I know this may be rehashing one of the most discussed topics, but I would like to take it a slightly different direction and solicit input. First of all, I am familiar with the electrical engineering and mechanisms associated with changing the camera gain, normally described in the DSLR community as ISO. I understand that it is changing the capacity of the electron well, which leads to a faster sensor for each pixel/photodiode. However, what I am trying to rationalize is theory vs. reality.
I have done due diligence with my Canon 70d and 6d and measured the read noise, thermal noise as a function of time for various stagnant ambient temperatures, measured the gain in electrons per ADU at most ISO's, determined the electron well size in photons per ADU and most importantly, made analogous measurements of the signal to noise ratio for various exposure times (in a dark frame) at various ISOs. In my measurements, as theory would predict, while the noise does increase as a function of ISO, the signal to noise ratio increases.
This would lead one to believe that for properly exposed high ISO images in post processing, while you may loose some signal effects by stacking, subtracting bias and dark-thermal frames and adjusting the luminance channel, that you will still come out with a superior image due to the fact you started out with much more signal-effect than noise. However, reality has consistently presented a different situation, where as my ISO increases the best image I can achieve after post processing is in many cases inferior to what was produced at a lower ISO, especially if shooting time-lapse where stacking and frame subtraction are much more impractical. For crop frames I have found that I can never get above ISO 1000 and for full frame cameras, I have a hard time getting above ISO 2500 without inducing granularity that looks un-natural, despite the fact I see many people shooting at 25,000 and higher. I would love to hear comments are explanations from any interested people describing why the higher signal to noise ratio rarely translates to a better image, or rebuttals to this idea if anyone has one.