Hi all,

We purchased a machine with 3 GPU unit, where the system parameters are listed below.

GPU0 (enabled) GeForce GTX 280, 1371 MHz, 1023 MB VRAM, Capability 1.3
GPU1 (enabled) Tesla T10 Processor, 1265 MHz, 4095 MB VRAM, Capability 1.3
GPU2 (upgrade) Tesla T10 Processor, 1265 MHz, 4095 MB VRAM, Capability 1.3

The GPU0 is installed on the motherboard of the computer while the other two are installed in an an expanded slot attached with the computer. However, the expanded slot is very LOUD, like a jet engine:)

Does anyone have some comments in this case? Does the hardware installation have any problem or it should be like that? Thanks a lot.

its a powerful beast, needs powerful cooling. doesnt suprise me that its loud. ever been in a server room? all the fans are on max as the computers need to be cooled effectively. the last thing you want is your tesla cards to overheat when you are three quarters the way through a calculation that takes a few days…

Are these two C1060s or a Quadro D2 or a S1070?

C1060s should be quiet, but the S1070 should be loud–it’s designed to put in a server rack, and like all server equipment, it’s very loud compared to your average workstation.

(also hooray carnegie mellon–are you using them in RI or something else?)

Thanks a lot! Actually,we put the machine in a sever room, but we are having some problem to visualize the data using GPU remotely, especially under MATLAB. A toolbox called Jacket enables the visualization using GPU computing in MATLAB, however, it does not support remote access. Could anyone here give us some hints about how to utilize the powerful machine with remote access, while entailing the GPU-based visualization of graphics, figures, images, etc. under MATLAB and C++?

Thanks for your information, tmurray. We do use the S1070, so I think it should be that LOUD. The GPU units are used for the computational modeling of olfactory pathways, rather than stuffs for robotics:)

Well, this is fairly far-fetched, but if you really wanted to, you could do something like this:

  • Compile your matlab program to a DLL (I don’t know if Jacket supports this, if not, you could actually use use interop/COM and get the same functionality as the DLL).

  • Write a program (I like .NET/C#, but you can use pretty much anything for this) that interacts with the DLL and also exposes a SOAP service to call functions on the DLL and return data.

  • Visualize the data remotely by setting up some kind of Windows Media Encoder tools to publish a streaming .wmv with the output of your matlab code.

Pretty complex, but it would give the flexibility to house the server anywhere (even a remote location) and control it from anywhere, and still get the same functionality as sitting down at the workstation. You could even automatically store the results in a database, etc. using the program in step 2.

Or you could just access the machine remotely with VNC or NX…

Of course, neither this nor profquail’s solution will work if you require the use of CUDA/OpenGL interop to display data from the GPU in real time. There really isn’t any way to do that except directly on a workstation attached to the monitor in front of you. But I really don’t know what Jacket is doing, so…

Maybe what you need is a KVM over IP device?

Eg. http://www.thinklogical.com/kvmoverip.asp

It’s just VGA video, but should do the job.

Jacket is doing CUDA/OpenGL interop, so you really need to have a monitor connected. Maybe a nice kvm extender???

This may be OS-sepecific, but I was able to use RealVNC from Windows XP to launch and view a CUDA/OpenGL interop application running on a remote Windows Vista machine. It worked well, though the visualization was fairly simplistic…