Not sure what this means exactly, can you explain further? >> My experience is that when you pitch up in software you lose lots of dynamic range, same as when you pitch downward. ISO 50 on the 5D Mk2 gives me exactly one stop less headroom then the same picture has at ISO 100. If they would make a 5D Mk3 with native ISO 1600 which would go up to ISO 25600 (=H4), that would mean I lost 4 stops of DR in my image. Not very nice..
Why would Canon create a new sensor that wouldn't go past a native 3200 ISO after having one that does a very nice ISO 6400 (5D Mk2)?
When you start having more than one value in the raw file per electron in the sensel, it does not make much sense to increase the amplification further, you can just scale in software. Unity gain (1 raw value = 1 electron) happens quite early, say around ISO 1600 or so. Why some cameras still do "native" amplification past unity gain I don't really know, perhaps there are some read noise advantages or something.
Of course, if I am missing something technically, please do explain! Always willing to learn
In photography, you want to gather as much light as possible to get the best image quality. "Expose to the right" is the concept used when you are not restricted with the shutter, that is you can saturate the sensor (without overexposing important features of course).
The reason why more light equals better image quality is because light contains noise naturally (photon shot noise), but the stronger the light the smaller part of it is noise. If you reduce light with 1 stop, photon shot noise increase with 0.5 stop - the less light the lower signal-to-noise ratio.
ISO setting does not alter the sensor in any way, it only changes the amplification at readout. If we had readout electronics that precisely could count the number of electrons in each pixel/sensel, we would not need to have a ISO setting at all. It may happen in the future that the cameras will be "ISO-less".
Anyway, when you shoot high ISO it is because you need to have short shutter speeds and then we will not fill up the pixels with light, so we get more photon shot noise. Actually, the noise you see in high ISO pictures is typically dominated by shot noise, not noise from the electronics.
The sensor pixels are not perfect either though, not all photons that hit it becomes translated to electrons, so with improved efficiency of that (quantum efficiency), noise levels would be reduced for low light photography.
Yes if you multiply in software you lose dynamic range. But past unity gain it makes no difference, you've already lost the range. If your RAW value is in the range 1 - 16384 you still cannot represent 16385 electrons if your gain is >1. It could be some practical difference anyway, there must be some reason why they usually amplify past unity gain, but I don't really know why. Perhaps someone else can help us on that...