C++ to python boundary slow?

[sb5m]

Hello Pybind11 community,

I recentily started to take a look at PyBind11 and I found it to be interesting. I decided to follow a Microsoft tutorial for my first project. Tutorial

And everything was smooth sailing. The thing is that my benchmark results look nothing like the one attached in the tutorial. I understand that I can possibly be microbenchmarking but at the end of the day I am doing the same thing, so I do not understand such a big difference.

Tutorial:

Running benchmarks with COUNT = 500000
[tanh(x) for x in d] (Python implementation) took 0.758 seconds

[fast_tanh(x) for x in d] (CPython C++ extension) took 0.076 seconds

[fast_tanh2(x) for x in d] (PyBind11 C++ extension) took 0.204 seconds

Debug:

Running benchmarks with COUNT = 500000
[tanh(x) for x in d] (Python implementation) took 3.565 seconds

[fast_tanh(x) for x in d] (PyBind11 C++ extension) took 3.289 seconds

Start without Debugging (Ctrl + F5):

Running benchmarks with COUNT = 500000
[tanh(x) for x in d] (Python implementation) took 0.274 seconds

[fast_tanh(x) for x in d] (PyBind11 C++ extension) took 3.170 seconds

Any ideas?