It's a language whose syntax makes it easy to pass around multi-dimensional arrays, bounds checking is a standard feature on most compilers, doesn't have the same aliasing problems as C so compilers can do more aggressive optimizations and it compiles directly to your native hardware's instruction set (no vms/sandboxes). Why wouldn't you want that for cpu bound numerical code?

Also, the PG (now nVidia) Fortran compiler makes it very easy to make your code run on a gpu with only adding some attributes to variables and some pramgas.