Quadro K4200 Multi-GPU

[TeslaK20]

Hello
Can an NVIDIA Quadro K4200 be used in a multi gpu setup? (Can you use two such cards in one computer?) Considering one can either buy a second Xeon CPU or a second Quadro GPU for about the same price, which will provide a better performance increase? (I am using Adobe, Autodesk, Blender)
Thanks

 

[Constantine Yevseyev]

It actually depends on:

• Whether you're using an OEM-built workstation (e. g. HP Z640) or custom one.
• Whether your software provides support for multi-GPU rendering or not.

In theory, you could create any kind of setup (for example, use two Quadro K620 accelerators with one of them serving as primary graphics adapter), as long as there are two seamlessly available PCI-E slots on your motherboard. The second K620 will be used as so-called "offload" GPU during the execution of GPU-intensive tasks. Same goes to K4200, and also to configurations with more than two GPUs.

Now, the trick is, only your OEM/IHV can guarantee that such configuration will work (providing all kinds of support in forms of drivers and consultations). For example, HP offers their customers with support for up to two K4200 in their Z640/Z840 workstations:

http://h20195.www2.hp.com/V2/GetPDF.aspx/4AA5-4274ENW.pdf

The second thing that you have to think of is that sometimes the application itself can't really benefit from having any extra GPU power at all. For example, you'd better off with a second Xeon if you're planning on using Autodesk AutoCAD (the explanation for that can be found here). At the same time, you should be totally fine with a couple of K4200s when using Adobe Premier Pro CC:

http://international.download.nvidia.com/adobe/pdf/9353_Adobe_CC_Solution_Overview_Sept29_R14_HR.pdf

Blender will use your cards separately (as opposed to using them as a single logical device), but you'll have to turn off SLI in control panel first in case when using SLI-compatible configuration.

So what does your machine look like, and what programs exactly are you going to use?

 

[TeslaK20]

I use a DELL Precision Tower 7910.

 

[Constantine Yevseyev]

Dell Precision Tower 7910 Workstation does support multi-GPU configurations composed of two NVIDIA Quadro K4200 products, both running PCI-E 3.0 x16 link speed-wise/x16 electrically. They will be wired directly to a PCI-E host controller found inside your 4th generation Xeon E5 processor, occupying 32 lanes out of 40 provided.

I understand that you already have one K4200 GPU in your system? Then you should have no problems adding a second one. Any retail-available K4200 will do (look for PNY and HP products). They are offered in reference design exclusively, with no variations in VRAM size, clock speed or cooling solution used, so you don't have to worry about picking the wrong part.

Once inside the chassis, locate your current GPU which resides in the second PCI-E 3.0 slot (blue). It's separated from the fourth, black PCI-E 3.0 slot (which you'll be using for your other K4200) with a single PCI. Contact this manual (page 50) for more details:

http://topics-cdn.dell.com/pdf/precision-t7910-workstation_Owner's Manual_en-us.pdf

Still, I would strongly recommend providing more information on the software you're planning to use, since there's still a chance that second Xeon will do more. And keep in mind that SLI is not supported with this particular configuration, so you might have to contact your software vendor for clarification on this topic. "Offloading" on a secondary GPU is widely supported by ISVs, but it still requires some programming to make everything work (while SLI can be usually resolved as a single computing device on a device driver level).

 

[TeslaK20]

My use is "media and entertainment". Rendering, 3ds Max, Adobe AE and Premiere, Blender, Dynamic Simulations (for entertainment, not science).

 

[Constantine Yevseyev]

OK, let's see.

Using 2x K4200 will do more than 2x Xeon E5 in 3ds Max:
http://international.download.nvidi...Autodesk3dsMax_SolutionOverview_US_FNL_HR.pdf

Same goes to Adobe After Effects:
http://blogs.adobe.com/aftereffects/2012/05/gpu-cuda-opengl-features-in-after-effects-cs6.html

Same goes to Blender (as I mentioned earlier) and many dynamic simulation suits (they will usually query the total amount of CUDA cores when creating warps).

Adobe Premiere Pro is a different story. You can only benefit from having multiple GPUs in specific scenarios, so having two Xeon E5 processors might be a better choice here:
http://blogs.adobe.com/premierepro/2013/05/improved-gpu-support-in-adobe-premiere-pro-cc.html

 

[GustavZ]

Hey @Constantine Yevsevev
I know this post is 3 year old, but i have a really similar question:

- I have Dell Precissiom Tower 7910 which already has one nvidia quadro k4200 in it.
- I also have a second exact same nvidia quadro k4200
- I want to use this Workstation for Training Neural Networks

My questions:
Can i put the second Graphics Card into the Workstation and use both for Training?
Is it necessary to connect them with a SLI Connector? Do they support SLI (here is said not http://www.nvidia.de/object/quadro-sli-compatible-systems-de.html)
Is SLI even usefull for training neural networks?
If SLI is not usable, will training be faster if in the two cards without SLI?

Thank you very much!