Great.
One thing to try, depending on your setup, is whether GitHub - JuliaLinearAlgebra/MKL.jl: Intel MKL linear algebra backend for Julia speeds things up, just so you are comparing apples-and-apples against numpy/matlab. It may not help, but it is easy to try. @ChrisRackauckas found a tweak to openblas which may make it comparable to MKL, but it wouldn’t be implemented in Julia yet.
It will be interesting to see if TF is faster. If so, we would need to figure out why it is using a different algorithm. There is no fundamental reason (maybe aliasing behavior if it interprets the code?) that TF should be faster on a CPU.