I've personally tested DR for a number of cameras (the right way). My results match results from sites like DPReview. DxO is always off, and always quite obviously wrong.
dtaylor: How do you test DR? When I use a 42 step transmissive wedge, my DR numbers, assuming a lowest acceptable SNR of 1 (which is what I believe DXO does), pretty much match DXO results for RAW capture within 0.2 EV. And to account for lens transmission variability & different ISO mappings between cameras, I take a number of shots at 1/3EV apart & select the shot of the wedge where the brightest patch is just short of clipping... I then use these shots to build an output vs. input curve, & quantitate the range from SNR 1 to brightest patch that isn't clipped.
Furthermore, your results match DPReview? DPReview doesn't test RAW dynamic range... so I'm confused: what are you testing?
Pretty sure Bill Claff's DR numbers also match DXO for the most part; also he uses a possibly more meaningful number: PDR (photographic dynamic range: lowest SNR of 20 adjusting for circle of confusion of sensor). His tests also show ~2.5EV better performance of the D800. Check out his stuff here: http://home.comcast.net/~NikonD70/Charts/PDR.htm
My own 'real-world' tests also show ~3EV better DR on the D800 when I do side-by-side shots of high DR sunsets with my 5DIII vs. D800; the D800 is able to maintain highlights that the 5DIII needs a 3 or 4-stop reverse graduated ND filter to reign in (while keeping shadow noise similar between the two). Put another way: I have to overexpose my 5DIII by 2 to 3 stops at the very least to get its shadows to match the cleanliness of lifted shadows of the D800 file that was underexposed to maintain highlights.
So, respectfully, I fail to see how DXOs DR & SNR, etc., numbers are the 'odd ball out'.
To clarify: I don't at all mean this to be a personal attack; just looking for clarification. Also, I'm tired of people harping on DXO when they're just trying to do good science (though, like others, I do not
pay any attention to the 'overall' sensor score... jus the raw SNR, DR, etc. numbers).