It is not so much the numbers themselves, but the result of successive MAC operations. Here is where you need to be mindful of the effects of underflow and overflow.The numbers you use in FIR filters have well-defined fixed range (e.g. 0 to 1023 coming from 10-bit from ADC). Floating point calculations only introduce a burden by keeping track of the scale when the scale is always the same anyway. Thus, fixed scale FIR is considerably faster.
dsPIC33E with hardware DSP engine can calculate 3 million 16-tap FIR filters per second. You cannot easily match that with comparably priced MCU with FPU.
More so on mcu's without a fpu -> most mcu's fall into that category.Are there any advantages of using a fixed point MCU when designing FIR filgers?