Does Forcing "High" DPM Performance Help Out Your AMDGPU Performance?

[phoronix]

Phoronix: Does Forcing "High" DPM Performance Help Out Your AMDGPU Performance?

A premium patron recently asked about testing the open-source Radeon driver performance when testing the forced "high" dynamic power management state rather than the default "auto" mode. Here are some benchmarks...

http://www.phoronix.com/scan.php?page=news_item&px=AMDGPU-High-DPM-Quick

[philips]

It so happens that few days ago I ran similar tests on my setup (AMD HD 7970 with older kernel+drivers since is SteamOS), and yeah, not much difference:

http://openbenchmarking.org/result/1805284-FO-AMD7970PE38

http://openbenchmarking.org/result/1805282-FO-TALOSMODE51&obr_sor=y&obr_rro=y

[VikingGe]

wined3d can tends to benefit a lot when there's a lot of GPU-CPU synchronization going on and the GPU isn't fully loaded as a result, but other than that there *shouldn't* be much of a difference.

[tildearrow]

For me setting it to "high" fixes some stuttering in Dolphin. Most of the time I keep it in "low" though.

[M@GOid]

For me setting it to "high" fixes some stuttering in Dolphin. Most of the time I keep it in "low" though.

What type of rendering you're using: OpenGL 2, 3.1 or XRender?

[skeevy420]

For me setting it to "high" fixes some stuttering in Dolphin. Most of the time I keep it in "low" though.

Ditto. But in my case I assume it's because I'm using AMDGPU with an R7 260x -- it isn't officially supported. I also don't have working overclock features even though the settings are present. It just sucks that it seems like AMD doesn't care about the older GCN cards anymore. That said, Feral games that require Vulkan and GCN 1.3+ cards work just fine with my GCN 1.1 card.

@M@GOid -- OGL 3.1, but all give slight lag when on auto or low. 3.1 passes the "seems & feels better" tests better than the rest....that could also be the result of my script that bumps the fan speed while setting High DPM too.

My anecdotal "feels and seems while watching GPU temps" testing shows the problem to be the default fan speeds and not the default DPM settings. I have similar "feeling" results when setting the fan to run at 80% or faster. There's a good chance the default fan curved just sucks and that forces the cards to thermal clock lower than necessary to prevent them from burning up.

These are my results from about 20 minute of random gaming sessions last month when trying to find the cause of random lag during Mad Max:
Auto DPM + Auto Fan = 80 degree C gpu when playing Mad Max, with predictable lagging from effects and random slowdowns; 55 C idle.
High DPM + Auto Fan = same as above
High DPM + 80% Fan = 55 degrees C when playing Mad Max with no lags or slowdowns; 40 C idle.
High DPM + 100% Fan = 50 C during Mad Max and 35C idle.

It isn't thorough, but it does show that the fan speed is a major factor. I've never bothered testing Auto DPM and Manual fan speeds because I'm too lazy. I found a setting that works for all my games so that's what i stick with and haven't bothered testing since (High DPM + 80-100% fan speeds depending on how GPU intensive the game I'm playing is).

Its about the same for me on Windows or Linux in regards to the temps, DPM settings, fan curves, etc.

[aufkrawall]

Clocking is very aggressive here with RX 560, having 0 performance issues with dynamic clocking since 4.17 (unlike on Windows...).

[rycodge]

Michael - Are you going to post an article regarding new amdgpu undervolting on kernel 4.17? The only reason I've been hitting F5 on phoronix this week

[tildearrow]

What type of rendering you're using: OpenGL 2, 3.1 or XRender?

As in KWin? OpenGL 3.1 always, plus custom patches to eliminate input lag and stuttering.

[monte84]

There are some titles, like War Thunder, where my frame rates double by forcing high to DPM. For some reason with this game, GPU clocks speeds average just under 500MHz without high DPM. I think borderlands was another game that had a benefit as well, although not as high as War Thunder and I am thinking Company of Heroes 2, would have to Double Check on that. Seems it is more of an issue with the port perhaps, instead of a power management issue. Thanks for looking into it.