Really? There are lots of architectures that have that already, and really when you're using the vector unit you're going to be limited more by bandwidth than by execution resources. I hear that it will let you have more precision in the intermediate state, though, which some scientific computing people will care about.
