My suggestion to approach this is to install everything with conda and then use the same environment for panthera and Python (in this way you get MKL for free)

Then I'd test short running things and long running stuff, because my wild guess is that for short things Python should be faster, but the overhead involved should be more or less constant and so negligible for long running stuff

I will try over the weekend! Some MOOCs stuff to finish before :) (And real life brother visiting)