Size does matter even in the digital age, does it not?
In an important article, written in 1993, G. Crawley of BJP fame (when the BJP still was worthy of study by all serious students of photographic science) compared the image qualities of the 35mm format with that of the medium format. He proceeded as follows. The 24x36 format has an aspect ratio of 2:3. The medium format has a size of 56x56mm. To compare the formats in a meaningful way, one has to crop the medium format to a size of 56x37mm. The frame areas then are 864 sq mm and 2072 sq mm. The smaller size then needs an image (pixel) density that is 2.4x higher than the bigger format for the same information content. The logic is this: if you take a picture of a scene with the larger format camera and you wish to transfer that same amount of information onto the smaller frame, then the information packing density must be higher. Optically this implies that the defining power of the 35mm lens must be much higher to allow for the smaller format. In practice this works, because the 35mm lens designer has to cope with an image circle of 43mm, whereas the medium format designer has to cope with a 83mm image circle.
The comparison was done with a medium format camera with a 75mm lens and a 35mm SLR with a zoomlens set to a focal length slightly longer than 50mm to give the same sectional view of the scene. After exposure and processing the prints were compared and there was no detail present in the larger negative that was not captured by the smaller 35mm size. So resolving power is roughly the same.
But Crawley noted a significant difference in qualitative terms. With more grains to cover the same picture area, the larger negative gave a much better quality of detail, with more subtle gradation changes and tonal depth than the 35mm negative, which had more bite (high contrast) but less subtlety of tonal imaging.
There was a remarkable difference in lens quality: the SLR was a high grade very sophisticated design, where the medium format had a cheap and simple three-element design. The bigger size apparently compensated for the lower optical quality of the lens as the definition of detail in both cases was identical.
Crawley then concluded that there is more to consider than just resolution and pixel count when comparing formats. But that was the old analogue era when silver halide had its own bunch of laws and photographic expertise and knowledge.
I was reminded of this article after glancing over the review on dpreview of the Nikon D200, where one can read that the choice between format sizes with 8, 10 or 13 Mp is one of convenience not of substance. This conclusion was based on a comparison that basically established equality between the three formats in terms of resolution and detail definition. My own testing showed more differences and was more in line with the Crawley article referred above. The question then is: does the old wisdom of the silver halide world still hold and does the test methods of dpreview overlook some aspects that are relevant for the comparison. The dpreview stance is not unique and one can read all over the world, in all kinds of magazines and websites, that size is not relevant anymore (with the possible exception of a much higher signal-to-noise ratio of the smaller sensors).
To study this issue a some depth I made the following comparison. I set up my standard test chart, with a test pattern that has been employed for ages in analog testing and should give fine results in the digital environment too. Then I put the Canon Macro lens EF 2.5/50mm on a Canon 20 D with 8 Mp and photographed the test pattern at such a distance that the chart covered the frame area completely.
The 20D has a sensor size of 22.5x15mm or 338 sq mm. The Canon 5D has a sensor area of 35.8x23.9mm or 856 sq mm. This happens to be a factor of 2.4. So both sensors relate to each other as 35mm versus medium format as in the Crawley comparison above.
The pixel size of the 20D is 0.0064 mm which allows for a maximum resolution of 156 lines/mm or 78 linepairs/mm.
The pixel size of the 5D is 0.0082mm and that gives a maximum resolution of 122 lines/mm or 61 linepairs/mm.

The 20 D has a total of 7.834944 pixels and the 5D has 12.719.616 pixels. The 20D then has more resolution by a factor of 1.3, the 5D has more pixels by a factor of 1.6. The picture of the test chart is as presented left.
Then I switched to my 5D with the same Canon lens and took a picture at the same distance. It is obvious that the lens now covers a larger chunk of the test chart. The factor of 2.4 is clearly visible: this again demonstrates that there is no change in magnification, only in frame size of the image area.

The third picture is now made with the 5D closer to the scene to cover the same area of the test chart, but now using the much larger sensor area of the 5D. Note that I do record the same amount of scene area on a larger image area, exactly as if I were switching from 35mm to medium format.
The results are most intriguing.
By the way: I used the Raw pictures without and with sharpening algorithms. As is to expected there is no difference in detail resolution between both sets of pictures.

