So, If NVDIA actively supports some API Python interface on top of CUDA C API (maybe reusing Pycuda and numpy), it would standardize and have new tools easily.
It also benefits back NVDIA
OPen Source can also support it, like tensor flow of google.
I think it is unrealistic to expect support for CUDA 8 features in PyCUDA at this time, given that CUDA 8 hasn’t even been finalized yet (the final version is expected this month, from what I understand).