GPUs problems in Windows 10

Why does Matlab still not work well with GPUs in windows 10, I am using a Titan V GPU and the CNN simulations are very slow, what is the problem? Because the Matlab staff does not solve this problem, versions of Matlab pass and pass and the problem persists. Right now I am evaluating matlab 2020a and I am still confirming that it is useless to have one or more gpu's if Matlab is used; This leads me to make the decision not to acquire a license and look for another development framework as an alternative to Deep Learning.

[Mohammad Sami]

Is Matlab able to detect your GPU ?

Type gpuDevice in command line to check.

[Mohammad Sami]

In addition using a GPU requires Parallel Computing Toolbox™ and a CUDA® enabled NVIDIA® GPU with compute capability 3.0 or higher.

[ANDRES FERREYRA]

Matlab if it detects the GPU.

>> gpuDevice

ans = 

  CUDADevice with properties:

                      Name: 'TITAN V'
                     Index: 1
         ComputeCapability: '7.0'
            SupportsDouble: 1
             DriverVersion: 10.1000
            ToolkitVersion: 10.1000
        MaxThreadsPerBlock: 1024
          MaxShmemPerBlock: 49152
        MaxThreadBlockSize: [1024 1024 64]
               MaxGridSize: [2.1475e+09 65535 65535]
                 SIMDWidth: 32
               TotalMemory: 1.2885e+10
           AvailableMemory: 1.0279e+10
       MultiprocessorCount: 80
              ClockRateKHz: 1455000
               ComputeMode: 'Default'
      GPUOverlapsTransfers: 1
    KernelExecutionTimeout: 1
          CanMapHostMemory: 1
           DeviceSupported: 1
            DeviceSelected: 1

[Joss Knight]

It's impossible to answer your question without knowing what you are doing. What is your model, your data, are you doing training or inference, and what is the code you are using. For instance, if your model is very small (a digit recognition model perhaps) then it is completely expected that the GPU, being poorly utilized, does not give a performance benefit over the CPU.

Also, from your other question it is clear you are using an external GPU enclosure, which means your memory bandwidth is significantly affected, by a factor of maybe 10.

It is also clear that you are using the much slower WDDM driver mode. It may not be available to an eGPU but if possible you should enable TCC mode.

You also do not say what your expectations of performance are, such that you believe MATLAB is slow. All I can do is assure you that we use Titan Vs and we get great performance when training and running CNNs on Windows 10.

Answer 1

[ANDRES FERREYRA]

Matlab if it detects the GPU

>> gpuDevice

ans = 

  CUDADevice with properties:

                      Name: 'TITAN V'
                     Index: 1
         ComputeCapability: '7.0'
            SupportsDouble: 1
             DriverVersion: 10.1000
            ToolkitVersion: 10.1000
        MaxThreadsPerBlock: 1024
          MaxShmemPerBlock: 49152
        MaxThreadBlockSize: [1024 1024 64]
               MaxGridSize: [2.1475e+09 65535 65535]
                 SIMDWidth: 32
               TotalMemory: 1.2885e+10
           AvailableMemory: 1.0279e+10
       MultiprocessorCount: 80
              ClockRateKHz: 1455000
               ComputeMode: 'Default'
      GPUOverlapsTransfers: 1
    KernelExecutionTimeout: 1
          CanMapHostMemory: 1
           DeviceSupported: 1
            DeviceSelected: 1