Missing Functionality From The Linux Graphics Drivers

[crazycheese]

I think we should at least make a valid list of what is present and what is not.
This would at least make this thread productive.

[crazycheese]

The preferred way to save power (for GPUs and CPUs) is ondemand / adaptive (i.e maximize idle time at low power by doing any work as fast as possible).

There is powersave governor for CPU. I repeat - there are GPUs that have a lot of processing power and default idle state by nvidia driver is just an overkill.
As example - my 260sp216 uses 300Mhz for GPU and memory with kwin composite on fullhd.
By manual downclocking I found out that even 150 Mhz was enough power.
In fact, more modern fermi chips downclock exactly there.

1) An option to persistently keep the clocks minimal MUST exist. There IS a selectable option to enforce maximum performance - reason left for card user. He may be testing new heat sink, burning the chip for stability or trying to disable the power management altogether for whatever reason. It ain't windows where "some policy" exists and you are forced to it, just because ms decided so.

2) The should be a possiblity to keep settings persistent. Currently theydisappear on xorg restart, relogin or reboot. The card user has to go through license scrolling and clock typing once again.

But KDE does not use openvg (and most distros don't enable it) so it does not help you much.

I heard gallium uses acceleration with KDE4.
At least nvidia is much slower in kde4 without bigger pixbuf xorg manual setting.
If anyone knows something on this topic, please comment.

To get both overclocking and fanspeed you'd want to set it to something like "5" i.e

Option		"Coolbits"	"5"

It is there

Another hidden useful option, why does nvidia hide it anyway? Is it developer-only?
Apart from behaving similar to overclocking in terms of resetting on every logout, forgetting the values and that I accepted EULA, it is nice to thing to have around.

[Dragoran]

There is powersave governor for CPU. I repeat - there are GPUs that have a lot of processing power and default idle state by nvidia driver is just an overkill.
As example - my 260sp216 uses 300Mhz for GPU and memory with kwin composite on fullhd.
By manual downclocking I found out that even 150 Mhz was enough power.
In fact, more modern fermi chips downclock exactly there.

1) An option to persistently keep the clocks minimal MUST exist. There IS a selectable option to enforce maximum performance - reason left for card user. He may be testing new heat sink, burning the chip for stability or trying to disable the power management altogether for whatever reason. It ain't windows where "some policy" exists and you are forced to it, just because ms decided so.

You missed the point, you are not saving power by permanently lowering the clocks.

I heard gallium uses acceleration with KDE4.
At least nvidia is much slower in kde4 without bigger pixbuf xorg manual setting.
If anyone knows something on this topic, please comment.

You heard wrong ... that is simply different ways the accelerating architectures work. Being slower is not "missing functionality" but a bug.

Another hidden useful option, why does nvidia hide it anyway? Is it developer-only?

No it is potentially dangerous, so they do that to make sure you do no lose the warranty by doing stupid actions. Also it is documented in the readme.

[yotambien]

You missed the point, you are not saving power by permanently lowering the clocks.

Why not?

..............

[Dragoran]

Why not?

..............

Read this ... I tried to explain this in my earlier post but apparently failed.

[yotambien]

I know that stuff, and it's a made up example. Notice how it says that it "tends to hold", and "it's generally better". I wouldn't take it as a divine law. A real life example showing how delicate a balance is to be played can be found here:

http://www.ibm.com/developerworks/linux/library/l-cpufreq-3/

For what is worth, and if judging from the fan noise is to be trusted at all, I am pretty sure I save power by locking the CPU to a lower frequency while not doing any demanding tasks.

[crazycheese]

Read this ... I tried to explain this in my earlier post but apparently failed.

Apparently you failed to understand me.
In case, the user is totally unaware of anything, the best option is, yes, to use adaptive clocks with lowest minimal clocks possible to the task.

