It is a sad state of affairs when presented with direct evidence...
I present a careful and reasonable calculation...
Where, exactly, is this evidence you've presented? Looking above, I see none. Calculation? Aside from model numbers, the only numerical part of your statement is "zero improvement."
If you'd like present a well-reasoned case, backed up by actual evidence, fine. Else, it's just opinion and hot air. I don't shoot messengers who bring bad news, as long as that news is accurate. But an unsupported (and knowingly inflammatory, at that) conclusion deserves the bullet.
Load the RAWs into IRIS (it recognizes 5D2 files completely and masks off their masking area so you need to use a black frame from the 5D2). For the 5D3 select a big chunk of the left masking area (the top masking area appears to have some data or something stored in it and won't work) and hit the statistics option. You find that the avg value is 2048 and that is the black level. You then see a standard deviation of about 5.95 or so (I was actually a little generous if anything since a patch higher up that I sampled said 6.0). Then you draw a box around the blown highlights in one of the IR ISO 100 images and notice that the raw saturation level at ISO 100 is 15283.
With those three numbers you do this Log[2] ((15283-2048)/5.95)) and get 11.1 stops.
For the 5D2 you find values 15760,1024,6.1 and get 11.2 stops.
banding looks worse on the 5D2 black frame but it's hard to compare to the tiny area of the 5D3 masking, the 5D3 looks to have noticeably less banding there, but the area is too small to really be able to much for sure, it shows promise, but it remains to be seen.
You can set the pop theshold and base values to a narrow range around 1024 for 5D2 and 2048 for the 5D3 to get a good look at the dark current banding and noise characteristics between the two in a fair way (other than nobody has provided a 5D3 black frame so the area is too small to be able to count as representative, the banding might show up more if you got to see a large chunk of the main area).
The masking area that I talk about is the portion of the sensor that does not get exposed to light but that does get recorded to the RAW files, the actual RAW files are slightly larger dimension than the exposed area size that most people are familiar with. The top masking area is unfortunately not quite as large as would be helpful to better aid RAW converters in removing banding.
Anyway, that is how I did it.
On DPR some others have also done it (using different software in some cases) and they got the same things to a tenth of a stop or two.
There is always a chance that the masking area will not match the main area, but looking at it more closely I see no sign of anything weird in the left masking area and I would be surprised if it turns out to not be representative. I hope I am wrong though since who wouldn't want the real result to shot a stop or two improvement here.
