Which NVIDIA card for CUDA with 2008 Mac Pro?

Hi,

I’m looking to get an NVIDIA card for my (2008) Mac Pro as I am interesting in using CUDA (I have an ATI card at the moment). However, I can’t figure out which cards would be compatible with my specific model of Mac Pro. The ‘8800GT for Mac’ on the NVIDIA GeForce list is for the first generations of Mac Pros, and the link to the card for second generation mac pros is broken. The NVIDIA card on the apple store is for early 2009 Mac Pros. Here are the specs for my computer:

Model Name: Mac Pro
Model Identifier: MacPro3,1
Processor Name: Quad-Core Intel Xeon
Processor Speed: 2.8 GHz
Number Of Processors: 2
Total Number Of Cores: 8

Does anyone have recommendations of cards that would be compatible? I’m running 10.5.7.

Thanks for any advice!

Thomas

I could not find the 8800 GT on the Apple store either.
Just the GT120… which is a bad joke considering how much faster the 8800 GT is.

Looks like the best think to do is wait and pray for that rumored GTX 285…

Mark

Two of my 2008 Mac Pros, same as yours, are equipped with 8800GT in slot 1 and PC versions of a GTX 260 and 285 in slot 2. You can use any of the PC cards under Windows with bootcamp, and it is fine with CUDA 2.2

Under OS X you have some options:

  1. Buy one of the few remaining 8800 from the UK Apple Store, being sure to get the correct version, at

http://store.apple.com/uk/product/MB137Z/A…mp;s=topSellers

  1. Wait a couple of weeks to get the Mac Edition GTX 285, now available to pre order at provantage and other world computing (mine is on order!)

  2. If you really need that level card, get the Quadro 4800.

I am sure that now option 2 is your best bet. The OpenGL test results show the 285 doing rather better than the expensive Quadro, and it has more cores.

There are also ways of hacking in ordinary PC cards to get them to work, but unless NVIDIA make the drivers they are shipping with options 2 and 3 more widely available, you will have to wait for Snow Leopard for 200 series support. How about it NVIDIA - all the drivers for other OS are on your web site - why not Mac? The Mac forums also have tricks to get 8series and 9 series PC cards working.

I would not bother with 120 as a CUDA engine - perfectly nice monitor driver, but too few cores.

In case you understand German:

http://www.myvideo.de/watch/6378735/Saugeile_Grafikkarte

(I could not resist)

The new 285 does work just fine in an 08 Mac Pro, and has been running CUDA 2.2 under 10.5.7, OpenCL under 10.6 seed, and both CUDA and OpenCL under Windows in some tryouts I did today. All very good. :thumbup: One minor issue is how to control when it goes into a low power mode, but I reckon that will become clear. Make sure you get the latest drivers from evga ftp site.

Hi MacFan,
this is great to know. I’m waiting for the 285 to become available in Germany…

Would you mind sharing the performance of the 285 under Leopard & CUDA 2.2.

I am especially interested in the performance of the histogram64 and histogram256 examples.
Also, does the 285 support concurrent copy and execution?

Regards
Mark

deviceQuery reports concurrent copy and execution and also compute level 1.3 as expected. I am a bit reluctant to report performance because I have noticed this card seems to have a throttle back mode where the clock speed drops and maybe even some cores are disabled. If you look at my screen shot you will see the MonteCarlomultiGPU code run twice. I have a PC 285 in there as well. There is factor of about 8 between full speed and throttle back on the mac card (check options per sec). I wake the GPU up by thrashing it with OpenGL benchmark between those two runs. Busy spamming other forums to see if there is a control panel somewhere to stop Green operation! On its own the Mac card managed over 100k options per second when woken up. I know this code much better than the histogram codes so am confident about these numbers.

Thanks a lot. Impressive, with my GT8800 I get about 40k ops/sec. So the GTX is really about twice as fast as the 8800 (as it should be).

If you have the time to post that histogram benchmarks sometime I’m still interested - especially because the histogram 256 is so hard to optimize, so I’m wondering if the GTX really has so much an advantage there. On the GT8800 I get:

histogram64GPU() time (average) : 9.565000 msec //10454.783522 MB/sec
histogram256_SM11() time : 19.650032 msec // 4853.296495 MB/sec

About the “Grren” operation of the GTX 285: I guess this only happens when you run the card not driving a display - right?

Mark

Suprisingly, no! The Mac 285 is the one I boot from and is attached to my main display - it is the PC card that is just cuda coprocessor and not driving a monitor most of the time, and that one stays up at 1.48GHz come what may.

With only the Mac card loaded I got it up to just over 100,000 options per sec - there is clearly an overhead for the multi operation.

I will post the histogram numbers as soon as I can - sending this from a different machine - I just need to make sure I have woken up the card! Someone on macforums has pointed out a possible bug in Apple power management tool so maybe this will get sorted out soon.

So with the GTX 285 “woken up” by some initial bashing with 3D stuff, the corresponding numbers are

histogram64GPU() time (average) : 4.556000 msec //21949.077018 MB/sec

histogram256_SM11() time : 12.808063 msec // 7445.890527 MB/sec

If you do it with the card in Green mode, running at 0.6Ghz and who knows how many cores up, instead you get

histogram64GPU() time (average) : 16.646999 msec //6007.088596 MB/sec

histogram256_SM11() time : 24.688562 msec // 3862.818342 MB/sec

I hope this helps - it clearly illustrates the need to control the power management - the card itself is great IMHO.

Here is MonteCarlomultiGPU run on an awake card on its own

bash-3.2$ MonteCarlomultiGPU
main(): generating input data…
main(): starting 1 host threads…
main(): waiting for GPU results…
main(): GPU statistics
GPU #0
Options : 256
Simulation paths: 262144
Time (ms.) : 2.468000
Options per sec.: 103727.717474
main(): comparing Monte Carlo and Black-Scholes results…
L1 norm : 3.083966E-06
Average reserve: 361.436058
TEST PASSED
Shutting down…

In Green mode it reports

Options per sec.: 10054.988578

so that is 1/10 as fast. I do hope someone from Nvidia is reading this…

I ran NiBiTor under bootcamp and it confirmed that the Mac card has 3 shader clock states: 1.48, 0.8 and 0.6 Ghz, with even bigger scale downs in the memory clock, so this is consistent with the Cuda times. But my PC card is set to the same three states and does not have the same narcolepsy problem. The states are “Extra: 1.48”, “3D 0.8” and “2D 0.6”.

Can someone from Nvidiai explain whether a BIOS edit will fix it, or is it the case that with the problem being Mac side only this is an EFI setting? Can we have an OS X to control the power better? This is a fabulous card but its performance is being screwed by the power management. I have to run nbody before I run one of my own codes just to kick the GPU into full speed.