Lets apply this to current situation with nvidia:
1) Speaking about IQ of driver


Uploaded with ImageShack.us



Uploaded with ImageShack.us



Uploaded with ImageShack.us

So, the people with manual transmission are not that dumb.

2) Not every user is "not knowledgeable". Absence of forcing minimal clocks adds very well to existing nvidia tradition to use their hardware in their own ways. Like forgetting OC and fan settings, forgeting he read the EULA already and sorts of stuff. Or they continue microsoft tradition? I don't know.

3) I'm very well aware of the fact, that throttling performance will make chip perform lower and will make it last ntimes more time to end the task, consuming same amount of power in the process(plus some %), compared to max performance. UNLESS the chip is really doing nothing important. So it is better the hardware adapts to load and tries somehow to "manage itself". But how is hardware or software is going to UNDERSTAND doubleV-tee-ef is usefull and what not? Through the driver... somehow. no?
We have seen many cases of DVD drives spinning like crazy in the past, preventing buyer to watch film and have seen a utility to force them at low speed.
Some manufacturers learned from the process, but ultimately it did not make those utilities useless.

4) When I have to encode some video for my mobile phone to watch later, I manually set cpu governor to performance. Because, and I'm sure you know it too, performance mode is much more "performant" than "ondemand", because there is no logic trying to watch over something. This coverts into the fact, that machine without adaptive logic gets thing done faster and then just gets a command from the script to go into adaptive again. So its FINE having adaptive, but it will rarely get things done best and very rarely will be really adaptive.

So the moral of the story is as easy and old as the world itself:
Allow people to use their brain.
Do not get into the way - do not conflict, but complement.
Allow humans to take the decision themselves when they want to take it.

Do not put every one in one group of "unknowledgeable", do not make idiocratic society.

But apparently, microsoft has been following this policy so long, it might have converted itself into own case in the process. And nvidia - follows?

[johnc]

So the moral of the story is as easy and old as the world itself:
Allow people to use their brain.
Do not get into the way - do not conflict, but complement.
Allow humans to take the decision themselves when they want to take it.

Do not put every one in one group of "unknowledgeable", do not make idiocratic society.

But apparently, microsoft has been following this policy so long, it might have converted itself into own case in the process. And nvidia - follows?

I think you're being too harsh here. The point is that the driver does indeed expose these functions (though no overclocking on fermi). If nvidia is going to provide a piece of software to interact with these settings, clearly they have to use caution, because it is the vendor that is responsible for replacing damaged cards, not nvidia.

If you want software that allows greater flexibility with fan speeds and overclocking, ask your video card vendor to release a linux version of their software, or you are free to write your own software to get the job done according to your preferences.

[md1032]

The driver cannot allow forcing the minimum power level because then there would not be enough memory bandwidth to perform all of the graphics functions and drive the displays at the same. It may have to raise the clocks dynamically to cope, which is exactly what the adaptive clocking is for.

[crazycheese]

I think you're being too harsh here. The point is that the driver does indeed expose these functions (though no overclocking on fermi). If nvidia is going to provide a piece of software to interact with these settings, clearly they have to use caution, because it is the vendor that is responsible for replacing damaged cards, not nvidia.

If you want software that allows greater flexibility with fan speeds and overclocking, ask your video card vendor to release a linux version of their software, or you are free to write your own software to get the job done according to your preferences.

Maybe.. On windows, there was that russian guy, Unwinder, who wrote Rivatuner... eventually he was bought by MSI and now Rivatuner exists under name "MSI Afterburner".
But maybe not. I see all this as limitation of product, which hinders its value for enthusiasts or overall linux buyers. I agree that you should be warned... once.

The driver cannot allow forcing the minimum power level because then there would not be enough memory bandwidth to perform all of the graphics functions and drive the displays at the same. It may have to raise the clocks dynamically to cope, which is exactly what the adaptive clocking is for.

Nope, dualhead should just have its minimal base clocks limit rised. This detection could be done automatically.