I agree with Erik, use fixed point if you are trying to do direct digital control. The multiplies and divides using floating point math take much too long to do direct digital control. The number of bits that you must use will dependent on the accuracy that you want to achieve from the system being controlled. A 16 x 16 multiply will need 32 bits for the result. I have used 16 bit fixed point calculations with very good accuracy.

Regards,

Jacob