I fully recognise the point about a full MWE - the census here seems to be that any MWE beats none - so I’ll try and come back with one. I was aware that MatLab implicitly multithreads - but it uses exactly the same resources as the Julia program - 100% of a single core and nothing more - so kinda doubt (but may well be wrong) that its gaining any significant average.
@fastmathshould be among the very last things you try. I don’t think I ever use it, and it may change the results numerically.
Very reasonable point - I was initially just shocked at the performance discrepancy and trying everything and anything in the performance tips reference (it helped that I didn’t have to rewrite masses of code to implement it either).