Way I understand it, computers can internally represent floating point numbers only as binary approximations with a finite number of binary digits. So since mantissa and exponent are stored only with an arbitrary precision, rounding errors happen. Hope that's basically correct.