Hi, I’m using matlab’s neural network package and trying to use a pair of nVidia Titan V GPUs. I used to use an nVidia 1080 GTX and that worked fine. But at the beginning of training of the network with the Titan V, MATLAB hangs for a while (30 seconds???) and then I get this error message:
Error using trainNetwork (line 140)
Failed to initialize the cuDNN handle. Return code was CUDNN_STATUS_INTERNAL_ERROR.
I have tried this both on Ubuntu 17.10, and 16.04, with identical results.
I have nVidia binary driver version 390.25 (open source), CUDA 9.1 and the three associated patches from nVidia website installed, as well as cuDNN 7.1, and all other packages are up to date on the machine, and I have installed “MATLAB” from the Ubuntu app store which is not MATLAB itself but a helper of sorts fixing some issues with the MATLAB distribution (old dynamic link libraries in the MATLAB distribution cause problems, e.g. plots do not appear).
Worth repeating: everything works fine with nVidia 1080 GTX board, the problem seems to be with the very new Volta architecture used in the nVidia Titan V GPU board.