Radeon DRM Power Management Moves Along

[BlackStar]

The tables look fine. How you interpret them and expose them varies based on the implementation. The kms code is pretty basic at the moment.

Are you referring to these tables or (I suspect) something else entirely?

[RealNC]

Can't you get info about this (how does Catalyst interpret this) from AMD or something since last time I looked they are supporting this OSS driver effort?

[agd5f]

We don't print all the info from the tables in the kernel log. We should probably interpret the tables slightly differently on mobile vs. desktop or check for additional attributes, etc.

[Zajec]

The tables look fine. How you interpret them and expose them varies based on the implementation. The kms code is pretty basic at the moment.

I don't see consistency in tables, at least where there is not Battery state. Then you can find lowest mode in Performance of Default, any interpretation for that?

[agd5f]

The tables look fine. How you interpret them and expose them varies based on the implementation. The kms code is pretty basic at the moment.

[BlackStar]

Hmm, strange why your card does it and none of the GDDR3 ones listed in the review. A board difference or a Windows/Linux difference?

Different BIOSes come with different defaults, but in this case I think the table is incomplete or mistaken.

For instance, my original 4850 BIOS contained three power states (2d, low 3d, high 3d), with 2d clocked at 750MHz and high 3d clocked at 993MHz (I think low 3d was 750MHz, too, but it's been while and I may be wrong). Later revisions of the card came with 2d at 500MHz, which prompted me to modify the BIOS manually.

I don't think that's a windows/linux difference either, because my power meter gives the same indication on both OSes.

[monraaf]

I don't think that's true. As far as I know, only the GDDR5-based HD4xx0 cards do not downclock memory, because GDDR5 memclock changes require some kind of recalibration that produces a short - but visible - flicker on screen.

Here's a review (in German) of the new HD 5670, if you scroll down a little bit you see a table with different cards, clocks and voltages.

http://ht4u.net/reviews/2010/sapphire_radeon_hd5670/index4.php

As you can see all the HD 4xxx cards in the list have the same idle as load memory clocks.

My GDDR-based HD4850 downclocks memory correctly out of the box, which makes a visible difference in power consumption: with fglrx this card consumes around 40-45W on idle in my system (which can fall another 4-5W using a BIOS with reduced 2d memory clocks). With radeonhd (no power management), the same card consumes around 65-70W on idle - ouch!

Hmm, strange why your card does it and none of the GDDR3 ones listed in the review. A board difference or a Windows/Linux difference?

[BlackStar]

AFAIK most of the HD 4xxx cards do not downclock the memory.

I don't think that's true. As far as I know, only the GDDR5-based HD4xx0 cards do not downclock memory, because GDDR5 memclock changes require some kind of recalibration that produces a short - but visible - flicker on screen.

My GDDR3-based HD4850 downclocks memory correctly out of the box, which makes a visible difference in power consumption: with fglrx this card consumes around 40-45W on idle in my system (which can fall another 4-5W using a BIOS with reduced 2d memory clocks). With radeonhd (no power management), the same card consumes around 65-70W on idle - ouch!

[monraaf]

On both Windows and Linux, when Catalyst downclocks my card (HD4870) for "non-3D mode" it only does that for the GPU. The memory clocks stay as they were.

On Windows this can be seen in the "Overdrive" section of the CCC. On Linux you have to use fglrx's command line tool to get the clocks.

AFAIK most of the HD 4xxx cards do not downclock the memory.

[RealNC]

On both Windows and Linux, when Catalyst downclocks my card (HD4870) for "non-3D mode" it only does that for the GPU. The memory clocks stay as they were.

On Windows this can be seen in the "Overdrive" section of the CCC. On Linux you have to use fglrx's command line tool to get the clocks.