Hi Markus,
There are two things related to SNR going on with hardware binning.
One that you mentioned is that you may get a benefit in reduced read noise. In a perfect world you'd get the same amount of read noise in a "pixel" whether you're doing x1, x2 or more. In the real world you'll often get less benefit than this. My experience with KAF-8300 and KAF-16803 sensors is that you're doing well if you only get 2x the read noise with x2 binning. The other thing to say about this is that if you're able to take sky limited subs then you don't care about read noise.
The second thing is that you're measuring more photons in a single pixel when binned. Shot noise is SQRT(signal) so binning x2 gives you double the per-pixel NR (at the cost of reduced resolution.) This can be done just as well in software. Another way of skinning this cat is software noise reduction - a controlled blur that loses some resolution in return for improved SNR. I like noise reduction better than binning because you can specifically target the areas where noise is most visible: the low signal areas and small spatial scales.
There's another potential benefit from not binning your RGB. You can use it to create a synthetic luminance or to improve real luminance data.
Summary: IMO, binning is usually a waste of time
The only justification I can find for it is when you're chasing dim NB targets and can't do subs that are long enough to be sky limited.
Cheers,
Rick.