The trainNetwork does not work with Pascal GPU. What's the next step?

Based on the bug report 1439741 the cuDNN library does not work correctly with the latest NVIDIA cards e.g. GTX 1080. And this is the reason why MATLAB has disabled trainNetwork in Neural Network Toolbox.

OK, but what is going to happen next? Once NVIDIA releases CUDNN that works with Pascal what can MATLAB user do to utilize the new library? Is it enough to download the latest cuDNN library and the existing MATLAB version 2016b will be able to consume it? Or shall we have to wait for new version of MATLAB?

Answer 1

[Joss Knight]

We have asked for a compatible version of cuDNN from NVIDIA with a backported fix. If they can provide that then we will provide a patch to MATLAB that incorporates it. Keep your eyes on the bug report, and your fingers crossed that a fix is actually feasible. If not, waiting for R2017a is the fallback.

Unfortunately simply patching in the latest version of cuDNN is not possible.

[Joss Knight]

The fixes have now been published and can be downloaded from the bug report link.

[Yanfeng]

The bug still exists after the workaround was installed. The following error was displayed:

"Error using vision.internal.requiresCUDAComputeCapability30 (line 24) GPUs of compute capability 6.0 and above are not supported due to a problem with this version of the cuDNN deep learning library. See bug report 1439741."

[Dalibor Knis]

Joss: Thank you, the patch works as advertised. :-)

Yanfeng: are you 100% positive that you replaced the required files?

[Joss Knight]

Sorry about that Yanfeng. The original patch for R2016b was missing a fix to CVST to remove that error. A new patch has been published which corrects it.

[Roshan Welikala]

The patch works perfectly within Matlab. But it doesn't appear to fix the issue in respect to running outside Matlab (exe and runtime). Any suggestions? I am using Matlab 2016b and gtx 1080. Thanks.

[Joss Knight]

We do not currently have any mechanism for patching the Matlab Compiler Runtime, so I'm afraid the only solution here is to upgrade to Matlab R2017a.