So saying that in floating point your position resolution is 2.5m is not quite true... it depends on what value you want to represent...
Sorry: I should have said: If you are at about zero degree lattitude but 179 degrees longitude and you store lattitude and longitude directly into a 4byte IEEE756 floating point representation, then your longitudinal resolution will be only 2.5m.
Yes, I thought you meant something like that...
But just to be clear for the others: this is is not only a float issue... (although with an appropriate fixedpoint representation you can do a better than with float, which was what Christophe wanted to point out).
It is also an issue of the coordinate representation used (always get's ugly when you approach the poles).
This is one of the reasons why INS filters are usually not implemented in lat/long ;-)