There is another difference between CMOS and CCD when we look at multiple read-outs, which is probably what Erik was actually referring to. The Read-out of a CMOS sensel signal is virtually non-destructive, so one could perform multiple read-outs and average the results to reduce the read-noise component further with each additional read-out. This will of course slow down the creation of each single image.
A CCD on the other hand, transports its charge out in bucket brigade fashion and collects noise on each shift to the next sensel. After the read-out, the original sensel signal is gone. So only a comparison with the Reset signal is possible.
Thanks for the additional information. Correlated double sampling is one way to reduce read noise, but there are alternatives as discussed in this post
which I bookmarked some time ago. Canon was the first major vendor to implement CMOS in their 35 mm sensors and their methodology was discussed in a white paper where they explained that no loss of image quality occurred, but it was unclear to me if they were using correlated double sampling or some other methodology. The Sony Exmoor sensors have very low read noise, but I have never read a clear explanation of the methodology (it may be a trade secret).
Thermal noise is often mentioned along with the suggestion that one should cool the sensor, but for most terrestrial exposures of 1 second or less thermal noise is negligible. Nikon does do some type of dark field subtraction for long exposures, but they also use hot pixel suppression, which does not please the astronomical community, because it also erases faint stars. Parallel processing allows slower readout and is one method used by the Exmoor. Cooling is often used in scientific and astronomical imaging, but these applications use slow readout to reduce read noise, but this allows additional time for dark noise to accumulate and makes cooling more necessary.
Hardware pixel binning can also reduce noise, but, as far as I know, the only company to achieve this is Phase One with their sensor plus technology. Apparently this is not feasible with CMOS. One can pixel bin in post (e.g.) 2:1 downsizing, but this involves 4 read noises rather than the one read noise that occurs with hardware binning where the superpixel is read out with only one read noise.
This is interesting to geeks such as myself and I would be interested in further discussion, but the general forum population likely has little interest in this arcane topic.