A 20D also seems to have its lowest Read Noise at ISO 1600, slightly lower compared to other ISO settings.
Does this mean that by applying that extra amplification (to "ISO 1600") early, right at the photo-sites, the relative impact of subsequent noise in the read-out and A/D conversion is minimized?
Naively, this makes sense, and suggests something like Samir's strategy, or at least my understanding of it, which is as follows.
1) amplify a lot and early, say to "ISO 1600", to minimize the effect of noise introduced later in the process (read-out, A/D).
2) use an A/D converter than can handle the resulting high maximum signal strengths, for example in situations where one is using "exposure index ISO 100", and thus amplifying to a level of four stops of "over-exposure". That is, add four bits or so of highlight headroom to the A/D convertor. So yes, go to 16 bits A/D or even more, in that direction.
Then selected ISO speed affects shutter speed/aperture choice in AE modes, but goes into the raw file as a mere suggestion for default raw conversion, like WB, sharpening, and such. With manual shutter speed and aperture setting, the light meter reading could be used to record a "suggested" ISO speed in the raw file.
My question is how large a signal-to-noise ratio an A/D convertor for a portable camera can have these days, or in the foreseeable future. That limits its useful bit range, regardless of what the spec. sheets say. No point the A/D being able to handle very high input levels if the noise added by the A/D convertor is much more that 1/65536=1/2^16 of that maximum signal, because then the least significant of those 16 bits are all A/D noise, no signal, and "16-bit" becomes purely a paper spec.
Aside: with current SLR sensors, the only possible benefit of 16 is covering up to four stops of exposure level error, which is related to what I (and Samir?) propose as a deliberate strategy. With correct exposure, abut 12 bits is enough to record the S/N ratio of any current SLR sensor itself, and trends at the high end (Kodak and Dalsa) seem to be holding the line or even dropping S/N ratio for the sake of increased pixel count. (Leaving the options of binning or down-sampling when the user prefers extremes of high DR at less high resolution.)