AMDGPU With PowerPlay Compared To AMD's Catalyst Linux Driver

[mannerov]

380 looks really great on amdgpu!
Really tempting to get one!

I have a 4GB one as well: the msi gaming one.
Fans shut down when not playing, and perf is very good.

Don't take a 2GB one, new games will eat more and more memory, for example for Ark I get 2.1 GB used.

[jf33]

I don't understand why OpenGL is still at 3.0 (3.0 Mesa 11.1.0-devel (padoka PPA) Gallium 0.4) for Stock and PowerPlay? For RadeonSI it should be 4.1, isn't it?

It is at 4.1. Type
glxinfo | grep core

[darkbasic]

What do you mean full potential? Just with newer Mesa Git than earlier this month Git? Aside from that it was at full potential with PowerPlay, etc. I didn't re-test new Mesa Git for this article as was using the previous article's results as a base as I didn't have time / machine handy for running some fresh AMDGPU w/ patched PowerPlay + Mesa 11.2-devel + LLVM 3.8 SVN + DRI3.

With patches, see bridgman's post

[legume]

Kano "I would like to see HEVC Main 10 benchmarks. Especially with the AMD Fury and mesa." (quote button not working for me) Maybe just the output of - vdpauinfo | grep HEVC would be quicker. Michael?

[ernstp]

Guessing that "full potential" means including the patches Alex posted here after the previous set of tests :

http://www.phoronix.com/forums/forum...630#post834630



There's a v2 of the mesa patch in Mesa recently (~12 days ago) and cc'ed to stable. The kernel patch is in Linus's tree now, shows as 11 days ago but it might have actually been pulled in more recently.

Michael knows about the Mesa patches, mentioned in the article...



... but I think the kernel patch is needed as well, so in a perfect world the "PowerPlay" numbers would have used latest git Mesa & kernel. That said, there's always going to be something that gets faster while you're running the last set of tests.

Anyways, thanks for adding the Catalyst results !

I'm updating and compiling the powerplay branch myself and that fix only gave a few percent of increase. I'm getting the same results as Michael, something is busted in the Fiji case...

[nadro]

boxie
I also tried to run kernel with powerplay shared by Michael, but it doesn't work (crash at startup) for me. I have Gnome Ubuntu 15.10.

[przemoli]

You should!

I got mine with 4-GB of VRAM for 200 EUR.
And with ~3.5 Tflops of single-precision performance, it's about twice as fast when compared with the PS4!

IMHO, the best GPU one can buy for ~200 EUR right now...

You need apps that use modern OpenGL (4.x) and OpenCL (2.x) to actually approach API sophistication of PS4. Also PS4 unit is what heavily inspired of GCN 1.1 So its hardware is not that bad. And its APU so zero copy and all those goodies...

Its still better hw, and that's why I will by it if my notebook wont wait till AI gpus/apus with self-destructing....

[bridgman]

I'm updating and compiling the powerplay branch myself and that fix only gave a few percent of increase. I'm getting the same results as Michael, something is busted in the Fiji case...

When you say "powerplay branch" do you mean git master for mesa & kernel or something else ?

[Melcar]

Man, I really want a R9 380 now, but was planning to get a 2k monitor this holiday. Decision, decisions.

[ernstp]

When you say "powerplay branch" do you mean git master for mesa & kernel or something else ?

I mean I'm pulling the kernel from freedesktop and building, http://cgit.freedesktop.org/~agd5f/l...dgpu-powerplay
Mesa and LLVM is from Padoka PPA.
I test every commit you make. :-)

I'm guessing that performance would be good if only all the functions of the driver were working properly with powerplay, which they seem to not do...
Opened a bug report on it: https://bugs.freedesktop.org/show_bug.cgi?id=92996