At ISO 100, the D4 gives the kind of histogram you’d expect (I’m pretty sure the gaps in the red and blue histograms are due to digital white balance after the ADC:
At ISO 200, there are also no surprises:
At ISO 400:
At ISO 800:
At ISO 1600, a little over twice the Unity Gain ISO, if the noise were less we’d expect to see a decrease in the population of every other histogram bucket. The noise is high enough to swamp that out, but remember that we’re looking at a 200×200 pixel selection, and probably have several amplifiers and ADCs represented:
At ISO 3200, the noise in the red channel is getting large enough to have serious clipping at zero. There’s so much noise that I had to change the upper bound on the horizontal axis:
At ISO 6400, it looks like the standard deviations of the noise have gotten to about 40 counts. Since this is about 10 times the unity gain ISO, we’d expect a periodicity in the histogram about every 10 counts. You need a good imagination to see that.