The training accuracy of "trainNetwork" decreases a lot when I used another computer

I ran EXACT the same code on two computers with function "trainNetwork" to train Convolutional Neural Networks. Both of the computers have Matlab R2016a installed. One with GPU nVidia Quadro 3000M have 99%~100% training accuracy. But the other one with newer GPU GTX 1070 only have about 65% training accuracy, and the accuracy doesn't improve through the training iterations. Is it possible that Matlab doesn't support the latest nVidia GPU?

Answer 1

[Joss Knight]

NVIDIA doesn't support the latest NVIDIA GPU:

https://www.mathworks.com/support/bugreports/1439741

Sorry :-(

[Dalibor Knis]

How does this information relates to official cuDNN release page information that "cuDNN is supported on Windows, Linux and MacOS systems with Pascal.." https://developer.nvidia.com/cudnn

I am confused.

[Joss Knight]

The latest version of cuDNN does support Pascal. The versions in MATLAB R2016a and R2016b do not. MATLAB builds with and ships with the version of cuDNN current at the time the product is completed. To upgrade an old version of MATLAB after product release is not possible. However, we have asked NVIDIA for a back-ported fix for the older versions of cuDNN, which we will bring to you as soon as we have received that. Keep your eye on the bug report.