Why not 16bit?

[Edwin Herdman]

I also question the assumption that DR always scales in a linear fashion.

CMOS sensors, like CCDs, are normally linear devices. Non-linearity (normally on the order of a few %) can be important if you are interested in accurate photometry, but for S/N calculations they are irrelevant. See, e.g., Hain et al. (2007), Comparison of CCD, CMOS and intensified cameras for CMOS linearity measurements.

To confirm - I wasn't talking about S/N, but the scaling of DR along with ISO. Is that what you are referring to?

Again, to link http://sensorgen.info/ - click on any two cameras, and compare the graphs of read noise and DR. For one camera, the graphs do not correspond - and across any two cameras, the graphs are usually quite different.

I will point out that I do not disagree that CMOS and CCD devices are essentially linear; I think what is going on is that gain is not applied evenly at various ISO settings in order to target specific areas of importance to the designers - in the "Hi+" ISO settings, for example, it appears obvious that the goal by that point is to have as clean an image as possible, even allowing DR to take a disproportionate hit.

 

[epsiloneri]

To confirm - I wasn't talking about S/N, but the scaling of DR along with ISO. Is that what you are referring to?

Ah - sorry, no, I misunderstood you. I was talking about the linear response of the detector with time (i.e., twice the exposure giving twice the number of photo-electrons). I assume you are referring to the break in the DR function close to the point corresponding to the unity gain (when one registred photo-electron corresponds to one digital number, DN). For e.g. 5D2 the unity gain is achieved at ~ISO 1600 (e.g. this graph from sensorgen). Going beyond the unity gain results in numerical oversampling of the number of electrons with very little benefit. At the same time you limit the numerical DR.

Basically, beyond unity gain, the DR will decrease linearly with ISO, due to the maximum number of electrons being limited by the numercial encoding. Below the unity gain this limitation in the DR due to the reduction in the maximum number of electrons possible to encode is countered by the number of electrons being better sampled, reducing the numerical noise.

That's why there's no IQ point in going beyond ISO 3200 on the 5D2 (for the 7D, the maximum useful ISO is ~2000). Check fig 6 from clarkvision. That's also why any ISO beyond 6400 for the D1X (with a pixel pitch of 6.9µm) is a gimmick.