There is a whole field of CS that deals with minimizing error when doing floating point math. They just probably use decent algorithms/encoding.