Nikon D810 Dual Amplifier Read Noise
Prepared 2014-12-20 by Bill Claff

Newer Nikon cameras take a new approach to amplification (conversion gain); they use two amplifiers rather than one.
Apparently it's possible to use two amplifiers in series, each with a lower top gain, and produce less noise than a single higher gain amplifier.
The approach is similar to DR-Pix (see white paper) except that the first amplifier in the chain is variable rather than simply being either Low or High.

Although good news for photographers; but it complicates the work of people, like myself, in trying to characterize sensors.
Because of variability of how the conversion gain is divided between the two amplifiers, the analysis is not simple!

Just a reminder that the key noise equation looks like this:

output_noise = sqrt((input_noise * conversion_gain)^2 + conversion_noise^2)  "the equation"

And with two conversion gain stages we have to consider that the output_noise of stage one is the input_noise of stage two.

The Nikon D810

I'll present my Nikon D810 findings.
For this article I'm using 14-bit values and I'm only showing one of the green channels for simplicity.

The analog conversion gain range is from ISO 60 up to and including ISO 1000.
From ISO 60 through ISO 250 the first amplifier remains steady (probably it increases so slowly that I cannot detect the change).
From ISO 320 through ISO 1000 the gain of the first amplifier slowly increases requiring less gain in the second stage.

The input noise at the pixel is 1.056e- and the conversion noise at the pixel is 3.654e-.
When the first stage conversion gain is 1 and the pixel output noise is 3.804e-.
Above ISO 250 the conversion gain increases approximately 1/20EV for each 1/3EV of ISO and the pixel output noise rises.

Conversion noise at the second amplifier is 0.696DN.

This table summarizes the calculations for all of the analog ISO settings:

 model measured pixel conversion conversion intermediate conversion conversion read read ISO noise gain noise noise gain noise noise noise e- e-/e- e- e- DN/e- DN DN DN 62 1.056 1.000 3.654 3.804 0.208 0.696 1.055 1.050 79 1.056 1.000 3.654 3.804 0.262 0.696 1.217 1.215 100 1.056 1.000 3.654 3.804 0.331 0.696 1.438 1.458 125 1.056 1.000 3.654 3.804 0.417 0.696 1.731 1.743 158 1.056 1.000 3.654 3.804 0.525 0.696 2.114 2.135 200 1.056 1.000 3.654 3.804 0.661 0.696 2.610 2.596 251 1.056 1.000 3.654 3.804 0.833 0.696 3.245 3.200 317 1.056 1.035 3.654 3.814 1.014 0.696 3.930 3.927 400 1.056 1.072 3.654 3.825 1.234 0.696 4.772 4.796 503 1.056 1.110 3.654 3.837 1.502 0.696 5.805 5.842 634 1.056 1.149 3.654 3.850 1.828 0.696 7.072 7.103 800 1.056 1.189 3.654 3.864 2.225 0.696 8.623 8.606 1007 1.056 1.231 3.654 3.878 2.707 0.696 10.523 10.479

You can see by examining the table or the following chart that the model read noise and the measured read noise agree quite well.
(This is a log-log chart and the small discrepancies are generally well within a 1/6EV tolerance.)

Conclusion

I have demonstrated that the read noise of the Nikon D810 follows the model of two amplifiers chained together.
This technique improves high ISO read noise but covers a much small range of ISO settings than the DR-Pix technology applied to the Sony A7S.

The analog range for the Sony A7S was from ISO 160 to ISO 25600, a factor of 160.
The analog range of the Nikon D810 was from ISO 60 to ISO 1000, a factor of only 16.

It's not clear why Nikon took this approach in the D810. Perhaps we'll something more DR-Pix like in the future.