As I expected... thanks for doing that Julian.
Several of us have been testing cameras back to back next to each other for years now, and have seen this trend:
1) Each generation seems to manage noise better at the processor level (software/firmware, what-have-you), not necessarily at the sensor hardware level
2) The only changes Canon has made to their sensors were changing the well sizes and then the gaps between each sensel, if I am using the correct terms. Nothing else radically has changed over the years.
3) Resolution differences between different formats across different generations forces us, the consumer, to try to figure out the real performance differences by resizing images one to the next, or trying to change the focal length to keep the same number pixels on target between two formats as best we can, etc. That is almost always an epic fail.
These ISO performance numbers that we throw around is much like car performance parts, that air cleaner gets you 10hp, the exhaust 40hp, the throttle body 15hp, and so if you do all 3, you get 65hp. NOT....
If the 5D2 is 1.3 stops better than the 7D, and the 1D4 is .5 stop better than the 5D2, and the 5D3 is .5 stop better than the 1D4 and the 1DX is 1 stop better than the 5D3, then the 1DX should be 3.3 stops better than the 7D.
As evidenced, it just isn't so. What is really the truth is that the newer generations are more consistent at getting great high ISO performance across a range of situations, so you can be more sure of your results and processing is more consistent afterwards. That is my opinion anyways. Flexibility at missing exposures and more consistent noise results at high ISOs is a great thing because you can get your bearings on how to shoot what and how to process it afterwards. Also, Canon is getting better at massaging those high ISO results from the sensor into something much more usable.
For fun, it would be interesting to see an ISO 3200 difference from a 20D to the 1DX.