance is excellent.

The screen, however, is very dark and desaturated. I thought this could be 30bit colour enabling itself, but specifying 24 bit depth in xorg.conf makes no difference. Interestingly, the darkness issue is also present in the console & plymouth -- Ubuntu's purple plymouth background is practically black.

I know it's early days, but has anyone else seen this?

The issue is present on both HDMI and DP. Windows on the same machine looks perfect.

Are you on 4.12 or -next? With the amd-staging-4.12 I didn't have this issue but did encounter it when trying -next.

[pal666]

That was unnecessary, two people already answered that question

but taking into account following

I couldn't read all posts because I was in hurry

it certainly increased chances

[leipero]

but taking into account following
it certainly increased chances

So it means you criticize me while making same "mistakes" as I do. Sometimes you simply can't read every post, especially when you are "late to the party" and there's 5+ pages already.

[anoother]

Are you on 4.12 or -next? With the amd-staging-4.12 I didn't have this issue but did encounter it when trying -next.

agd5f amd-staging-4.12 branch...

I'll try the kernel you compiled and see if that makes a difference.

[anoother]

Are you on 4.12 or -next? With the amd-staging-4.12 I didn't have this issue but did encounter it when trying -next.

Just tried your 'vegaphx' kernel -- it works

I guess there has been a regression in the amd-staging-4.12 branch since your build. Can you let me know the commit hash you built from, so I can build with my own config and maybe even try to track down the issue?

[artivision]

This is what I said:



1) It is a great value as long as you don't buy one of the price-gouged ones. On a new architecture, there are always growing pains. We saw that with Ryzen and memory stability as well as some overclocking settings that didn't work right. It's not like this is an AMD only thing.

2) Linux drivers (while getting very good) are still in second class. Windows is where the first class performance is right now, and even they've got their share of growing pains. For day 1 performance, on Linux, that is this good and without having to install blobs is simply incredible.

3) The size of AMD is a dwarf compared to the money behind Nvidia and Intel. I like competition, but I also like cheering for the underdog.

Correction: open source drivers are faster than windowz closed source drivers. The small decrease in performance from D3D11 to OGL is because titles are ported and they do not use native OGL benefits. There is no chance for native OGL titles to run faster on windows nor for Amd and nor for Nvidia. There is also no chance for native D3D11 titles to beat native OGL titles if just some Azdo extensions are used. The same is true for low level apis to (or will be shortly). People I know say that a D3D11 state tracker will be faster on Linux even than Nvidia on windowz.

[Michael]

Just tried your 'vegaphx' kernel -- it works

I guess there has been a regression in the amd-staging-4.12 branch since your buils. Can you let me know the commit hash you built from, so I can build with my own config and maybe even try to track down the issue?

Don't remember the commit, was from last ~Thursday, basically a few hours before that post on Phoronix with vegaphx download link.

[Karbowiak]

AFAIK the SR-IOV feature was not enabled on the consumer cards, just server cards.

This is such a shame really, even if limited to just 2 instances, it would have been amazing.

I wonder who we'd have to blow to get it /o\

[bridgman]

AFAIK the problem is that it is all or nothing, no ability to say "only enable a couple of VF's".

I have a couple of requests in to see if a cut-down version is possible though.