GaborSch
wrote:

The question is, how much difference in EV does this difference in

noise represent. To find it out, I searched for a patch in the same

5D2 image with closely comparable noise. I found one with noise

15.55% (very close to the 15.4%), at the intensity -8.29 EV. This

means, that the 5D2 requires 8.62-8.29=0.33 EV more light to have the

same level of noise as the D3X. Measurements on different spots show

like or slightly larger difference, and it can be somewhat greater in

even darker spots.

What you are measuring mixes together a number of different effects. First of all, by measuring the red patch, you are including variations in the reflectance of the source (surface roughness) as well as gradients of the overall illumination in the measurement patch in the standard deviation. This apparently is what Thom meant by "noise in the target". When I've used images such as the ones you are examining to get rough estimates of camera properties, I've used the difference of the two Green subarrays in order to minimize contamination by these sorts of effects. This contamination calls into question why you are reporting 3-4 significant figures in your measurement results.

Second, since the patches being measured are not particularly close to the black point, there is a significant contribution from photon shot noise. How much photon shot noise there is depends on how much light is being gathered -- the sensor quantum efficiency. The 5D2 is similar to the 1Ds3, whose efficiency is not far from the D300 which is a close cousin of the D3x. So let's for the sake of argument assume that the two sensors have roughly equal efficiency. This means that the photon shot noise for a given exposure value will be roughly equal.

At a given EV, the total noise will be a combination of read noise and shot noise, combined as RMS. The 5D2 has substantially higher read noise for low ISO, at least according to DxO. This is why the SNR plots are much steeper in shadows for the 5D2 than the D3x -- the D3x is shot noise dominated down to much lower exposures due to the lower read noise at low ISO (the slope is 1 for pure read noise, 1/2 for pure shot noise, and in between when there is a combination of the two).

Because the slopes of the SNR plots are different, and assuming as we have that the signals S are roughly the same, compresses the distance in S one must travel to find a value of N which is the same for the two cameras when S is not very close to zero. Your measurements are using S values of a few tenths of a percent of saturation. If one for instance looks at the D3x at about .4%, the SNR is about 22, and for the 5D2 it's also about 22. On the other hand, if one really goes close to the black point on their plots, the 5D2 hits SNR=1 (0dB) at about .027%, while the D3x has DNR=5.5dB at this point; the D3x hits SNR=1 at about .015%. That's about a one stop difference near black.

Your measurements are substantially affected by photon shot noise, which hasn't been accounted for in your analysis; and I doubt that the RAW files you are using are sufficient to back out the photon noise to get at the substantially smaller read noise underneath, which is what controls the DR near black.