Because of how floating point works, it doesn't matter. The difference between modelling in meters or millimeters disappears in the exponents. What matters for precision is the mantissa, which is 23 bits in single precision floats. So you get ~16 million of your smallest unit before you're losing precision, counting the implied leading 1. You can double that to ~32 million by using the sign bit.

Edit: this translates to 320m if you want 0.01mm precision.