From my experience, and the documentation it is not f2py
that makes the difference - it is your or the final users Fortran compiler - f2py
generates a wrapper not compiles the code itself. For those Fortran dialects that are not mentioned i.e. 90, 2003 & 2008 as specifically supported you may, (depending on the features that your code uses), have to generate the interfaces yourself.
I have not tried f2py
under python 3 but the current numpy
notes suggest that most packages are available. Again it will be the Fortran compiler that will determine the speed and memory footprint of the extension.