AMDGPU vs. Radeon DRM With GCN 1.0/1.1 GPUs On Linux 4.11, Mesa 17.1-dev

[phoronix]

Phoronix: AMDGPU vs. Radeon DRM With GCN 1.0/1.1 GPUs On Linux 4.11, Mesa 17.1-dev

With the upcoming Linux 4.11 kernel release as well as for the next cycle (Linux 4.12), the Radeon DRM driver remains the default for AMD GCN 1.0/1.1 GPUs while the newer AMDGPU DRM driver continues offering "experimental" support for these earlier generations of GCN GPUs. As it's been a while since our last Radeon vs. AMDGPU GCN 1.0/1.1 benchmarks, here are some fresh tests today with Linux 4.11 Git.

http://www.phoronix.com/vr.php?view=24488

[mibo]

Thank you for this test.
So, there is no regression with your 290 anymore?!

The 270/370 seems not to be ready for AMDGPU.
And, why did you test Tomb Raider with low settings? :-)

[Tomin]

The Tomb Raider performance was close to the same but appears to be CPU found in this scenario.

Typo.

[M@GOid]

I have installed Kubuntu 17.04 and noticed a performance regression in at last on game (Euro Truck Simulator 2) with the stock 4.10 kernel. Using DRM-NEXT 4.11.99 there are no problems.

I believe Michael did one benchmark these days that showed another performance regression with the 4.10 kernel.

[Namenlos]

You might want to mention that there is no UVD and no VCE support on amdgpu yet (at least on 1.0).

[LeJimster]

Didn't realise performance was worse on AMDGPU for my R9 270. Only switched over last month to test radv. Seemed pretty decent performance. Labelling it experimental is a shame as it works fine, very stable.

[geearf]

You might want to mention that there is no UVD and no VCE support on amdgpu yet (at least on 1.0).

Oh I had no idea :/

[boxie]

Seems that the team is doing really well with porting over the old devices - well done!

[schmidtbag]

So Michael, it seems you can start including the 290 in your tests again. None of those benchmarks seemed to show any major performance problems.

What else is left that's keeping these GPUs as experimental?

[smitty3268]

So Michael, it seems you can start including the 290 in your tests again. None of those benchmarks seemed to show any major performance problems.

What else is left that's keeping these GPUs as experimental?

Just look at the article, 2 of the 4 GPUs tested show performance problems and 2 of them had problems with output as well requiring Michael to treat them carefully. Plus 1 test caused crashes.

I'd say that's 3 reasons right there to keep it marked experimental.