Shipping fat binaries with different versions for every different Atom/Core/Xeon/FX/Phenom/Opteron Intel/AMD/Nvidia + chipset glue combination would be quite a feat.
This could go well beyond what compiling every package for your specific machine can do.
That's not what I was talking about. I said that, if all you support is a specific Intel Atom processor, you can tweak your kernel to support every bit of energy-saving performance-enhancing silicon in there.
Shipping fat binaries is not new. I remember them (not very fondly) from the MacOS 7/PPC days.