Catalyst 10.1 and Xorg 7.5 / 1.7.x?

[Darksurf]

My point is that the OSS driver is not good for my card because of this and why I keep using Catalyst. The OSS drivers don't do the stuff AMD's/ATI's hardware designers intended a driver to do. Correct driver-level power management is an important part. I do not want the card to be useless in a year from now. I intent it to be used in older machines when I get a new card at some point. I do not trust the OSS drivers at this point; I believe they will result in hardware damage in the long run unless they implement power management correctly.

Don't take this the wrong way, but have you payed much attention to the temp capabilities of AMD and ATI hardware? Hell, the Turion64 single core cpu was allowed to get up to 80C without thought, and was set to power off at 90C, AMD and ATI have always ran a little hot, but thats due to lots of power in a little area .

Its not like they didn't expect the gpu to get that hot. Try playing high-end video games for hours, they didn't make this gpu expecting you to not put a load on it, or constantly run in it in performance mode. Otherwise you would be running an X200m lol.

[mirv]

However, heat generation is proportional to the square of the voltage, so the effects are much more apparent. I don't know what the ForceLowPowerMode in the OSS driver does, but I know that using fglrx in low power mode is much more effective than using rovclock with radeon to simply downclock the card (even when pushing it to the very low limit, next to screen corruption).

The same goes for the CPU, I undervolted my laptop's processor by a tiny amount at each frequency and now the fan only kicks in when it gets stuck at 100% usage (read Flash, heh).

Actually, power loss through heat is proportional to the square of the current - it's just that applying more voltage in high frequency circuits can also change impedance, leading to it also being related to the square of the voltage. Just wanted to clear that up a little bit. In any case, downclocking can't match switching components off that aren't in use, which is where I suspect the fglrx driver shines over the open source driver in power management.

[bridgman]

I should also mention that the voltage changes here are pretty small, maybe 20-25% reduction, so even if you square that you aren't making a huge difference in power dissipation. It all helps, though.

The ForceLowPowerMode option currently reduces the engine clock and the number of pcie lanes IIRC.

[yotambien]

Actually, power loss through heat is proportional to the square of the current - it's just that applying more voltage in high frequency circuits can also change impedance, leading to it also being related to the square of the voltage. Just wanted to clear that up a little bit. In any case, downclocking can't match switching components off that aren't in use, which is where I suspect the fglrx driver shines over the open source driver in power management.

Well, actually you lost me there. I was thinking in terms of Ohm's law. If R is constant V is proportional to the current...But anyway, I guess we can agree on P = C V^2 f.

I should also mention that the voltage changes here are pretty small, maybe 20-25% reduction, so even if you square that you aren't making a huge difference in power dissipation. It all helps, though.

You are thinking it the wrong way. For a 25% reduction in voltage you have:

Pa = k Va^2
Pb = k (3Va/4)^2

Pb/Pa = 9/16

Which is a lot less.

[mirv]

Well, actually you lost me there. I was thinking in terms of Ohm's law. If R is constant V is proportional to the current...But anyway, I guess we can agree on P = C V^2 f.



You are thinking it the wrong way. For a 25% reduction in voltage you have:

Pa = k Va^2
Pb = k (3Va/4)^2

Pb/Pa = 9Va/16

Which is a lot less.

Oh that...yes, well, it has been a while since I did the math directly (sadly, I always did want to go more in that direction, but there weren't any opportunities down under). Anyway, "R" isn't always constant (often changes with heat, which is always fun)...but I won't go into more detail, as that's straying from what the main focus is - the link you provided explains rather well why reducing the frequency will consume less power over a time period - but reducing voltage will allow less instantaneous power (at the risk of not driving the circuits with sufficient power, but that's another story).

[bridgman]

You are thinking it the wrong way. For a 25% reduction in voltage you have:

Pa = k Va^2
Pb = k (3Va/4)^2

Pb/Pa = 9Va/16

Which is a lot less.

We are thinking about it the same way, so we're either both right or both wrong

I was thinking .75*2 = .5625 so ~44% reduction in power dissipation going from highest to lowest voltage, which is IIRC about the same as you can get by dropping engine and memory clocks. Again, I'm not saying the power savings from voltage reduction is not important, just that it's in the same ballpark as the power saving from clock reduction and that both are equally important.

[not.sure]

So, is there a 10.1 in the pipe with xorg 7.5 support?

[RealNC]

OK, I'll try experimenting again with the OSS radeon driver and see if things changed. Question though: how to monitor GPU temps with that driver?

[yotambien]

We are thinking about it the same way, so we're either both right or both wrong

I was thinking .75*2 = .5625 so ~44% reduction in power dissipation going from highest to lowest voltage, which is IIRC about the same as you can get by dropping engine and memory clocks. Again, I'm not saying the power savings from voltage reduction is not important, just that it's in the same ballpark as the power saving from clock reduction and that both are equally important.

Ohhh, I see what happened. For a moment I actually thought that you sucked at math : D

Anyway, I think it's all clear now. One cool think to have when the voltage reduction gets implemented in the OSS driver would be to allow users to adjust them manually. If GPUs are anything like CPUs they probably are overvoltaged by a safe margin. Tweaking that sort of thing is always fun.

[bridgman]

Apparently I suck at typing, however. I typed .75*2 rather than .75^2.

Some of the newer boards seem to have fine grained voltage control (not sure if that's what our PM code uses though), but most of the older boards seem to just have a single GPIO line that lets the driver choose between high and low voltage presets.