Actually, how any plug in handles non-integer values (ie fractions) is a very big deal. Simply rounding to the closest whole number available is the second worst solution. The worst solution is to simply chop off the numbers at an arbitrary cut off point (truncation). Plugs that truncate sound like ass. Plugs that round simply sound cheap.
As much as folks would like to forget dither, and as wierd as it is to handle conceptually, it's a necessary evil. Waves Renaissance plugs do it automatically, and Waves IDR is designed to avoid "build up".
Some folks believe that dither below 24 bits is inaudible and thus a waste of time. This perspective can be supported in multitrack mixing: The sum of 10-100 tracks of audio generate enormous numbers, and the low level detail on any individual track is lost, and the least significant bit (ie quietest sounds) is even LESS significant when you consider the dynamic scaling necessary to fit 48 or 56 fixed or 64 float bits into a 24 bit signal. Hence the digi recommendation not to worry about it until the very last step, where this scaling occurs.
This idea falls apart in a couple circumstances: If you're mastering or working with just a few tracks, that detail remains in the number space, and can easily be fit and can be necessary. Second, if every track is using lots of plug-ins and were recorded at non-optimal levels chains of cheap truncated plugs will push the truncation errors into audible range (normalizing at the track level will exacerbate these problems, by making quantization distortion audible, alongside the truncation).
The dirty little secret: These accumulated errors are quite audible, and are responsible for the hard to nail down myth about ProTool's mix bus sounding bad. PT sounds fine. It's the crap you put between the fader and the master, and the narrow plug window that mess things up. Digi's actually done a great job cleaning up that bus and making a solid number cruncher. They've done a less good job defining TDM2 as yet another 24 bit bus. 32 bit float is computationally a little neater, so VST, MAS and AudioUnits all have a huge edge at the channel level over TDM systems. DirectX can do 64 float I think, which is near ideal. Had they bumped it to 48 they'd have halved track counts (already halved by HD), but truly removed the need to worry about dither between plug ins (the LSB at 64 or 48 bits is so far down it's truly irrelevant). As it is, your mix will only sound as good as the most prominent cheap plug, or longest chains of them. Doh!
When you need to dither: Anytime your signal changes bit depth. When does this happen? Any gain change turns your 16 or 24 bit fixed point source file into a 32 float or 48 fixed point word. Any plug in or DSP. Any mixing operations. Virtually EVERYTHING you do expands the remainder/product of math operations. Good coders know this and design plugs to elegantly deal with remainders in a way that won't build up noise.
Unfortunately sticking a Waves IDR or MOTU Quan Jr between your T-Raxx plugs isn't going to get it done. You need plugs written properly to start with. Better to just forget it if it's not there, and rely on the bus dither to do it's best. Most Waves plugs fall into that category: the Renn line automagically dithers, while other plugs give you a choice of IDR options. Typically the ones with options are things like L1 or L2 which are used at the end of the chain.
In general, flat or non-noise shaped dithers are best at the track level. Gentle noise shaping is ok for 24 bit mix busses. Aggressive noise shaping (IDR t2 Ultra, POW-R, Megabitmax, SBM) should be avoided until mastering.
If anyone would like a primer on how/why it works, I'll give it a shot.
-d-