They've also claimed both on the page and in the podcast that certain testing reveals a 68,000x speed up vs Python. That's probably true in a select case, but this is why you use libraries in Python - to make use of all the hard work that already exists in C/C++ in those libraries, with Python serving as a high level and far easier to write functional wrapper. Ie anyone serious about performance isn't going to be looking at that metric seriously, rather they're going to be asking what the real performance gains are vs standard practice. In that case the answer seems to be rather roughly a 1.5x speed up based on what's on the performance page of their site when comparing to PyTorch and CUDA, which is a more realistic scenario.

I am curious about how much work the compiler does to automatically identify available hardware and distribute the load on its own vs how much you need to specify as a developer.