Tesla m2090 on 5,1 Mac Pro help needed

[DoctorZoidberg]

Greetings from Argentina!
I'm in process of building myself a budget CUDA/OpenCL based color-correction/video post production workstation. After countless hours of reading this forum, I conclude in a 2009 4,1 Mac Pro. After updating it to 5,1 status, I'll throw an hexacore (maybe a W3590), and a dozen (or two) GB of RAM, using Mavericks or Yosemite.
I'm planning to use a nVidia Telsa m2090 as video accelerator, complementing the GT120.
I understand nVidia don't make Tesla drivers for the MacOS, but in a post by MacVidCards on Netkas Forum I learned the Quadro 5000 driver can control it. Is it true? If I don't missed the point, PCIe speed will be 2.0 in all 4,1/5,1 cMP. Am I correct?
I don't care about boot screens ("m" series Teslas has no monitor out; only "c" series have), but I'm a little confused about TDP. Some places says 238W, some other 250, and the manual says "<= 225 W". Assuming the Pro can output a maximum of 225W using PCIe and the two "booster cables", is it possible to limit the TDP flashing the ROM? Perhaps it sounds like a silly noobie question, but I will not stress the PSU (nor install another supplementary PSU for the video card).
So, if the ROM has to be flashed, manual says the BIOS is a "2Mbit Serial ROM", I assume there is plenty of room to flash whatever code is needed.
I also understand a GTX780 is more powerful than the Tesla, with its aging Fermi architecture. Fact is, here down under video cards have excessive import duties and I can have the m2090 for peanuts. Besides, I won't be using Windows on this machine (maybe some day I'll try some RHEL like CentOS)
Thanking in advance!!

 

[xcodeSyn]

I'm planning to use a nVidia Telsa m2090 as video accelerator, complementing the GT120.
I understand nVidia don't make Tesla drivers for the MacOS, but in a post by MacVidCards on Netkas Forum I learned the Quadro 5000 driver can control it. Is it true? If I don't missed the point, PCIe speed will be 2.0 in all 4,1/5,1 cMP. Am I correct?

... but I'm a little confused about TDP.

nVidia only makes one single set of drivers for Mac OS X, and they should work for Tesla GPU as indicated in this thread running Tesla C2075, whether they were optimized for Tesla is another question. BTW the Quadro drivers are the same ones used by all GTX cards as linked in the NVIDIA FAQ thread. You just need to download the appropriate drivers matching your OS X version.

As for the TDP issue. The Mac Pro is safe to run any video cards with 225W TDP or below (75W from the PCIe slot and 75W from each auxiliary video power connector). Since the Tesla M2090 is rated at 250W TDP and using both 6-pin and 8-pin video connector, it may be safer to use an external PSU. On the other hand this document on page 14 seems to imply that using a 6-pin to 8-pin connector may also be acceptable. It's up to you to decide which way to go. You may want to send MacVidCards a private message to ask for his opinion since he's quite knowledgeable about the TDP issues.

One last thing, the CPU you want to upgrade is either W3690 or 3680 since there is never a W3590 six-core CPU made by Intel.

Edit: The above document from nVidia only stated power consumption <= 225W instead of 250W cited else where. So you should be fine using the internal PSU and two 6-pin power connectors.

 

[Surrat]

The bit about the c2090 being able to use just the 8pin cable, or 6+6 or 6+8 is interesting, never knew of a vid card that could do that.

My big question tho, is how do you cool this thing? it has no built in fan? Its a blow thru design, but is the macpro pci card fan strong enough to keep it cool on its own?

I wonder if MacVidCards has experimented with a rom for this card, maybe just to insure pcie 2.0.

 

[netkas]

wait, m2090 has no video outputs, you can only use it for cuda/opencl (if can) in mac, no opengl.

 

[Surrat]

wait, m2090 has no video outputs, you can only use it for cuda/opencl (if can) in mac, no opengl.

Yeah, its not a vid card at all, just a compute processor, should have called it that earlier.

 

[philliplakis]

The Tesla m2090 is just a GTX 580 with 6gb ram and no display output but at double/triple the price.... 512 fermi cores. Is there a reason you need the 6gb?

Personally you should look into maybe a pc GTX 680 4gb? 1500 kepler cores. and you could flash it yourself. Just a though..

