Does MATLAB 2018b Support NVIDIA GeForce 2080 Ti RTX For Creating/T​raining/Im​plementing Deep Learning Neural Networks, And, If So, Then Which Version Of CUDA Should I Use?

Does MATLAB 2018b Support NVIDIA GeForce 2080 Ti RTX For Creating/Training/Implementing Deep Learning Neural Networks, And, If So, Then Which Version Of CUDA Should I Use?

Answer 1

[Walter Roberson]

https://www.mathworks.com/matlabcentral/answers/426316-cuda-with-turing-gpu-and-neuralnetworktoolbox-in-2017b

suggests a workaround to a known problem .

Answer 2

[Joss Knight]

Yes, you can. Use the following workaround to avoid annoying warnings and errors:

warning off parallel:gpu:device:DeviceLibsNeedsRecompiling
try
    gpuArray.eye(2)^2;
catch ME
end
try
    nnet.internal.cnngpu.reluForward(1);
catch ME
end

You can put this in your startup.m, for instance.

It is a common misconception to think that you need to install a particular version of CUDA. The vast majority of users just need to get the most recent driver appropriate for their GPU.

[Walter Roberson]

MacOs High Sierra and a certain prerelease, refused to run gpuArray until I installed Cuda 10. Newest display driver was not enough. Details of the needed display driver and cuda version available in one of my cases .

[Joss Knight]

Okay. I'm not saying you dont need to install a CUDA driver (although it is only on Mac where that is not packagd with the display driver), I'm saying you jt need the latest, not a particular version. For instance, your CUDA 10 setup should work on any previous MATLAB version, even though we have not yet released a MATLAB version built with CUDA 10. I admit however that there have edge cases and you have often been the one to point them out to me!

In R2019a we have gone to some effort to improve the language in the documentation to explain the system requirements, although there is always more we can do.

Apologies for typos, there's a bug in the editor and . it doesn't work properly on Android.

[Walter Roberson]

"Each MATLAB release has a minimum CUDA driver version, but any later version can be used as well, provided that NVIDIA still supports your architecture. As NVIDIA continually improves their drivers, we recommend that you use the most recent CUDA driver version that supports your architecture."

Which I would then follow by a list of releases and corresponding minimum CUDA version (annotated with information about when architecture families started and stopped being supported by NVIDIA.)

[Joss Knight]

Ah, well we already have that on our GPU support by release page. But yes, your wording is along the lines of what we'll end up with - I hope you'll be happy with it!

[Walter Roberson]

Cool, I've never seen that page before!

[AILEEN SENGUPTA]

I am sorry this problem seems to still have an appearence in my code:

Error using DAGNetwork/classify (line 193)

Failed to initialize the cuDNN handle. Return code was CUDNN_STATUS_NOT_INITIALIZED.

Error in SeriesNetwork/classify (line 577)

[labels, scores] = this.UnderlyingDAGNetwork.classify(X, varargin{:});

Any help is appreciated.

Thanks,

Aileen