If the depths are floating point values, they can simply be in "real
world" units, and do not need to be normalized.
However, 16-bit floats are not really sufficient for "real-world" depth
data, with units in feet, meters, or centimeters. When used as an input
to something like a depth-based fogging node in Nuke or Shake, you'll
see noticeable banding for a scene of the scale of a building or larger.
If you can make it so that the 'z' plane is 32-bit float, with the rest
being whatever precision you need for R,G,B,A - it will improve the
"depth-based fog in comp" kind of usage considerably.
If you do want to normalize the depths, you could put the near & far
values into the header's metadata.