can't find GPU in MJS

I have set up a MJS in my own PC, which has a CPU (Intel Xeon E3-1230 V2, 4 cores 8 threads), and two GPUs (GeForce GTX 670 and 560). Matlab is in the version of R2013a, including PCT and MDCE toolboxes. Visual Studio 2012 ultimate and CUDA 6.0rc are also installed.

I use "admincenter.bat" under the default setting of "mdce_def.bat" to set up a MJS with 4 workers, named "Pattern_CUDA". And I run the following commands

>> matlabpool open Pattern_CUDA

Starting matlabpool using the 'Pattern_CUDA' profile ... connected to 4 workers.

>> spmd disp(gpuDeviceCount); end

Lab 1: 0

Lab 2: 0

Lab 3: 0

Lab 4: 0

It seems these 4 workers didn't find any GPUs.

Then I open the "local" scheduler with 4 workers, and run the same commands as

>> matlabpool open local 4

Starting matlabpool using the 'local' profile ... connected to 4 workers.

>> spmd disp(gpuDeviceCount); end

Lab 1: 2

Lab 2: 2

Lab 3: 2

Lab 4: 2

This means this PC are installed with CUDA and any other related correctly.

Even if I leave just one GPU alone, namely GeForce GTX 670, the workers of "Pattern_CUDA" still can't find the GPU, but the workers of "local" still can.

Is anything I missing when set up the MJS? How can I let the workers of MJS find GPUs?

I am very appreciate for any help!

Answer 1

[Edric Ellis]

The MJS workers cannot access the GPU because they're running as a Windows Service, and you need to use the TCC mode driver to allow them access. This section of NVIDIA's documentation tells you more. As noted in that page, "NVIDIA GeForce GPUs do not support TCC mode".

[Xiang Gao]

Thanks a lot for this reply!

So, it means Geforce GPUs can not be used by MJS worekers in Windows?

Can I try them in Linux (e.g. Ubuntu)? It seems the documents related to Linux didn't mention any differences between Tesla and GeForce cards.