Radeon Power Management Still An Incomplete Mess

[libv]

"Just release the code already, and arrange for anyone who tries to block it to find a new position where being an idiot is acceptable."

Thanks Dave!

That's nice. Now think back 5 years and wonder why ATI was more than keen to work with Dave than with us. We at SuSE were aiming too high (as in, reducing the relevance of fglrx) and we were asking too many nasty questions for ATIs liking (many of those ATI were simply incapable or unwilling to answer). We were getting blocked on so many sides by the end, not only by ATI, but also by supposed open source people. So I am still waiting to see that logic applied comprehensively, which in this case, includes the person quoted.

So stop complaining, Dave. You played a key role in reducing the set of requirements AMD had for ATI. Now go and sit on the blisters. (and before you try to complain; why don't you get back to me when more documentation than just the shader isa, released by the AMD gpgpu people, and not by the mess-formerly-known-as ATI, is being released again - as we at SuSE required)

[tstrunk]

And i don't think the corporate/large scale care about open source OpenCL. They'll probably just use the blob and get their job done. And even the average joe will install a blob because he want's VA, Games, PM etc.

They don't care about open source, true, but they care about hassle-free. Implementing OpenCL using Catalyst on a HPC architecture is a pain and you need to break some best practices to do it, like install a X-Server and make it permissive. I can only assume the open source driver would allow for a better solution.
Still of course, Performance is king. If an open source driver wouldn't attain performance parity with the closed source solution, it wouldn't be used also on HPC no matter the security holes.

On the other hand for office terminals the open source drivers are working great and hassle free.

[89c51]

People that just want to watch youtube videos probably aren't running Linux. The percentage of Linux users that want fast OpenCL is actually pretty large. Most large scale Linux users are not consumers.

I was (and probably Adarion) referring to the average joe linux user and not the corporate/large scale type.

And i don't think the corporate/large scale care about open source OpenCL. They'll probably just use the blob and get their job done. And even the average joe will install a blob because he want's VA, Games, PM etc.

[agd5f]

Probably you don't see people complaining about OpenCl because the computer already does what is supposed to do (period). I mean it won't matter much if a computation takes a bit longer to complete for the average user. On the PM and VA things are different. ie. People see their batteries draining much earlier and CPUs hit 100% when playing Youtube videos hence you get more complains.

People that just want to watch youtube videos probably aren't running Linux. The percentage of Linux users that want fast OpenCL is actually pretty large. Most large scale Linux users are not consumers.

[89c51]

At least here on Phoronix I rarely see anybody complain about missing OpenCL features but daily requests about power management and video accelearation issues (UVD/shaders). So maybe the AMD-ATI devs reading here should note to their superiors that these things and some 3d performance improvements seem to be a really pressing point among users. To give that more priority.

Probably you don't see people complaining about OpenCl because the computer already does what is supposed to do (period). I mean it won't matter much if a computation takes a bit longer to complete for the average user. On the PM and VA things are different. ie. People see their batteries draining much earlier and CPUs hit 100% when playing Youtube videos hence you get more complains.

[Adarion]

Oh. Well. From my personal pov:

More than a year ago my old Sapphire HD 3870 (Ultimate ed.) was doing during idle periods
about 75 W (whole computer w. 4850e CPU, HDDs, some USB, PCI cards, ... without screen) "from the wall" with (Gentoo) Linux (dunno which kernel) and fglrx.
80 W in Windows (NT 5.1 aka XP) w. Catalyst
85 W free driver stack (not sure if I had KMS at that time, otherwise same Gentoo installation).
So fglrx in Linux was best, 5 W more in Windows, and another 5W more with free drivers. Not super shiny for the free drivers and visible difference but not screaming bad.

With just the onboard HD3300 the difference was less visible (all around 45-50W or something).

I should really measure stuff today again with my HD 5670, again Sapphire. But iirc. the difference last time I tried was not so much. Same mainboard still with a mediocre DSDT and stuff (MSFT compiled).
So I think it actually got better, but to be exact I would have to find my old 3870 and compare again. (The 6xxx series was out already but I chose the 5xxx series since I knew it was better supported at that time.)

But then, I have a dedicated card, from a vendor that normally sticks close to the AMD-ATI reference design. (And with my poor user-only-knowledge) I guess a lot of problems might also come from vendors doing their own tweaks and modifications, especially on laptops. Then any default mechanisms might not get a good grip so you need a lot of quirks in the drivers.

At least here on Phoronix I rarely see anybody complain about missing OpenCL features but daily requests about power management and video accelearation issues (UVD/shaders). So maybe the AMD-ATI devs reading here should note to their superiors that these things and some 3d performance improvements seem to be a really pressing point among users. To give that more priority.

About the lack of coders: I wish I could code. Then I'd like to spend some time on that stuff. But I decided to study chemistry instead of informatics / coding. And now I am partially even working on weekends on my PhD stuff, revising student's lab protocols and revising proposals for scholarships from foreign students and whatsnot. :/
So no chances for me to get into coding rand contributing.

All I can do is to spend some beer, cola or Mate tea for devs to keep them in a good mood. ;-)

[agd5f]

I use rovclock to manually clock my graphics chip down at boot, and then I clock it up and set HYPERZ whenever I run a game.. I think it's actually *BETTER* than the AMD Catalyst drivers on Windows XP simply because the AMD Catalyst drivers on Windows XP *ALWAYS* ran the GPU at full clocks when the laptop was plugged into the wall.. So whenever my laptop was plugged into the wall, it was a lot hotter than on battery, but now I have a lot more control over it than I ever had in Windows.. So I don't complain..

I've gotten my Mobility X700 all the way down to 100/150 from 350/350.. Anything less than 100/150 is unstable...

You are probably actually using lower clocks than you think. The problem with rovclock is that it doesn't take into account the post dividers on the PLLs. So unless the post divider is set to 1, you are actually programing much lower clocks than you think. Additionally, many of the vram related registers changed between 1xx and newer generations so adjusting the mclk can be problematic on newer cards. For atom based cards (r4xx and newer), I'd recommend using the atom tables to adjust the clocks manually.

[remm]

I also switched to Intel thanks to Ivy (not great on paper to be polite, but very decent on LInux "with no proprietary kernel module" in practice). Looks like this was the right decision, ATI is really going nowhere.

[Sidicas]

I think the fact remains that for a lot of the Radeon Mobility hardware, it's not possible to fix the power management simply because the driver doesn't know how to properly clock the graphics chips..

ATI gave the laptop manufacturers freedom to clock some of the mobility chips any way that they wanted to meet their heat/power usage capabilities of the laptop. So some laptops are designed around the idea that the graphics chip is going to be clocked a little bit lower than ATI's "stock clock" for the chip..

I use rovclock to manually clock my graphics chip down at boot, and then I clock it up and set HYPERZ whenever I run a game.. I think it's actually *BETTER* than the AMD Catalyst drivers on Windows XP simply because the AMD Catalyst drivers on Windows XP *ALWAYS* ran the GPU at full clocks when the laptop was plugged into the wall.. So whenever my laptop was plugged into the wall, it was a lot hotter than on battery, but now I have a lot more control over it than I ever had in Windows.. So I don't complain..

I've gotten my Mobility X700 all the way down to 100/150 from 350/350.. Anything less than 100/150 is unstable...

[DanL]

The ATOMBIOS vs radeonhd thing in this article is bullshit from the usual phoronix shit.

You know journalists... They love "angles"