You would need to break the laws of physics to have a two-stop improvement in noise at the iso12800 in that comparison tool setting! At high isos, the limiting factor in noise is mainly statistical fluctuations in the low number of photons captured (photon shot noise rather than electronic noise) and the signal to noise is given by sqrt(number of photons captured). Modern sensors have been greater than 80% efficient at capturing photons for several years and so there is room for only a tiny improvement, a fraction of a stop, and which is why you have seen only some improvement.
Edit: Michael, your reply was posted while I was writing mine and so sorry to repeat what you said.