Pixel noise actually increases a little bit at ISO 1600 at the pixel level.
I am sorry, I posted my original reply when it was mal-formatted and had some errors in it, and I have tried to edit it but it looks like the forum's edit function is inoperable at this time. The edits never send.
So, let me put the body again here, and I will get rid of the body of the other post when I can edit again:
I find it interesting that the 40D does not appear to fare any better, noise-wise, than the (already excellent) 30D - again, Canon sacrifices advances in high-ISO sensitivity to resolution. Then again, I guess, they have 'improved' it since they kept noise constant despite an additional 2 Megapixels.
Pixel noise actually increases a little bit at ISO 1600 at the pixel level, but not little enough to be cancelled by the extra pixels. Banding seems to be a bit weaker though, so in some ways you might say the noise isthe same or a little less for the 40D. ISO 100 noise is much lower, though, relative to the RAW clipping point; down to near-1D* levels. There should be about 1/2 stop more RAW DR at ISO 100.
Am I correct in assuming that the 1D MkII / MkIIN still performs better at high ISO?
They have the same RAW pixel noise levels as the 20D/30D at ISO 1600, and 1/2 stop less at ISO 100, so they are about similar for image noise at 1600, and slightly inferior to the 40D at ISO 100, electrically. Optically, however, they have a weaker AA filter and have a higher pixel pitch, so they are generally sharpened less and therefore have their noise sharpened less as well. Lots of factors here; no simple answers.
(because of the larger photosites).
The larger photosites gave no photon advantage to the 1D cameras at any ISO as opposed to the 20D and 30D, because their quantum efficiency is rather low. With the newer series, however, it looks like Canon is starting to hit a wall with their multiple-transistors-at-each-photosite technology. The mk3 had room left for improvement, the 40D did not. The mk3 maintained full well capacity per pixel over the mk2; the 40D does not, over the 20D/40D. The 40D seems to drop full well capacity inversely proportional to megapixel increase (pixels increased 1.25x, photons per pixel divided by 1.25).
I still do wish that, whatever Canon does with the replacement to the 5D, they'd keep resolution the same and improve the noise performance substantially (keep the niche as a great low-light camera). I'd hate to see a 16Mp model that has the same 'per-pixel' performance as the outgoing model - ala the 40D situation.
Well, it would take 26 MP to fill a 36x24mm frame with 40D pixels. I really don't think that a 26MP fullframe with 40D pixel quality would be something bad. Those extra pixels result in less image noise than the pixel noise would suggest. Look at my Panny FZ50, looking at it at 100% zoom, with minimum noise reduction in the JPEGs, I can see chromatic noise in the midtones at ISO 100. Downsample properly (NO NEAREST NEIGHBOR OR HYBRIDS!) to fill the 1600*1050 LCD screen, however, the noise completely disappears, as far as can be seen. The same thing happens when you print or view a 26MP image at the same size you'd view a 12MP; the extra noise of those pixels is probably so much less that the increase in pixels results in lower *image* noise.
Try it yourself; take a 40D high-ISO image, downsample it properly to display it, or print it at 39% the area that you'd normally display a full print. The noise at the viewing level will be less than it is when viewing most other cameras' images at the fuller size. Or stitch together enough pixels from 40D images to make a 26MP.
On the other hand, a 16MP or so model, with the same excellent performance as the 1D MKIII, and with an 8MP or so 'Small RAW' mode, should do the trick nicely for those of us that don't want higher resolution, but even better ISO3200! I imagine perceived noise in a SRAW image is very good, because one is effectively 'binning' pixels?[/QUOTE]
Keeping the pixel pitch coarse does still have benefits for ISO 1600, 3200, etc, because it gives Canon more room to place high-gain transistors and better quality traces. The trade-off is that there seems to be a very hard limit with current technology concerning how low you can get read noise at ISO 100. A full-frame sensor filled with pixels and circuitry from an FZ50 would have more image DR than a 16MP 5D2 at ISO 100, if you compensated for the extra pixel readout with parallelization (to keep the pixel read noise as low as it is on the FZ50, which is actually quite low). Lacking optimization for high ISO, however, the FF filled with FZ50 pixels would have about a stop higher *image* read noise than what the 1Dmk3 is doing now at ISO 1600 and 3200.
I just want to make it clear that when I speak of image-level DR, I totally reject the idea as popularized by Roger Clark that pixel noise relative to pixel capacity directly determines RAW image DR. Pixel DR has to be scaled by the displayed pixel pitch to describe image DR properly. The only problem with a lot more pixels with slightly more read noise each (shot noise has no relevance here, as shot noise never results in photon captures less than zero photons, while read noise results in negative "photons") is that if you clip the RAW data at the blackpoint (and most brands of DSLR do this before they give you the RAW file - not Canon, though), there will be a non-linearity in the mean luminance as you approach optical black. This, however, can be profiled for cameras based on their read noise level and compensated, and negative noise can actually be fabricated (by making a certain percentage of pixels with a RAW value of zero a negative number) to keep the mean closer to linearity near black, allowing converters to generate color at a lower resolution without biases due to white balancing.
As you can see, I have a lot of thoughts on these subjects, but they're probably hard to follow. I've simulated or tested everything I've said here, however, and found it all to be true, even if sometimes subtle in effect.
In summary, my main point is that ISO 100 image quality can improve with more and smaller pixels (even when each has more noise relative to saturation), while Canon-style high ISO may already be hitting a noise wall.