Using CUDA 6 for GPU computations (R2015b + parallel toolbox)

Hi,

I'm trying to use the Parallel Computing Toolbox with MatlabR2015b in order to perform computations on NVIDIA Tesla K40 GPUS.

I've read on your web site that "Parallel Computing Toolbox vous permet d'utiliser des GPU avec MATLAB. Il prend en charge : les GPU CUDA DE NVIDIA avec capacité de calcul version 1.3 ou plus, y compris les GPU Tesla (ces GPU prennent en charge les calculs double précision) ;"

My GPUs have a CUDA Capability of 3.5. Here is what returns the Cuda "deviceQuery" utility command for my cards :

 CUDA Device Query (Runtime API) version (CUDART static linking)
Detected 2 CUDA Capable device(s)
Device 0: "Tesla K40m"
  *CUDA Driver Version / Runtime Version          6.0 / 6.0
  CUDA Capability Major/Minor version number:    3.5*

When I try to detect my GPUs in Matlab R2015b, I get :

>> gpuDevice

    Academic License
Error using gpuDevice (line 26)
There is a problem with the CUDA driver or with this GPU device. Be sure
that you have a supported GPU and that the latest driver is installed.
Caused by:
    *The CUDA driver was found, but it is too old. The CUDA driver on your
    system supports CUDA version 6. The required CUDA version is 7 or
    greater.*

>> parallel.gpu.CUDADevice

ans =

CUDADevice with properties:

                      Name: ''
                     Index: NaN
         ComputeCapability: ''
            SupportsDouble: 0
             DriverVersion: NaN
            ToolkitVersion: NaN
        MaxThreadsPerBlock: NaN
          MaxShmemPerBlock: NaN
        MaxThreadBlockSize: [0 0 0]
               MaxGridSize: [0 0 0]
                 SIMDWidth: NaN
               TotalMemory: NaN
           AvailableMemory: NaN
       MultiprocessorCount: NaN
              ClockRateKHz: NaN
               ComputeMode: ''
      GPUOverlapsTransfers: 0
    KernelExecutionTimeout: 0
          CanMapHostMemory: 0
           DeviceSupported: 0
            DeviceSelected: 0

I'm surprised to see that I need CUDA 7.0 (with CUDA 6.0, I have a CUDA capability of 3.5).

Could you please :

• tell me what are the requirements for using GPUs with the parallel toolbox ?
• if there is a version of the parallel toolbox that supports CUDA 6.0 (I do not feel like installing Cuda 7)

Have a nice day,

Regards

Anne-Sophie Mouronval

[Joss Knight]

Can I ask you for a link to the website where you found those system requirements in French? They appear to be out of date.

Answer 1

[Walter Roberson]

R2015b requires CUDA 7.0 (and does not support CUDA 7.5)

R2015a requires CUDA 6.5

R2014b requires CUDA 6.0

See http://www.mathworks.com/help/distcomp/release-notes.html

You will thus need to use R2015a instead of R2015b if you do not wish to upgrade to CUDA 7.0 from your existing CUDA 6.5

Answer 2

[Edric Ellis]

Unfortunately, there are at least two different numbers here:

1. Your device has a CUDA Compute Capability of 3.5. This refers to the CUDA functionality that the device is capable of. For example, Compute Capability 1.3 devices were the first to support double-precision arithmetic.

2. The other numbers refer to the CUDA Toolkit release - in this case, 7.0. CUDA drivers support up to a given CUDA Toolkit. In this case, you need a driver that is capable of supporting the CUDA Toolkit version 7.0, or later.

In general, NVIDIA guarantee backwards-compatibility of their CUDA drivers, so it is always a good idea simply to install the latest CUDA driver for your device.

If you wish to compile CUDA Mex files for use with MATLAB, then you must use the CUDA Toolkit precisely corresponding to the release of MATLAB (this is described in the Parallel Computing Toolkit release notes).