Fortran Outsmarted Our Billion-Dollar AI Chips
medium.com"The language assumes contiguous memory, column-major order, and no hidden aliasing between arrays."
One out of three ain't bad.
"Column-major order", that's the one. That just says that there is an implied order of elements of an array. This enables efficient mapping of an array of the language definition to storage with a linear address space. It does not require storage to be organised in any particular way.
"Contiguous memory". Not really, the language definition does not even use the word "memory". The language is carefully defined so that you can implement it efficiently on a linear address space in the presence of caches (no accident that IBM machines in the 80s were pushing caches while CRAY was pushing vector processors). The term "contiguous" in the language definition just means a collection whose parts are not separated by other data objects.
"no hidden aliasing between arrays". This is a crude mis-statement of the actual rule of argument association across subroutine/function caller-callee boundary. The rule takes pages to describe fully. A language that still has EQUIVALENCE statement (although marked obsolete) cannot be said to disallow aliasing. It is still quite hard to find a compiler that will reliably diagnose inadvertent aliasing at runtime. The actual rule in Fortran says something like "thou shall not cause any effect on (dummy name) A through a reference to some other name, unless (some-clause-x or y or z)".
Relevant: A GPT-2 implementation in Fortran and an experience report: https://ondrejcertik.com/blog/2023/03/fastgpt-faster-than-py...
It's very likely that there is some serious autovectorisation going on behind the scenes.
Telltale slop: "And honestly? I didn’t see it coming."
The bolding of subjects of the sentences, the pacing of the various sections, the overuse of em dashes.
Something felt weird with the whole thing. Like I was supposed to read it and have my mind blown? But there wasn’t enough substance to go around for an entire article.
Likely AI slop.