Clock rate throttling on Mac 285

[10THzMac]

There are some very odd things going on with my Mac 285 installed half an hour ago. It seems to have two speeds and I wondered if anyone knows how to control this. I am using the drivers released by Nvidia in the last day or two, which are better than the ones on the CD.

I have been running the OpenGL Extensions viewer benchmark and also some CUDA timing examples. The first time I ran OpenGL I got rubbish results, then the card seemed to speed up and was much better. Similarly I ran the CUDA deviceQuery, and also the MonteCarlomultiGPU code and got two wildly differing numbers.

For example, deviceQuery reports a clock rate of EITHER 0.6GHz OR 1.48Ghz, and the "options per second" reported by the MonteCarlo code differs by a factor of 10 (TEN!!!)

Sorry if I am being dumb - but do we know if there is a sleep mode and how to keep the card awake. Screen dump below illustrates the issue. I ran the top code after hammering the card with OpenGL benchmark, then waited and ran the next one. Note the change from 1.48Ghz to 0.6GHz. The OpenGL scores are pretty good in awake mode.

I am trying to compare the Mac version with an injected PC one and the results are rather confounded by the throttling.

 

[10THzMac]

Here are some CUDA numbers for illustration - it is quite shocking. Output is from MonteCarloMultiGPU with a pair of 285s in an 08 Mac Pro.

Test 1 (DOG mode)
Mac 285: 9613.22 options per second
PC 285: 82580.65 options per second

Test 2 (awake mode, after waking up with OpenGL benchmark test)

Mac 285: 83441.98 options per second
PC 285: 81893.8 options per second

In awake mode the cards have identical clock speeds of 1.48Ghz.

 

[peskaa]

So the Mac 285 seems faster, once it has "woken up"? Have EVGA said anything on this?

 

[Cindori]

They are supposed to downclock themselves in idle, no?

 

[peskaa]

I can see the logic on downclocking when idle to reduce power load, but doesn't really help when the card doesn't actually wake up fast enough to be at full power when being used the first time round.

PS: what's the difference between the drivers off the EVGA site and some nVidia ones? (I also can't find the NV ones on their site, where are they from?)

 

[10THzMac]

The Nvidia ones were on the CD (they might even have been the same ones that netkas found for the Quadro) and I think are now irrelevant (dated early May). The new ones are dated June 10 and are noted on the EVGA forum by an EVGA forums administrator at

http://www.evga.com/forums/tm.asp?m=100725204

and the driver link there is the same as noted earlier by netkas, i.e.

ftp://ftp.evga.com/Driver/MAC285_2F16.dmg

so it's all nicely official at last (I might have confused things by referring to Nvidia instead of EVGA - I think Nvidia must have made them though EVGA posted them!!). Who knows exactly what changed - I just noticed that now both my monitors work and the OpenGL scores are a few % better on several tests. I am trying to nail down the throttle back issue on other forums. I am all for being green but need the full speed on demand. If you run OpenGL benchmark it seems to take most of the 1.0 test to wake up.

 

[Cuda]

The Nvidia ones were on the CD (they might even have been the same ones that netkas found for the Quadro) and I think are now irrelevant (dated early May). The new ones are dated June 10 and are noted on the EVGA forum by an EVGA forums administrator at

http://www.evga.com/forums/tm.asp?m=100725204

and the driver link there is the same as noted earlier by netkas, i.e.

ftp://ftp.evga.com/Driver/MAC285_2F16.dmg

so it's all nicely official at last (I might have confused things by referring to Nvidia instead of EVGA - I think Nvidia must have made them though EVGA posted them!!). Who knows exactly what changed - I just noticed that now both my monitors work and the OpenGL scores are a few % better on several tests. I am trying to nail down the throttle back issue on other forums. I am all for being green but need the full speed on demand. If you run OpenGL benchmark it seems to take most of the 1.0 test to wake up.

I'd have a look at, or remove for testing:
/System/Library/Extensions/AppleGraphicsPowerManagement.kext
/System/Library/Extensions/AppleGraphicsControl.kext

Odd thing is, under:/System/Library/Extensions/AppleGraphicsPowerManagement.kext/Contents/Info.plist

MacPro4,1
Vendor10deDevice05e2 (GeForce GTX 260???)
Vendor10deDevice0640 (GeForce 9500 GT???)

You may just want to try to edit that 05e2 with your 05e3 and see what that does.

