First off, I am VERY skeptical of the results provided by DXO-Mark. I have never understood their numbers, and I don't really think their results reflect real-world performance or behavior. They are probably extremely accurate purely scientific results, relative to their own domain, but I don't think that is helpful to normal people doing normal photographic work. My own rather cheap Canon 450D, with its pretty basic, entry-level sensor, was rated as having 10.8 stops wroth of dynamic range, and 21.6 bits of color information. I know that neither of those facets of information are true, as I most certainly do not get 21.6 bits of color information, and I have to work pretty hard to barely get 9 stops of dynamic range...I usually get 7-8 stops at best.
That said, I started getting skeptical with the article when I read the following:
When you look at the structure of CMOS
sensors, each pixel as basically a
tube with the sensing element at the
bottom. If a light ray that is not
parallel to the tube hits the photo
site, chances are the light ray will
not get to the bottom of the tube and
will not hit the sensing element.
Therefore, the light coming from that
light ray will be lost. It appears
from this graph that when using large
aperture lenses on Canon cameras,
there is a substantial amount of light
loss at the sensor due to this effect.
In other words, the "marginal" light
rays coming in at a large angle from
near the edges of the large aperture
are completely lost.
Outside of considerably older digital cameras, all digital sensors these days use microlenses above their pixels. These microlenses are designed to direct off-axis light down into the pixel well. The "marginal" light rays coming from large angles are not completely lost. Some are reflected, some are captured.
For all of DXO's talk about the accuracy of their tests, and their down-talk of camera manufacturers "cheating", they don't really tell their own customers how their own product really works. How exactly are they measuring this light loss? Is it truly accurate?
In my experience, and admittedly I have only used Canon bodies, so I can't speak for others. If I set my ISO to automatic, I get some oddball ISO values in my pictures based on the EXIF data. ISO 160, 240, 320, 480, etc. If I set my ISO to a specific value, it is always that value in the EXIF data. Granted, it is certainly possible for a camera manufacturer to truly try and cheat, tell you it is using ISO 100 when in actuality it is using ISO 200, but it is a little hard to believe they would actually explicitly change the EXIF data to hide that fact from their customers.
It should also be pointed out that ISO "settings" and actual analog readout levels are never in sync in the first place. On a Canon body, an ISO 100 is close to that, but I've seen various tests that indicate the analog readout is anywhere from 80 to 120 depending on the sensor. There have been similar tests for Nikon sensors as well (which probably apply to all Sony sensors given thats what Nikon currently use.)
I don't think the story is as cut and dry as Camera manufacterers are gaming the system. There are physical difficulties in manufacturing sensors that prevent the analog readout from exactly matching the chosen digital ISO setting, fine microlens structures that mitigate a lot of this supposed light loss at the photosite, and fairly advanced algorithms that, to my knowledge, work to maintain the accuracy of the settings you have chosen, not the other way around.
[NOTE: I would like to provide a more accurate description of what DXO-Mark actually does, however, predictably, their site is not accessible at the moment. I'll have to do some research to see if they do offer any detailed specifications or other information about exactly how their measurements work, to see if DXO-Mark are the ones trying to "game the system" as a marketing ploy.]