isn't the noise floor only in the a/d conversion ?
If you define noise as an error
in the signal (something that was not present in the original signal), you get errors on both ends (A/D and D/A) of the conversion. You have quantization error (rounding off error) that is noise. There's thermal noise. There's nonlinearity errors, more so on chips that were not laser trimmed. Modulation noise on higher order delta sigma ADCs. And so forth.
Now if you do multiple A/D and D/A conversion such as digital mixer with its analog outputs feeding a DSP, and the DSP analog outputs feeding a processing amp (amp with built-in DSP), noise could be apparent depending on overall system gain.
Dither on the other hand is random, low-level noise added to increase apparent S/N and create an apparent lower noise floor on systems with lower bits such as 16 bits. The analogy I use is how people cook rice. You can just use plain rice and water, or you an add a bit of dithering agent such as butter, sticky rice (to increase apparent starchiness, or vinegar (as in sushi rice.)
Use of floating point versus fixed point arithmetic: that's the architecture of the DSP, not the A/D or D/A converter. Floating point does not mean better S/N or dynamic range. Floating point can represent a larger value within given word length (e.g. 32 bits), but you can just as easily use wider words (e.g. 40 bits) fixed point.
I agree with what everyone said...with modern digital mixers using 24 or 32 bit converters, there's not much to worry about w/r/t noise floor, even with feeble input levels.