Recording at 32 bit (float)

Researching the web for information on the benefit of tracking at 32 bit (float) rather than 24 bit (fixed) I came accross an informative page on Audacity's website:

"Normalised floating point values are quicker and easier to process on computers than fixed integer values and allow greater dynamic range to be retained even after editing. [...] With floating point, rounding errors during intermediate processing are negligible.

The (theoretically audible) advantage of this is that 32-bit floating point format retains the original noise floor, and does not add noise. For example, with fixed integer data, applying a compressor effect to lower the peaks by 9 dB and separately amplifying back up would cost 9dB (or more than 2 bits) of signal to noise ratio (SNR). If done with floating point data, the SNR of the peaks remains as good as before (except that the quiet passages are 9dB louder and so 9dB noisier due to the noise they had in the first place)." [Link]

Still, this benefit is only relevant when...

• handling >24 bit (fixed) data
  
• within an engine with a wider path than 24 bit (fixed)
  
• that interfaces transparently with your system.
  

If you want to record the output of a 24 bit (fixed) ADC you need only track with 24 bit (fixed) accuracy.

As soon as you manipulate that data, before (!) recording it--which IMHO might not be advisable--you will increase the resolution, dependant on the internal bus depth of the engine / interface. In the case of Metric Halo's line of audio interfaces (2882, ULN-2, ULN-8) the bus is 80 bit (fixed), so you will gain some precision by recording the data at 32 bit (fixed). [Thanks to B.J. Buchalter for the clarification]