How AMD's Open-Source GPU Driver Performance Evolved In 2015: Big Wins

[phoronix]

Phoronix: How AMD's Open-Source GPU Driver Performance Evolved In 2015: Big Wins

One of the most requested end-of-year articles by Phoronix Premium readers was to compare the performance of AMD graphics cards at the end of 2014 on the open-source driver compared to how they compete these days with the very latest open-source driver code. Well, as one of our Christmas 2015 articles, here's this comparison with a few different Radeon GPUs.

http://www.phoronix.com/vr.php?view=22619

[F1esDgSdUTYpm0iy]

Bit of a shame there seems to be some regression going on for the 6870 though. Otherwise, looking OK. Lot of room for improvement still but with recent developments, such as the move to AMDGPU, I'm quite confident the open source driver stack will eventually prove to be a reliable and performant substitute. I'm already sold on it, had really nasty window corruption in Google Chrome when using the proprietary driver.

[eydee]

The difference is a bit smaller than I expected. At this rate catching up to the now legacy (non-crimson) catalyst build is still years away.

[geearf]

The difference is a bit smaller than I expected. At this rate catching up to the now legacy (non-crimson) catalyst build is still years away.

Not necessarily, if most of the work was on features and not performance, once the driver gets feature complete, more work can be done on the performance side.

[mao_dze_dun]

Not necessarily, if most of the work was on features and not performance, once the driver gets feature complete, more work can be done on the performance side.

They still have a long way to OpenGL 4.5 and Vulkan is due in a few months. Then they'll have to catch up to it. That's the vicious circle of the open source drivers - they're always a few years behind and performance is always last concern. Let's face it - the only future AMD gaming on Linux has is AMD half-assing its way till all GCN 1.1 and older cards are no longer supported and then maybe finally providing a half decent AMDGPU based driver for newer models.

[bridgman]

Can I remind everyone that the goals of the open source stack were to (a) cover all of the common usage scenarios for typical consumer Linux users, (b) to do so with at least 60-70% of Catalyst performance while remaining "clean" and maintainable, (c) be open source and upstream, so that they could naturally flow into distributions and (d) serve as a basis for supporting other OSes and non-PC systems. Other than a recent crop of games bumping the GL level requirement up a bit faster than expected (which is great news overall because it means the arrival of advanced games is ramping up) I think it would be fair to say those goals have been and are being met.

It seems pretty clear that open source shader compiler work over the last few years, Vadim's sb work and the LLVM work done by Tom and others, is going to let us skip past that original 60-70% performance estimate but that doesn't mean that the open stack goals should now be redefined so that "success" is determined primarily on whether the open stack outperforms Catalyst.

There are clearly some bottlenecks in both drivers on some recent games, where we hit either a CPU limit or a synchronization/parallelization limit sooner than the NVidia drivers (you can see the NVidia drivers hit a similar wall at any res below 4K, just at a higher frame rate), but my guess is that when those are addressed you will see performance on both drivers jump ahead. Moving to a common kernel driver will help there, but I expect that collaboration between the driver teams and an increased focus on Linux gaming will help even more.

[Kano]

70% of catalyst, that would be like getting midrage performance for highend price. Well for Fiji even Catalyst feels slow compared to Tonga. The benchmarks with CPU limited results are really boring, i always hope that Michael would OC the CPU or use a different one. But that problem is system-wide for AMD - that's why AMD introduced Mantle. Time will tell if Vulkan will be good for AMD hardware.

[smitty3268]

There are clearly some bottlenecks in both drivers on some recent games, where we hit either a CPU limit or a synchronization/parallelization limit sooner than the NVidia drivers (you can see the NVidia drivers hit a similar wall at any res below 4K, just at a higher frame rate), but my guess is that when those are addressed you will see performance on both drivers jump ahead.

So are you saying that IS going to be addressed? Or "if" it's addressed?

I guess I'd given up hope for getting that fixed by AMD with Vulkan coming out and presumably taking the some of the focus away from getting the best possible OpenGL performance.

[artivision]

Can I remind everyone that the goals of the open source stack were to (a) cover all of the common usage scenarios for typical consumer Linux users, (b) to do so with at least 60-70% of Catalyst performance while remaining "clean" and maintainable, (c) be open source and upstream, so that they could naturally flow into distributions and (d) serve as a basis for supporting other OSes and non-PC systems. Other than a recent crop of games bumping the GL level requirement up a bit faster than expected (which is great news overall because it means the arrival of advanced games is ramping up) I think it would be fair to say those goals have been and are being met.

It seems pretty clear that open source shader compiler work over the last few years, Vadim's sb work and the LLVM work done by Tom and others, is going to let us skip past that original 60-70% performance estimate but that doesn't mean that the open stack goals should now be redefined so that "success" is determined primarily on whether the open stack outperforms Catalyst.

There are clearly some bottlenecks in both drivers on some recent games, where we hit either a CPU limit or a synchronization/parallelization limit sooner than the NVidia drivers (you can see the NVidia drivers hit a similar wall at any res below 4K, just at a higher frame rate), but my guess is that when those are addressed you will see performance on both drivers jump ahead. Moving to a common kernel driver will help there, but I expect that collaboration between the driver teams and an increased focus on Linux gaming will help even more.

I remember well that it was 80% when we were on Shader_Model_4, and never about CPU overhead.