 

[xcodeSyn]

I wonder if MacVidCards has experimented with a rom for this card, maybe just to insure pcie 2.0.

If I'm not mistaken, NVIDIA web drivers have provided PCIe 2.0 link speed since OS X 10.8.3 or 10.8.4 for MP4,1 and later models. The thread mentioned earlier about the Tesla C2075 only showed PCIe 1.1 speed in CUDA-Z simply because it was run under OS X 10.7.4 using Darwin 11.4.0 kernel.

 

[xcodeSyn]

The Tesla m2090 is just a GTX 580 with 6gb ram and no display output but at double/triple the price.... 512 fermi cores. Is there a reason you need the 6gb?

Personally you should look into maybe a pc GTX 680 4gb? 1500 kepler cores. and you could flash it yourself. Just a though..

As the OP stated that he could get the Tesla for cheap and you can't compare Kepler CUDA cores directly to Fermi cores as explained by this AnandTech review.

Consequently, this is why we can’t just immediately compare the CUDA core count of GK104 and GF114 and call GK104 4 times as powerful; half of that additional hardware is just to make up for the lack of a shader clock.

Generally speaking, the GTX 580 (Fermi) has better CUDA and OpenCl performance than the GTX 680 (Kepler) as indicated here. You can get a Tesla M2090 on ebay for US$169 and 221 were sold already. Apparently there is still some demand for these used cards for GPGPU work.

 

[DoctorZoidberg]

First of all, thank you all for the replies.
Step by step:
1) xcodeSyn, I also read on the manual the ">=225W TDP" statement. nVidia site lists 238W (or 250) and wikipedia lists it as 250W. Strange. But, I suppose if it supports a modded a 8 (150W) to 6-pin (75), it could run safely on 225 max. Oh, and the "3580/3590" was my mistake. A typo. Thanks for the correction.
2) I plan to use Da Vinci resolve. I prefer the 6GB because higher resolution color grading needs more VRAM than CUDA cores. And it's true: Fermi cores can't be compared in number to Kepler architecture.
3) It will be used as a CUDA "engine" only. I'll use the native GT 120 for GUI (at least until MVC come back the 2GB flashed GT640!!!)
4) I have to admit I didn't consider the "no fan" cooling. Have to do some research.

 

[netkas]

First of all, thank you all for the replies.
Step by step:

3) It will be used as a CUDA "engine" only. I'll use the native GT 120 for GUI (at least until MVC come back the 2GB flashed GT640!!!)

Then you need no efi, auto-init should work just fine.

 

[DoctorZoidberg]

Then you need no efi, auto-init should work just fine.

I only asked if EFI would be needed just in case of TDP exceeding 225W

 

[Surrat]

First of all, thank you all for the replies.
Step by step:
1) xcodeSyn, I also read on the manual the ">=225W TDP" statement. nVidia site lists 238W (or 250) and wikipedia lists it as 250W. Strange. But, I suppose if it supports a modded a 8 (150W) to 6-pin (75), it could run safely on 225 max. Oh, and the "3580/3590" was my mistake. A typo. Thanks for the correction.
2) I plan to use Da Vinci resolve. I prefer the 6GB because higher resolution color grading needs more VRAM than CUDA cores. And it's true: Fermi cores can't be compared in number to Kepler architecture.
3) It will be used as a CUDA "engine" only. I'll use the native GT 120 for GUI (at least until MVC come back the 2GB flashed GT640!!!)
4) I have to admit I didn't consider the "no fan" cooling. Have to do some research.

I'm guessing that you might end up dropping down to a c2075 instead of the 90, its the same GPU, just little slower clocked and it does have its own cooling fan. Specs and clock speeds for the different cards are online.

 

[xcodeSyn]

4) I have to admit I didn't consider the "no fan" cooling. Have to do some research.

This GPU was designed to work in servers and that's why it only has passive cooling. You can probably use either smcFanControl or Macs Fan Control to manually adjust the intake, exhaust, and PCI fan speed when the card is running. This link has a detailed discussion on how to add active cooling if that's what you are interested. One problem when using non-Apple GPUs is that Mac OS X does not provide any temperature readings on them, if I'm not mistaken.

 

[DoctorZoidberg]

Thanks again for your time and patience.
I've been thinking on put an axial blower on the end of the card casing. Don't know how to deal with cfm or noise, but I think it could be done.