8Bit nomals maps vs. 16Bit ones


i saw the Ferrari Demo on a R9700 now and i have a question on the normalmap format they are using.
I know perpixel lighting on fragment basis, but i used 8x8x8bit components to compute the normalvector. This little demo show the usage of 16x16bits instead. Ofcourse it looks amazing precise, but i don’t understand how they compute the third component out of the texture. Can some1 help ?
I know that it uses floatingpoint precision, but how can i recontruct a third component for the normal from a 16x16 texture ?