If you want infinite precision, be aware that AI2 has a sub-layer that knows rational numbers, where
division of integers is deferred as long as possible and both the numerator and denominator are kept internally.
See the attached example for proof.
A floating point intermediate result would be incapable of supporting 50 digits after multiplication.
Mind blown yet?
ABG