This is still confusing for the layman, Emil. You wrote that the the lower noise at high ISO was due to the lower circuitry noise before analog pre-amplification.
If this is the case, whatever the analog gain later applied, the lower circuitry noise, prior to gain and with or without gain, should also have an effect on lowering noise whatever the ISO setting, which is merely an instruction to apply gain.
[{POST_SNAPBACK}][/a]
Sorry, I didn't want to resort to using equations. The point is that the gain is not applied to all noise sources, only to those upstream of the amplifier; downstream noise sources (including the noise of the amplifier itself) are not amplified.
As I understand it, the sensor readout has an associated noise N; this feeds into the amplifier that does ISO, which applies a gain G and adds its own noise N'. The output has two independent contributions to the noise, G*N from the amplified sensor readout noise, and N' from the amplifier itself. For instance, take the 1D3; the noise G*N is about .8*ISO/100 and the noise N' is about 4.2 raw levels; at ISO 100, N' is the bigger contributor, while at ISO 1600, G*N~13 is bigger.
Because the low ISO and high ISO noise are determined by different circuit elements, reducing one of them does not necessarily result in improved performance for all ISO. And note that it is the noise from circuit elements *before* the ISO amplification that determine the high ISO noise performance; this I think is why CMOS provides better high ISO performance, by having lower noise in the sensor element readout.
BTW, I probably shouldn't have said the ISO amplifier is off-chip. It can be and often is on the sensor chip itself, but I don't think it's part of the pixel circuitry itself, and that is the only distinction that was important for what I was saying...
[a href=\"http://en.wikipedia.org/wiki/Active_pixel_sensor]http://en.wikipedia.org/wiki/Active_pixel_sensor[/url]