Not sure I agree. What is the point of the benchmark? Figure out how to make the fastest Mandelbrot set? Probably not - then we could just copy-paste the solution posted by Kristoffer.
To me the interesting thing is to figure out why Numba can create better code than Julia, while the compiler presumably have even less information.
What is the Julia compiler missing? How can it be improved? This benchmark shows there are some gains to be had in a usecase that is probably hit quite often.