A convolution filter can theoretically change the volume radically.
Normalization is simply a method to make the volume after convolution roughly -6dB compared to the volume without convolution. The -6dB is used to provide a little extra headroom since the signal will likely need it once in a while due to the variance added by convolution.
In other words, normalization is just a straight volume adjustment. It's done in 64-bit, just like everything else. The volume adjustment is fixed, and consistent between uses.
You might not really need it if your convolution filters are crafted carefully, but it also won't hurt anything.
Finally, since we're talking about signal headroom, I'd recommend using Internal Volume if you use convolution for this reason:
http://wiki.jriver.com/index.php/Volume#Internal_Volume_Headroom
