I wouldn't say that the community is more involve with nvidia, there is lot of non paid people working on AMD GPU (and some not paid by AMD but by Red Hat). I think the big difference is that on nvidia you need powermanagement to have decent performance a reference nvidia bios clock gpu slower than amiga (bit of an exageration). And also most of the community working on AMD GPU prefer working on fancy stuff such as 3D and gladly let AMD or Red Hat do ground infrastructure such as power management.
Originally Posted by darkbasic
One of the thing that puzzle me is that in the end having AMD helping the community is 2 side coin. On one side it is very helpfull and help a lot, on the other there is lot less reverse engineering going one mostly because when we are lacking something you know that AMD is already helping on this but they can't release anything, so you kind of wait for everythings to clear up review. While on NVidia, well no help, means you can reverse what ever you want when ever you want. That said nothing forbid people in the community to reverse engineer what fglrx do for power management like for HDMI. It's just that part of the people that used to do reverse engineer on AMD can't anymore.
A good point, what I still don't understand is why power management isn't amd's priority: lack of proper pm hurts even cpu sales (who wants an apu which drains battery on a laptop?). My next laptop will probably be 100% intel just because of that.
I have 4,5 year old laptop with mobility hd 2400 and dynpm is still not usable. Because of that my new laptop will be intel cpu+intel gpu too.
Originally Posted by darkbasic
Btw fglrx just sucks on my laptop.
It's easy to say "XYZ should be a priority" (I imagine that means "a high priority") but the real question is "higher than what ?". Higher than lighting up the display ? Higher than making X and EXA work ? Higher than enough 3D acceleration to run a modern desktop environment ? Higher than providing low-level compute support for all the developers who want to work on GPU compute ? Higher than video acceleration (the perennial "top priority" in user surveys) ?
For the AMD-developed driver (fglrx) PM *is* a priority, and most tests suggest the PM on Linux is comparable to that on Windows with the exception of highly OS-specific areas like switchable graphics.
For the community-developed-with-help-from-AMD open source driver PM doesn't appear to be a short term priority (although good work is being done in foundation areas like adding multi-gpu and cross-render support to the graphics framework), and I'm as puzzled by that as you are. Maybe PM has too much of a reputation for being frustrating and fiddly, or it isn't as satisfying as work in more "visual" areas, or it's just the fact that kernel work in general is less approachable than userspace work... not sure.
Then again, in the user surveys video acceleration usually ranks higher than power management, so maybe it shouldn't be a surprise at all (but it *does* make it all the more surprising that people are complaining about us "wasting our time" working on video acceleration). We did commit to making a serious effort to get UVD acceleration into the open source drivers, and I felt it was important for us to do what we said.
Alex had already written and debugged initial PM code for both UMS and KMS, implemented sample code for more advanced functions like dynamic switching, and listed the tasks that still needed work. We figured that someone in the community would pick up the code from that point, but it didn't happen although a lot of great work is being done in other areas.
I suspect the core issue is that the current PM support is "good enough" for a lot of users and fglrx works for many of the folks who find it isn't good enough, so there are relatively few people left who are motivated to roll up their sleeves and take a shot at improving the PM in the open source driver. It's not like there aren't community developers working on the driver, they're just working on other parts of it.
All that said, I think there has been enough progress in other areas that PM has become one of the weaker areas again, and APUs in particular have a big gap between what *could* be done with publicly available information/code and what *has* been done, so it's probably a good time for us to do some more work in the PM area (and hopefully get to the point where new developers feel like getting involved) once we get through the current tasks.
Last edited by bridgman; 04-08-2012 at 06:56 PM.
I have an E350 Mini-ITX, running Xubuntu 12.04 beta and fglrx. I measured the power draw of the box, and for 95% of the time the difference between fglrx and open source video drivers is 2w. Really working hard the difference is hard to measure. Having said that I still prefer fglrx as it does save that small amount and that was the purpose of building the computer in the first place. Given that the box only pulls 24w most of the time, using fglrx is still about 8% saving. A high efficiency power supply did more than anything else to reduce power usage.
At the high end of energy sucking cross fire monsters the savings are no doubt much greater.
But, I do agree that adding power saving functions to the open source driver would be a good thing.
Just curious, what power setting are you running with the open source drivers for those tests (low/med/high/default/auto etc..) ?
I did not realise that the settings can be altered. So, the answer would be default. I take it the settings are somewhere in the power profiles.
I will have a play and see what changes.
EDIT: I just changed the AI setting and dropped another watt. Now I will revert to Open Source and see what happens.
EDIT2: With open source radeon driver the general idle is 26w, with fglrx the reading is 24w and fglrx with AI on max it drops to 23w. For general web browsing those numbers do not change. So, now I need to see if I can change the open source settings.
Last edited by grege; 04-08-2012 at 11:41 PM.
There's some info about PM settings on the RadeonFeature page :
I just remembered that on the APUs the default setting is pretty low already so you may find that the power settings don't make much difference. Sorry, didn't think of that until just now. On all but the most recent discrete GPUs, however, default was usually at or close to full speed & power.
You are correct. I played with dynpm and various profiles and if the were any changes my watt meter did not measure them. I have reinstalled fglrx and now I am back to 23w as the idle and normal desktop power draw. Full screen flash video can push it to near 32w.
Originally Posted by bridgman
Last edited by grege; 04-09-2012 at 12:38 AM.
Sounds like Radeon PM would have been a good task for gsoc.