as the op of this conversation i got lost about page 3. all i was really asking was the possibility of a 16 bit sensor on a dslr a better option than say the rumored nikon d4x at 54mp. i also read on this conversation that 16 bit really isn't 16. some of you have said that 2 to 3 bits is, if i read correctly, noise.
Hi David,
That is correct, and the noise is coming from several sources.
Suppose that a sensel could collect the charge of 60000 converted photons then, by the random arrival rate of photons, photon shot noise will be sqrt(60000) = 244.95 photons (0.408%) at the high exposure end. That's not much, but will become noticeable when we start reducing the level of exposure. Cutting the exposure in half (or doubling the ISO ) will leave a maximum of 30000 photons with a shot noise of sqrt(30000) = 173.21 photons (0.577%).
By the time we have reached an exposure level of 14 stops below maximum (= deep shadows) there will be only 3.66 photons of exposure recorded on average, with a photon shot noise of 1.91 (52.26%). At 15 stops below maximum the signal will become 1.83 photons with a noise of 1.35 (73.9%). That is only caused by the random nature of light particles as they build up the exposure over continuous time. Note: discrete particles,
continuous time, hence light is not a digital but an analog signal. It becomes digital after quantization.
You can imagine that
any additional electronic noise from circuits that process (e.g. read-out and dark current) that already noisy signal, and slight amplification differences between the per sensel transistors (PRNU) will immediately wreak havoc on the already marginal signal/noise ratio. It will be very difficult to exceed 14 bits of real signal.
Only by reducing that additional (read, PRNU, etc.) noise (e.g. by super cooling and very high component quality), or increasing the storage capacity (well depth) and exposure time, can we approach between 15 and 16-bit signal accuracy per sensel. With the shrinking sensel sizes, I think that 14 to 14.5-bit signal accuracy will be the practical limit, for which 16-bit components would be required (14-bit components will be too close to the best signal level with no room to spare).
do you also lose the same amount on a 14 bit sensor such as the d800/e?
The D800/D800E can approach 14-bits of actual signal with a well depth of some 45000 electrons, but it clips part of the Read noise before writing the Raw data. So the maximum signal is something like 15.46-bit, minus 1.4-2 bits of noise.
Do note that this is all before demosaicing, gamma, and tone curve adjustment, which may amplify or reduce the visibility of noise in the final image.
Cheers,
Bart