It has nothing to do with Canon files. Linear encoded data is linear encoded data.
The math is undeniable. End of story.
There are some differences between a Canon with a 12 bit ADC and the medium format backs with 16 bit ADCs. For example, the Kodak 39 MB KAF 39000 chip used in some of these high end backs has a full well capacity of about 60K electrons. At base ISO, the unity gain would be 1 electron/ 16 bit data number.
Unity gain for the Canon 5D with a 12 bit ADC is at ISO 1600. When one is exposing according to light meter readings with the meter set at 1600 ISO, one must increase the ISO on the camera setting to 1600 so that the electronic gain will increase enough to enable the ADC to capture all the information in the image. The 5D has a full well of 80,000 electrons, so the gain at base ISO is 19 electrons/12 bit data number. The granularity of the measurement at base ISO is 19 electrons and this would lead to significant imprecision at ISO 1600 where fewer electrons are collected. At ISO 1600 the 5D would collect 80000/16 = 5,000 electrons in the highlights and half as many in each darker f/stop as one goes towards the shadows.
With the 16 bit MFDB, one could leave the ISO setting of the camera at base ISO and increase the exposure in the raw converter and still get all the information. This is explained by [a href=\"http://www.clarkvision.com/imagedetail/digital.sensor.performance.summary/]Roger Clark[/url] on his web site.
However, signal to noise ration in digital cameras is limited primarily by photon counting statistics. If you double the number of photons collected, the S:N will increase by the square root of 2, 1.414. This applies to Canon and Nikon 35 mm style cameras, MFDBs, and the Hubble space telescope. This is the main rationale behind ETTR. Perhaps you boys with MFDBs are satisfied with your S:N, but you could do better with ETTR. That is a fact of physics and mathematics.
Another consideration is that noise with a 39 MB file will be much finer grained than with a 12 MB file when the images are printed at the same size, even though the noise expressed as a standard deviation of the pixel value might be the same.