Previewing The Radeon DPM Performance On Linux 3.11

[liam]

Maybe that's why the article title started with "Preview".

Also, looks like Urban Terror is bottlenecked somewhere other than the GPU (or at least the GPU/VRAM clocks didn't make much of a difference).

That was the most interesting of the lot (besides seeing the general degree of improvement). At the 1080 there was only around a 15% difference, but at 1600 it jumped to better than 70%, with the dpm'd driver experience only a slight fps change between resolution.
So the clock speed made the biggest difference when it comes to larger textures. I'm guessing this is down purely to the increased memory speed?

[synaptix]

Gonna try this out later on my HD4200 IGP. Setting power profiles never seemed to stick and previous attempts at dpm was failure as well.

[bridgman]

The performance difference shouldn't be nearly like that on dedicated cards. But it's not supposed to be, the DPM code is for energy consumption and thermal management.

It's going to depend on the card. Guessing we'll see similar improvements for the GCN cards, since they also default to pretty low clocks, and *some* of the earlier discrete GPUs, particularly mobile parts.

[bridgman]

PS : A special compliment goes to bridgman for concentrating on development and not on tweaking the old PM code.

Actually that credit should go to Alex (and probably Dave too). I wanted us to tweak the old PM code a bit more to improve things while we were trying to get approval for DPM release, but I kinda got outvoted.

Taking the flak from all of us must have been a bit harsh

Yeah, guess I still get credit for taking the flak

[project_phelius]

I've been recommending AMD products

I've been recommending AMD products mainly because their OSS support, it feels great to be proven right.

You were recommending products with no proof?
You're not a very good friend!

j/k

[DanL]

Power consumption graph is not interesting because DPM works very same way as Catalyst. Comparison to Catalyst is actually irrelevant.

Well, it's supposed to work the same way in theory, but it would be nice to confirm that it works that way in practice.

[GreatEmerald]

As a question, is there any major difference in performance between the new DPM code and the old static PM on high? And if so, will this apply to all the supported cards?

From what I can see, this was tested on an APU, which have had the issue where static PM didn't do anything (the clocks were always set to default). The performance difference shouldn't be nearly like that on dedicated cards. But it's not supposed to be, the DPM code is for energy consumption and thermal management.

[Hamish Wilson]

As a question, is there any major difference in performance between the new DPM code and the old static PM on high? And if so, will this apply to all the supported cards?

[brosis]

Do I understand it right? The open source drivers will be in a few cases faster the propertiary ones?

Correct, but in some cases slower.

Really? This must be the most noninteresting benchmark I've ever seen!

Everybody already knew, that higher clocks get higher performance. Where is the power consumption graph? Where is the comparison to Catalyst? Did you set the old Power-management to powersave or performance (what clocks were chosen)?

Ha, its one of the most epic benchmarks in history. The performance is result of DPM+Vadim patch+Marek+Deutcher and many others! Its by far not only clocks.

Power consumption graph is not interesting because DPM works very same way as Catalyst.
Comparison to Catalyst is actually irrelevant. Performance is roughly same. But its opensource driver, that's the prime advantage. Support till hardware is dead, support for other operating systems, full integration, full security, no kernel tainting. All those not possible with catalyst, ever.

Powersave is not working right now, so it was performance.

[Serafean]

This can never be said enough: Thank you Bridgman,Tom, Alex, and anyone else involved in the OSS effort.

I've been recommending AMD products mainly because their OSS support, it feels great to be proven right. Going to test DPM on a laptop right now.


Serafean

PS : A special compliment goes to bridgman for concentrating on development and not on tweaking the old PM code. Taking the flak from all of us must have been a bit harsh