These statements are hardly contradictory at all. You have it backwards: read noise is lower at high ISO than low ISO. As discussed in the Kodak white paper below, read noise has two main components: white noise and flicker noise. In the equation for white noise, the amplifier gain is in the denominator and hence greater gain (used at higher ISO) results in less white noise. The flicker noise has to do with with speed of the readout--it is greater when the pixels are read out faster. The discussion is for CCD, but the same principles apply to CMOS (according to Ron Parr).
Maybe I have got it backwards. There's a lot of misinformation on the net. Ultimately, one can only go along with what makes sense and what gels with other concepts one thinks one has understood. The idea that a signal can be amplified with a result that noise is less in absolute terms rather than relative terms does not make sense to me, in the absensce of a noise reduction technique. Yet the formula you refer to in the Kodak paper certainly suggest that this is the case. I'm not a mathematician but I cans see that a numerator enclosed in a square root is not going to escalate in value greatly and a denominator consisting of what is essentially a constant (efficiency) multiplied by a gain figure, can only have an effect of reducing the over all value as gain increases.
Having just looked at Roger's table of measurements you've provided a link to, it's clear he is not measuring the readout noise of equal photon counts at different ISOs. As ISO increases, the absolute value of readout noise is shown as falling, but dynamic range and signal-to-noise falls much more dramatically. As you can see from my example images, in situations where the photon count is the same, the higher ISO image has
better S/N and DR.
What appears to be happening here (it's the only interpretation I can think of) is the signal being read is approximately the same level at all ISOs. The difference is, the signal at lower ISOs has resulted from a higher photon count combined with lower gain. It sort of makes sense that a signal that has been amplified in a controlled fashion might be easier to read than a signal that is just stronger initially.
Nevertheless, I still find it difficult to accept that there is no additional noise reduction going on at high ISOs. If one looks at the history of Canon DSLR development, there appears to be no dramatic improvement in S/N and DR at base ISO. The 1Ds actually had slightly worse noise than the D60 at ISO 100, on a pixel for pixel basis. Nor did the later 10D have better noise characteristics than the D60 at ISO 100, but it certainly did at ISO 400 and above, and this improvement at high ISOs has continued with the 20D and 5D, so it seems clear to me that that simple formula you refer to in the Kodak paper is not telling the whole story. I would also find it difficult to believe that that equation represents any recent development in preamplifier and/or CMOS chip design.
Whilst doing a bit of research on Google, I came across an interesting thread on Photonet, which addresses some of these issues. The article
here is mainly about dark frame subtraction during post processing, which doesn't appear to produce consistent results because this sort of thing is best done in-camera prior to demosaicing. However, the author, Jeff Medkeff, seems convinced that modern CMOS imagers have on-chip noise reduction devices for readout noise which he refers to as 'bias noise'.
Following are a couple of relevant quotes from his article.
Bias noise is also highly repeatable - but since it is a result of reading out the sensor, it does not even depend on shooting conditions being the same. Practically the only variable affecting readout noise in a digital camera exposure is the amount of amplifier gain. As long as the amplifier gain remains the same, readout noise will be nearly identical from shot to shot. In general, doubling amplifier gain can be expected to approximately double the amount of readout noise.
Is Jeff dead wrong here?
CMOS sensors allow the placement of both photosites and transistors on the sensor itself. (CCDs cannot have any processing circuitry built into the sensor - just transfer gates and the like, which are controlled by off-sensor control circuitry.) Because of this, CMOS sensors generally have at least the readout amplifier built in to the photosite. There may be other transistors as well, which perform other processing steps. It is now very common for a CMOS sensor to include noise-reduction circuitry directly on the sensor alongside the readout amplifier. In some designs, a sort of small dummy photosite, shaded from light, is used to quantify the likely dark noise level in the actual photosite, and this quantity is subtracted during readout. In other designs, a constant - corresponding to the tested dark current of the sensor - is subtracted from the photosite value during readout. If anything like this is happening, expectations such as "dark noise will double with twice the exposure duration" may turn out to be false.
This makes sense to me but what about the next statement?
In addition, this on-sensor circuitry can be designed to subtract the amount of bias noise that the sensor designer expects will be contributed to that particular pixel. This is a design-time decision, so bias noise may still be introduced due to manufacturing variations, erroneous expectations on the part of the designer, changes in other circuitry at a later point in development that the designer decided not to compensate for, and so forth. In any case, if bias noise is being addressed in a CMOS sensor camera - and it is being aggressively dealt with in all known current DSLRs - the relationship between ISO and readout noise in a particular camera's images might not be as simple or as repeatable as expected.
Finally, there's one factor which might have a much greater effect (than lower read noise) on noise reduction of equal initial signals at high ISOs, and that is the number of bits available to describe the signal at the A/D conversion stage. Whether the signal is pushed to the right as a result of greater photon count or greater in-camera amplification, it's pushed to the right nevertheless and more levels are available during digitisation.