@Mike Risedal: Thank you for sharing! I am not a tech and non-anglo as well. But do I understand correctly, that Sony sensors apply kind of a black frame ( as we do it in long exposures at night) to get rid of all the noise before ever a new photograph is taken? Well, this allows much more noise free high ISOs then...So, when is it likely that Canon apply a similar system? My 5D3 seems to be a good investment until they are able to present a similar solution.
Every camera applies CDS, or Correlated Double Sampling. The point of CDS is to determine the reset dark current noise present in the sensor, before an exposure is taken. The charge of each pixel is registered and saved, and when the actual exposure is read out, the registered reset charge is subtracted from the exposure charge. That, effectively, eliminates noise caused by dark current present in the sensor at reset time. That is a moderate mount of noise, and exists in all sensors. It is not, actually, the worst form of electronic (read) noise, so CDS only does a partial job of removing noise in hardware.
At the moment, as far as I know, Canon only employs analog CDS, via circuitry in each pixel. The use of an analog register makes their implementation of CDS succeptible to heat, charge leakage, etc. So it is not perfect, and the registered dark current noise can be "infected" by other sources of noise. Sony Exmor, on the other hand, employs Digital CDS. A reset read is performed via CP-ADC after each exposure, clearing the sensor. CDS is still employed, however instead of registering the reset charge of each pixel in analog form in CDS circuitry, each pixels reset charge is read, converted via the ADC, and stored digitally in a memory bank per column, associated to each pixel of that column. Since the reset charge is stored digitally, it is effectively immune to infection from other sources of noise such as current leakage, heat, etc. Again, that is only part of the story.
Other sources of noise, which tend to produce more prominent noise, are the ADCs themselves, and output differential from multiple ADCs. Additionally, if the ADCs are off-die, then the analog signal from the sensor itself has to travel along a bus, which can introduce its own noise. The general concept with Exmor is to turn analog information into digital information as early as possible. Digital information is bits, 1's and 0's, which are more "resilient". Error correcting data channels can transmit digital information in a reliable manner via, say checksums, and if a chunk of digital data is received that does not have a matching checksum, then it can be resent. Thus, a digital signal is always "pure". An analog signal can be "infected" during transmission, and thus can never be truly "pure".
It also aims to do so in as highly parallel a nature as possible. High frequency components tend to generate more electronic noise, and when you have fewer ADC units, each one has to process more pixels. By hyper-parallelizing the ADC with its CP-ADC patents, Sony is able to run each ADC at a much lower frequency, since each one only had to process, say, 4000 pixels (one column) rather than 4000 * N. Say, in a 6000x4000 sensor, there are only 16 ADCs...that would be 6000 columns / 16, or 375 columns processed per ADC...a total of 1.5 megapixels per ADC. An ADC responsible for processing 1.5 million pixels per readout needs to operate much faster to be capable of processing at a high enough rate to support the frame rate of the camera, than an ADC responsible for processing 1/375th of that much (4000 pixels).
By switching to a CP-ADC approach, and converting the analog signal to a digital signal at the earliest possible opportunity, Sony Exmor has made their readout more immune to infection by noise. They leave the shortest window of opportunity to allow noise to be added to the signal, this eliminating the majority of it "by default". Additionally, slight differences in ADC operation can cause banding. I have not read any explicit indication that Sony CP-ADC does this, but I believe each ADC is able to determine differences with its neighbors and eliminate any non-uniformity related noise as well (and, thus, reduce banding to a level where it never exhibits in their images.) Other forms of NR could also be performed digitally, such as PRNU (pixel response non-uniformity). The use of low-frequency parallel ADC also eliminates a prime source of heat (high frequency ADCs operate at a higher temp), thus reducing thermal contributors to noise.
Can Canon get around this patent? Well, as Mikael states, sure...they could bucket rows instead of having one ADC per column. Is that good enough? Well, parallelization is only part of the story. The real source of the low-noise operation is the DIGITAL NATURE of Exmor sensors. Column-Parallel is really more of a speed thing than a low-noise thing...by reading each column out in parallel, you can do more work in less time at the same frequency as an off-die ADC that processes 375 columns. If the on-die ADC's operated at double, triple, quadruple the frequency they operate at now in say the D800 sensor, you could achieve extremely high readout rates. You would also experience an increase in noise, however thanks to the Digital NR it is unlikely that even a higher frequency CP-ADC would produce as much noise as purely analog readout systems.
Whether Canon can get around Sony's patents for Exmor really remains to be seen. A lot of the same concepts can be employed via analog readout...column-parallel read, CDS, non-uniform response normalization, etc. could all be employed in an analog version of CP-ADC on a Canon sensor. But such a readout system would still be an analog signal, and still succeptible to infection by noise during transmission from the sensor to the image processor. Canon might be able to combine an image processor right onto the sensor die, or stack them. That would reduce the transmission distance. It would also likely increase heat, and heat is a contributor to noise in an analog signal. I dunno...at the moment, it seems like Canon is in a tough spot to really directly compete with Exmor. They could probably compete with other manufacturers using various forms of bucketed parallel readout...20 rows per on-die ADC, stuff like that. They will still experience banding (as the Nikon D5200 does), but it would certainly be a step in the right direction, and hopefully a significant improvement over their current sensor technology.