Also, the drawback of anything written in this way is that it doesn't compose well. The individual operations may well have been optimized into oblivion, but their composition may not necessarily be what you'd like it to be. One can also say goodbye to higher-order functions. But I'm not exactly a Fortran person when it comes to code style, so I may be prejudiced against overly loopy code.
(BTW, am I the only one who thinks that the Sourceforge (http://sourceforge.net/projects/math-atlas/) comments are machine-generated? I mean, comments like "An essential Windows program." or "The best program that I've ever used." are hardly what I'd expect for a numerical library.)
Composability is a huge issue: I think the whole reason SciPy exists is to provide an interoperability standard for this sort of work. That said, I think your average scientist can more intuitively see ways of composing _using_ matrices than the rest of us.
(Checks the Sourceforge reviews. ROTFL. What's the purpose of this spam?)