But he was design in different way – so it hold as many float part as he can.

Exactly, the point is to maintain the most significant part of a (sub)result. The simplest example is multiply, if we think of our two numbers as consisting of a 24 bit integer (all fractional part so: 0.24) and an exponent, then the product simply adds the integer exponents and to store the new fraction part would require: 0.24 x 0.24 = 0.48, so 48 bits for the significand. Since we're using a fixed number of bits those 48 bits are rounded to 24.