>>3892519>>3892688>>3892830You're reading the graph wrong, common rookie mistake. And shot noise only starts to dominate after the knee of the curve. If you move one stop to the right in the RAW exposure axis (vertical lines in the grid) and to the next ISO curve you see the resulting SNR for that obtained exposure. Naturally this depends a lot on the sensor. 0EV is the point where S/N is 0, and corresponds to 0dB [log(1)=0]. 12dB is 2EV, and it's the point where you start getting some usable texture. So by raising the expo ISO you gain S/R in those deep yet usable shadows.
Here sensors with different behaviors were selected to illustrate how it works. On the Pentax, raising ISO is nearly useless. The Nikon shows improvement until ISO 400. The Canon up to ISO 1600. These are old cameras but I'm lazy and it's the best I could find. DxO's "full SNR" has this kind of curve but they made some unfortunate choices like using a base 10 logarithm as the scale for the x-axis. You need a base 2 logarithm to easily go up and down in stops.
Shot noise only dominates at high signal levels, as noise is determined by the square root of the sum of read noise squared plus the ratio of signal to gain (which has read noise baked in). When signal is low, read noise squared is more important. The change in the slope of the SNR curves for a given ISO correspond to that.
And the dynamic range of the scene being captured absolutely matters.
