If you need absolute precision then floating points aren't going to be optimal. In particular your constraints need pretty much all the precision a 32-bit floa

If you need absolute precision then floating points aren't going to be optimal. In particular your constraints need pretty much all the precision a 32-bit float has to offer, and you still need a handful of exponent bits, so your total savings would be like 4 bits.