Maybe we should separate circuits and settings ... the fact that the circuitry is there does not mean that the firmware allows the top counts to go all the way through the shoulder into the raw files.
Hi Edmund,
Indeed, circuitry (some of which is not present in all designs)
is different from firmware settings. The settings used in the ADC process will determine how much of the total captured signal is transferred to discrete values. Noise is added to the already shotnoise riddled signal during that process. Some conversions are therefore cleaner than others.
In addition, e.g. Sony have started using a non-linear encoding of the in essence linear ADC output in their Raw format. So now it seems like the camera's response curve is vastly non-linear if one were to look at the Raw data in isolation. However, during Raw conversion this added non-linearity is removed again, because demosaicing requires a linear gamma data space.
My take on digital SLR cameras, after issues with highlights on my D4, is that the top stop of a saturating channel is problematic due to non linearity, the bottom 2 stops also due to noise and amp mismatch, 1 stop is often lost to channel mismatch eg. incandescent or crepuscular illuminant, so a 14 stop camera is a 10 stop camera under less than ideal conditions, even at low ISO. Exposure lattitude is thus 1 stop each way from the center, move more and you will see damage effects at one end of the histogram. I'm sure my reasoning is wrong somewhere, but this corresponds to what I see in practice
There is probably also a Raw conversion effect in play here, which is not a sensor characteristic. I rarely look at the Raw data, unless I need to analyze it before the demosaicing and conversion process turns the data into an image. During that process a lot can happen to the tonecurve, especially in the LR and ACR Process 2012 (which significantly compresses highlights), or in Capture One when a filmcurve is used instead of a linear tonecurve. But that is all postprocessing, and is not a sensor characteristic.
Interestingly, I (most people?) seem to be very sensitive to near-white highlight hues, so non-linearity of a channel at the top means bad color discrimination in the clouds and highlights and me unhappy; on the other hand, a shoulder does mean no hard burnout. I don't think ETTR really makes sense because of this, as eg. sky and cloud nuances will look very different from reality if a channel goes non-linear at the top
I'm also sensitive to highlight rendering. Using a linear tone curve instead of compressing highlights helps a lot, no, it's essential. It's easier to add a shoulder roll-off, only if and when/where it is needed, than to create linearity where the detail has already been lost. A fantastic tool like Topaz Labs Clarity helps a lot to regain some sparkle in the image by altering local contrast in a very intelligent (and halo free) way, if necessary to be applied locally by clever built-in masking tools, but it won't invent detail where there was none to begin with.
Cheers,
Bart