Radeon Driver Power Management Has Room For Improvement
Phoronix: Radeon Driver Power Management Has Room For Improvement
Today on Phoronix are not only new benchmarks of the proprietary Catalyst graphics driver compared to the open-source Radeon Linux driver alternative when looking at the OpenGL frame-rates, but also metrics on a number of other fronts. In this article is a graphics driver comparison when looking at the system power consumption, GPU operating temperature, and CPU usage too. The results are quite interesting and not commonly looked at on Phoronix or by users.
What I found the most interesting in these graphs is the idle power management.
I'd always thought that my graphics card (5870) was using much more power with radeon than it would with catalyst (never tried it though).
According to this, that isn't necessarily the case.
Thanks for the benchmarks.
Without catalyst temperatures I do agree.
Originally Posted by Welsh Dwarf
It seems every card is a different story. Dynpm did not work for your cards, but it works with my HD3870. Low profile did not work for HD4870, but it works flawlessly for the others.
Anyway it seems power management is in good shape, we just need more testing and a better dynpm which can be enabled by default.
The other thing I find bizzar is that according to the DRM, my card doesn't have a low profile (or at least, it's the same as mid) but according to the benchmarks (comparing voltage) it's the mid profile that my card is lacking.
Being on a laptop, these things have increased importance.
Thanks Michael, very interesting. Seems the biggest conclusion is that dynpm doesn't work, period.
DynPM does work, but not well...
I have a HD 5400 card, and with power management set to DynPM, it switches between High profile en Mid profile.
Problems (black screens) occur with refreshes of the screen on lower (low, mid) profiles because of sync problems.
High profile makes the card hot, and makes the fan spin loudly even when your compositing desktop is not doing anything.
I think that dynpm is simply extremely slow at switching, probably because it's hard to do all the work during blanking periods or something.
I've had the opposite experience. If I run an idle system with dynpm, it consumes little stuff, but games are really slow. It's like running them with the low profile. It takes quite a while before it adjusts to the load, it seems.
I did some tests when this stuff was new, and it took a few minutes of running gears or something for the frequencies to ramp up, and a few minutes of idling for them to drop to idle levels again. If Michael did the switch from default profile to dynpm and tested immediately, then this would explain why he was getting basically the default profile performance.
I agree that dynpm could use lots of tweaking, but apparently, this is far from easy.
I had also the same experience of slow switching between profiles with dynpm, and sync in blancing period is off when in low profile.
Thanks for the article. As an owner of a legacy card (X1600 Mobility, upgrading means upgrading the laptop which is not an option atm), I wish that a similar test could be made to legacy cards as well (you could use Debian Lenny or Ubuntu 8.04 as a base for Catalyst and a recent distro for Radeon, these are the real alternatives for us financially restricted ).
Personally I am interested only in power consumption, but it might be interesting to see the different framerates the Catalyst gives with its different power profiles, too.
I am inclined to believe the differences are smaller for R300(-R500) cards, but I am still playing it safe and using a "legacy" distribution, until convinced otherwise (for example by a Phoronix article).
If I had any free time of if the testing could be fully automated (I have a plenty of computing time available), I would be glad to contribute (HP Compaq NC8430).
The article has a funny title. "Room for improvement". lol. Seeing the performance and power consumption charts, the only conclusion is: the open drivers totally suck ass.