The full native support for Pascal architecture cards exists in CUDA versions 8.0 and higher. This version of CUDA is integrated by MATLAB in R2017a.
All MATLAB GPU code is forward-compatible to future architectures using a byte-code representation that allows it to be compiled by the graphics driver at runtime. No NVIDIA CUDA 8.0 compiler was available when R2016b was released for native compilation, however this runtime compilation offers limited support.
When MATLAB encounters a card with an architecture the GPU code was not originally compiled for there will be a one-time delay for GPU computing commands whilst the byte code is compiled and cached. Users may find the need to increase the CUDA cache size to prevent this delay recurring.
We recommend achieving this by setting the environment variable:
CUDA_CACHE_MAXSIZE 536870912
This sets the CUDA cache size to 512MB which is the minimum we would suggest. The default for this is 32MB. For further instructions on how to change environment variables on your system please consult here If you are intending to use an NVIDIA Pascal card with Convolutional Neural Networks in MATLAB please be aware of the following bug report https://www.mathworks.com/support/bugreports/1439741 where Pascal cards produce incorrect output for versions before R2017a. Other functional problems are currently being investigated.
Best Answer