Noise Primer by Eric Fossum

Eric Fossum posted a nice 4T pixel noise explanation on dpreview forum:

"In an active pixel sensor with intra pixel charge transfer, the photodetector is pretty much decoupled from the floating diffusion and SF amplifier. The latter's sizing is usually determined by design rule (e.g. 0.25 um or lower) whereas the former is determined by sensor physical size and resolution, unless it is for camera phones in which case the photodetector is made as small as possible yet still have decent QE, full well, etc.

So, true, gain in the SF is inversely proportional to the FD capacitance which in turn scales like the area of the SF gate and FD node, and parasitics, and sometimes intentional capacitance.

In this amplifier, the kTC noise is mostly removed by correlated double sampling. In theory, totally removed, but only if the sampling is perfect and the SF transistor itself has no noise. In fact, that transistor does have several noise sources. The dominant one these days is the so called 1/f noise. Actually, it is only loosely 1/f. The noise is believed to come from fluctuations in the carrier density under the gate. As carriers get trapped at the Si-SiO2 interface (or within a short distance thereof) they partially screen the gate field and change the electrostatics of the structure. The resultant change in semiconductor carrier density results in a change in current. The trapped carriers dont stay trapped forever. It depends on the trap energy and the distance of the trap from the Si SiO2 interface with lots of time and temperature dependencies. So, carriers get trapped, carriers get released, and the current thru the source-follower fluctuates "randomly" in time. If you are lucky enough you can see the effect of a single trap being filled and emptied as the current fluctuates between 2 different levels (small fluctuation mind you). This is the "random telegraph signal" or RTS noise. The ensemble of traps can lead to 1/f like behavior in the spectrum of the noise.

Actually, trap densities are of the order of 10^10 per cm^2. So, a gate that is, say, 0.25 um long and 1.0 um wide is 0.25x10^-8 cm^2 and contains a handful of traps, on average. So, it is easy to see that some transistors may be more noisy than others depending on the number and type of trap present.

The effect of these traps increases as the gate area decreases and 1/f noise increases with smaller transistors. So, to the extent smaller pixels have smaller transistors (not necessarily true, of course) then smaller pixels would be noisier than larger pixels.

This depends on how you measure the noise. I have been referring to fluctuations in current around some nominal operating point, relative to the nominal current. If you measure the current fluctuation as an effective gate voltage change, and then relate that via capacitance to an effective number of electrons, the read noise could go up or down or stay constant. It depends on how the other things scale. You could also measure noise in input-referred electrons as a fraction of full well, in electrons. This would then bring the photodetector into play if the full well was limited by the detector. The full well could also be clipped by a combination of low sense node capacitance (FD) and low voltage swing (amplifier limited full well). Usually the design tries to match these things.

1/f noise also depends on the process, and design, and also biasing. It is generally lower for p-channel devices compared to n-channel, but then to get teh same current drive, the p-channel has to be larger since the hole mobility is like half that of the electron mobility. Buried channel transistors have lower 1/f noise than surface channel but have higher operating voltage requirements or have lower swing for the same operating voltage. Many tradeoffs are available to the designer. Also, these things evolve in time esp. once they become limiting factors. Nothing like hundreds of engineers trying to fix a problem.

So, combing the literature and comparing noise and pixel size is not going to give you definitive answers. It is hard to get consistent results in your own fab doing your own test chips from run to run. Apparently the phase of the moon and sunspot count, and butterfy activity in South America all count.
What was the question? Oh yes, what can be expected.

Sub electron read noise has already been demonstrated in rare cases, and a few electrons in several cases. Low read noise is not so important is DSLRs because most photographers are going to be limited by shot noise or something else rather than read noise. However, for light-starved small pixels in camera phones that people want to use in poorly lit bars, low light is a persistent problem and low read noise is always desired."