I think what is going on is AppleGraphicsControlClient is loading for the Mac 285GTX, but the Power Management is not defined, thus odd goings on, or I could be way off.

 

[10THzMac]

Neat idea Cuda, but no luck so far. My Mac is 08, i.e. 3,1 so I am not sure I am triggering anything at the moment.

The only other thing I have found out is that the clock has at least three states, 1.48Ghz, 0.8GHz and 0.6GHz. I ran the Cuda deviceQuery several times in succession after some heavy graphics and observed it winding down through these levels. There are some posts on similar matters on EVGA and CUDA forums so I am going to wait now to see if there is a mod for thresholding on the power.

 

[netkas]

Dump bios with gpuz, open it in NiBiToR and see clocks for 2d/3d modes, show the sshot of nibitor with openned powerstates.

 

[10THzMac]

Dump bios with gpuz, open it in NiBiToR and see clocks for 2d/3d modes, show the sshot of nibitor with openned powerstates.

That's a new one on me. Getting NiBiTor and will report back. Thanks (AGAIN)

OK - so netkas suggestion reveals the issue. The shader clock and memory have three speed states. See screen dump below - exactly the clock states revealed by CUDA, but also the memory speed drops right down, which explains the factors of 10 in some perf. Now - how do I get it going faster - fewer states? Or will this fry my card? (And how is anyone supposed to manage any of this without a handy bootcamp partition...)


Further comment - my Palit 285 has similar settings - so it is the transition between them that is problem - that card tends to stay at 1476.

 

[netkas]

That's a new one on me. Getting NiBiTor and will report back. Thanks (AGAIN)

OK - so netkas suggestion reveals the issue. The shader clock and memory have three speed states. See screen dump below - exactly the clock states revealed by CUDA, but also the memory speed drops right down, which explains the factors of 10 in some perf. Now - how do I get it going faster - fewer states? Or will this fry my card? (And how is anyone supposed to manage any of this without a handy bootcamp partition...)


Further comment - my Palit 285 has similar settings - so it is the transition between them that is problem - that card tends to stay at 1476.

if u dont care about power cosumption, then dump whole rom with nvflash , edit clocks wit nibitor, and flash the card, also share rom if not mind

 

[Cuda]

if u dont care about power cosumption, then dump whole rom with nvflash , edit clocks wit nibitor, and flash the card, also share rom if not mind

Yes, if you edit the Bios side for the Rom, with all PStated(PerfTables) to values you want, and flash it back, we'll know if the Mac card gets its PStates from the Bois side of the Rom under OS X.

I'm thinking that it won't work, bet you a $1 PStates are defined in the EFI Side.( I'd be happy to be wrong. )

Post a Dump of the ROM, and we can look and see what it holds.

For me, with a 8400m gs, OS X read the PerfTables in the PC Bios and sets them on the fly( POWER Mizer )

It was trouble before 10.5.6, after that it works great.

 

[10THzMac]

if u dont care about power cosumption, then dump whole rom with nvflash , edit clocks wit nibitor, and flash the card...

I could edit the clocks as you say, but what about the point that my PC card has the same three states, as I discovered a bit later, but that card is running at 1476 all the time and never seems to throttle down. Is there a cooling issue and might I fry the card or even the Mac? I took out the Palit PC 285 to give more room for the air to circulate and the Mac 285 is still down at 0.6GHz. I'd just like to get views on this before I fiddle with the heat state!

Edit:
Just saw Cuda comment - I think he might be right as the Palit card having the same values in nibitor but behaving differently would be consistent with it being set elsewhere in the ROM.

Give me some time guys - I need to be careful with this new and expensive thing.

 

[netkas]

ioreg -l -w0 -p IODeviceTree | grep device-properties

pm me the output of this command, will look later what keys ur efi inject into ioreg.

 

[Cuda]

I could edit the clocks as you say, but what about the point that my PC card has the same three states, as I discovered a bit later, but that card is running at 1476 all the time and never seems to throttle down. Is there a cooling issue and might I fry the card or even the Mac? I took out the Palit PC 285 to give more room for the air to circulate and the Mac 285 is still down at 0.6GHz. I'd just like to get views on this before I fiddle with the heat state!

Edit:
Just saw Cuda comment - I think he might be right as the Palit card having the same values in nibitor but behaving differently would be consistent with it being set elsewhere in the ROM.

Give me some time guys - I need to be careful with this new and expensive thing.

Try this

Works for most Nvidia Cards under OS X.