X.Org ATI Driver Supports New Power Options

[signor_rossi]

@Vighy
There is a typo in the bugzilla link posted...

@elsie
I use it on a laptop, but mine hasn't a GPU fan.
glxgears (not a benchmark TM) gives me around 1400 fps on my X600 with ForceLowPowerMode true. The GPU stays fine cool... )

Direct Rendering and the xv overlay seem all to work.
KDE doesn't start up properly because plasma crashes, I haven't figured out yet why Xfce hangs...
EDIT: A clean config (I moved the .kde4 folder makes KDE start up fine...

[Niagra]

The DynamicPM option shouldn't affect performance. It only switches to a low power mode during idle (DPMS active). When you are using the system, it switches back to the default power mode.

At the moment, this doesn't work for me. With activated DynamicPM option, the card always uses the lowest possible mode. But perhaps this isn't the right place for bug reports and of course it's still development code.

But nevertheless thank you for bringing power management to R700 based cards on linux! This saves nearly 17W for me in regular system operation, since the Radeon 4670 is only actually powersaving when those powerplay features are activated.

Niagra

[bugmenot]

This is awesome! Thanks!

[Louise]

The DynamicPM option shouldn't affect performance. It only switches to a low power mode during idle (DPMS active). When you are using the system, it switches back to the default power mode.

You know a lot about this topic. Was it you who implemented it?

Does this mean that AMD have finished writing the power management specs and released them to some developers?

[elanthis]

You know a lot about this topic. Was it you who implemented it?

Does this mean that AMD have finished writing the power management specs and released them to some developers?

Alex works for AMD, so him having info means nothing about it being released publicly.

[bridgman]

Yeah, agd5f wrote the code.

I believe all the information used was already available, except for changing the # of PCIE lanes on 6xx/7xx (the registers moved).

This is the functionality we have been saying for months that someone could write

[Louise]

Alex works for AMD, so him having info means nothing about it being released publicly.

I see

Otherwise I'd say that would justify an article and a few comments from the AMD folks

[Louise]

Yeah, agd5f wrote the code.

I believe all the information used was already available, except for changing the # of PCIE lanes on 6xx/7xx (the registers moved).

It's pretty cool that these things can be changed without crashing the GPU!

How many pages would you say the final power management specs would be? 10 pages, 100 pages, 1000 pages ?

[bridgman]

It's pretty cool that these things can be changed without crashing the GPU

This code only adjusts the engine clock so far, not the memory clock; changing engine speed is pretty safe because you can block acceleration while changing the clock. Changing the memory clock reliably is more difficult. You also need some pretty complex logic to make sure that the memory clock doesn't get too low for your display bandwidth and latency requirements, or you start to get artifacts on the screen very quickly.

Remember that dynamic power management is governed by Schmidt's law -- "if you mess with something enough, it'll break".

Note that this code probably needs to be moved into the drm if we want to have full activity-based power management. The X driver doesn't know about 3D driver activity, just 2D and video, but the drm knows about all the acceleration activity.

How many pages would you say the final power management specs would be? 10 pages, 100 pages, 1000 pages ?

We'll probably try to identify a subset of the information which gives the biggest power savings without being too painful to code for on all the different GPUs and systems. As with 6xx 3D, we will probably release a mix of documentation and code.

Probably more than 10 pages but not a lot more, although if you count things like display bandwidth and latency calculations the numbers go up quickly.

[Xipeos]

Doesn't seem to work (yet?) for my 4870x2. I get a blank screen/hard lock with ForceLowPowerMode, and the problem seems to be

[drm] wait idle failed status : 0xA0003028 0x00000002
last message repeated 174 times

but I can't be sure due to the hard lock (messages may not be sync'ed to disk).

Can't reach buzilla either, so I can't report it.