AMD's Radeon Gallium3D Starts Posing A Threat To Catalyst

[bridgman]

Still really interesting if there any chance for any improvements for 6950.

bridgman, can you comment on this please?

I'm sure HD69xx performance can be brought in line with HD68xx (ie faster rather than slower), but I don't know the reason for the performance delta.

What I can say is that AFAIK there aren't any secrets being held back which would cause a performance difference, so presumably it's just a matter of driver/compiler optimization ie a matter of time. It would be interesting to see if the same performance delta exists with the VLIW4 APUs (Trinity and Richland) relative to the VLIW5 Llano.

[dffx]

AFAIK there aren't any secrets being held back which would cause a performance difference, so presumably it's just a matter of driver/compiler optimization ie a matter of time.

Out of a mix of curiosity and ignorance, is this the same case with radeonsi, that we'll just see performance improvements as we get progressive optimizations? Or are there missing bits for RadeonSI that makes it more difficult to improve on performance?

[smitty3268]

I actually play Dota2 and L4D2 on my 7970M using the latest radeonsi git pulls with no problem.

Marek just posted patches for UBO support today. Seems like full GL 3.1 support probably isn't far off.

[Marc Driftmeyer]

And now do the same with a HD7970 or a R9 290X using games like Serious Sam 3, DOTA 2 or Left 4 Dead 2.

I think everyone knows that the driver for the 4000 to 6000 series cards is wonderful and that there's no real reason to not use it over Catalyst. The situation for newer cards (and thus for gamers who bought their AMD card after 2011 and want to switch to Linux) is still dire if you don't use Catalyst (and Catalyst is dire itself).

I'm so happy that I was able to replace Catalyst with the open source driver on my laptop which runs the latest Mesa git snapshots and 3.12-rc7 right now, but it also has a Mobility Radeon HD5650 - I don't think I would've done that if I had an actual up-to-date GPU.

So yeah ... keep up the good work AMD, but please, start focusing on newer GPUs as well. I don't think recommending a 6870 for playing games on Linux sounds that nice, especially if everyone else already shouts "use NVIDIA on Linux only".

Really? No reason to use Catalyst and openCL-ID over the FOSS driver? Hardly.

Run those two drivers using ANSYS 14, COMSOL 4.6, etc. Then talk.

[smitty3268]

Out of a mix of curiosity and ignorance, is this the same case with radeonsi, that we'll just see performance improvements as we get progressive optimizations? Or are there missing bits for RadeonSI that makes it more difficult to improve on performance?

Also, according to Anandtech, Hawaii (and presumably Bonaire) include a much improved hardware PM abilities compared to previous architectures. Is the OSS driver able to take advantage of that? Or is it treating them the same as the previous GCN 1.0 architecture cards?

[smitty3268]

Really? No reason to use Catalyst and openCL-ID over the FOSS driver? Hardly.

Run those two drivers using ANSYS 14, COMSOL 4.6, etc. Then talk.

I think everyone realizes that the OSS drivers OpenCL support is still immature. Most people are referring strictly to OpenGL support, since that's all most people ever use. And stability/other desktop uses, like UVD decoding, etc. of course.

That said, there have been a few recent changes that might show some improvements here - such as:

http://cgit.freedesktop.org/mesa/mes...fcb61fbe087af7 and http://lists.freedesktop.org/archive...er/047264.html

Edit: does ANSYS even run on Catalyst? The blog post i saw mentioned something about only supporting Tesla cards, but maybe it's an old one.

[dffx]

Really? No reason to use Catalyst and openCL-ID over the FOSS driver? Hardly.

Run those two drivers using ANSYS 14, COMSOL 4.6, etc. Then talk.

I think everyone realizes that the OSS drivers OpenCL support is still immature. Most people are referring strictly to OpenGL support, since that's all most people ever use.

That said, there have been a few recent changes that might show some improvements here - such as:

http://cgit.freedesktop.org/mesa/mes...fcb61fbe087af7 and http://lists.freedesktop.org/archive...er/047264.html

Really, between Catalyst and radeon there are a lot of give-and-takes. For me it's that switchable/hybrid graphics are *much* better supported in the OSS stack. Desktop users might find Catalyst to suit their needs more for the time being.

[Alejandro Nova]

It would be nice to see where the APU performance is at. I have a laptop with an E-350 and abandoning Catalyst seems to be a more interesting proposition every day -- especially if/when the VDPAU Hi10P support gets merged.

I'm using an E-450 with the Radeon free software stack, and the only thing I'll tell you is: why are you still waiting? On the blob: support for brightness keys is sloppy, there are random crashes on resuming, suspend is unreliable and freezes the computer, and XvBA works in a specially patched MPlayer and little else. None of those issues is present in the Radeon driver and VDPAU (yes, Radeon uses VDPAU) works everywhere.

If you use DPM, power consumption is similar between the blob and the free driver. Also, multiscreen support is super reliable and in kernel 3.12 you can switch off HDMI audio, to use less power.

[duby229]

Still really interesting if there any chance for any improvements for 6950.

bridgman, can you comment on this please?

I really think it's just a case of optimizing how instructions are packed into the shader cores. In GPU terms the compute units are called stream processors, so when you are running a game all the opengl stuff is compiled into units that can be executed on those processors. I think at this point it really is just a case of making sure that the shaders are doing as much work as they can on the stream processors. It's just a matter of optimizing right now.

This is a really good article if you are interested in the differences between VLIW4 and VLIW5.

http://semiaccurate.com/2010/12/14/look-amds-new-cayman6900-architecture/

[bridgman]

Out of a mix of curiosity and ignorance, is this the same case with radeonsi, that we'll just see performance improvements as we get progressive optimizations? Or are there missing bits for RadeonSI that makes it more difficult to improve on performance?

Yep... for radeonsi the performance gaps are mostly known issues, HW optimizations like hyperZ and some tiling support (tiling may be done now) that need to be ported across from r600 and adapted to GCN hardware. There are also some new capabilities in GCN that are getting supported and queued up for review as we learn about them.