First the 20D picture. Looking at the testpattern that is cleanly recorded, we find the pattern with numeric value of 2.8 with 3.2 just resolved. The calculated maximum resolution is 57 linepairs/mm, somewhat less than the absolute Nyquist frequency of 78 lp/mm. Now I enlarge the test pattern to 800% in Photoshop to be able to count the individual pixel rows that make up the pattern. I can count 15 to 16 pixel rows that are needed to capture the 10 lines of the 3.2 pattern. This number of pixel rows has a physical extension of 15 x 0.0064mm = 0.096mm for 10 lines or a real resolution of 52 linepairs/mm. In fact the ultimate we can have is one pixel row per line, so this 1.5 pixel row per line is as close to maximum performance as we can get given al the interpolation that the software has to do. This is all accomplished on the 3.2 pattern, which is in fact beyond the visual resolution threshold. It would be safe to state that the 20D can handle 40 to 45 lp/mm.
Now the 5D picture. The 4.0 pattern is clearly visible, but now we are at a closer distance to the scene, the magnification factor is lower and we have more detail to capture. Counting the pixel rows again, we see the same pattern: here we have 15 pixel rows for 10 lines of the black and white test pattern. With the larger pixel size of the 5D, we can now calculate that the physical size for the 10 lines is 0.123mm or 40 linepairs/mm. 

Again we are at the limit of practical resolution, but because we are recording the same test pattern from a closer distance, we are able to capture finer detail! The 20D stops at a pattern of 3.2 where the 5D can record even finer details at 4.0, while having the theoretically lower resolution limit. This can be explained quite easily by the fact that you can capture the same scene on a larger amount of pixels or higher pixel density per subject element. You see this clearly in the test pattern with square with the horizontal, vertical and diagonal line patterns. The 20D is just able to record that pattern in a recognizable way, but the same rendition by the 5D sensor shows a crisper and cleaner version of this pattern.
And not to be neglected: the larger size of the 5D picture allows fro a larger sized print with equal or even better imagery.
Quality versus quantity.
We may then conclude from the analysis above that there is more than just the numbers to analyze, when comparing lenses and sensor areas. The 8Mp sensor does indeed capture most of the information that is available in the scene, but it looks quite strained where the 13 Mp sensor looks easy going. And we all should know that resolution is not the end game, just one of many image quality parameters.
There is even more to look at. Recently Peter Karbe, head of the Leica Optical Department gave a series of lectures at the Leica Open Days in Europe, where he presented a seminal and in-depth review of the issues at stake when comparing sensor sizes. His main thrust is this: when lenses are designed and used for small area sensors, there is an inherent increase in depth of field that restricts the esthetic use of the aperture as a means for selective focus. Where the physical aperture of a lens has the number of 2.8, used on a small sensor this aperture becomes 8.0 when looking at the DoF extension. So a sensor with a large size (ideally the classical 35mm frame size) is best when one wishes to play with the DoF parameter for esthetic purposes. This fact does show that the large sensor may have advantages that are not easy to test in a comparative way.
The 4/3 debate
In an earlier article I noted that the Olympus four-thirds approach was a most sensible one. With the current flurry of new 4/3 lenses and bodies announced at PMA 2006, the issue might be revisited. I stated that the 4/3 approach allowed for much better optical designs in relatively small physical packages. This is still true: generally the 4/3 lenses are at least twice as good as the normal 35mm designed lenses. The new Leica tagged Panasonic lens for the Lumix L1 body is an excellent design, but good as it is, it has to work with a sensor size that is one quarter of the 35mm frame. However good the lens, you cannot make up for the much smaller size. The Olympus approach is well thought out, and it works quite well up to ISO100, everything above this sensitivity reads noise! And that is the problem.
As I noted in the analysis above, the normal 35mm lens for the Canon system, the 2.5/50mm macro lens performed as expected till the limit of 50 lp/mm. That is indeed the target figure for most 35mm lenses designed for silver halide capture. But for best results on the small sensor sizes, a much higher resolution is required. We saw this above, where the theoretical resolution limit of about 80 lp/mm was not even approached closely. To design lenses to be able to have this level of resolution, they have to be physically larger than current designs. Compare the Canon 2.5/50 with the EF-S 2.8/60mm!
That the 35mm size sensor (as seen in the 5D) cannot be used in an optimal way at least with current classical designs is true: the 2.5/50mm barely resolves 40 lp/mm, where theoretically 60 lp/mm are available. On the other hand we may state that 40 lp/mm is in itself an excellent result and one that is hardly needed in every day shooting requirements, even for exhibition targets.
The upshot
Sensor size does matter (as it did in the past) and the 5D has clearly many advantages above the 20D (and for the record the Nikon D200 too)
Most testing as done in current practice does neglect important issues and must be regarded as too superficial to be able to discriminate between aspects that are relevant for most high-quality results.
Theoretical resolution figures should in practice be reduced by a factor of 1.5 to conform to real life results Resolution figures must be accompanied by qualitative remarks to become meaningful when comparing camera systems
Depth of Field analysis is a prerequisite for the choice for a camera system differentiated by sensor size alone
Current 35mm lens designs cannot exploit the full resolution limits of the modern full frame sensors, but the question is whether the photographer needs this limit for his work
It is a pity that much of the valid photographic knowledge of the past is no longer between the ears of the persons who currently conduct the testing of digital photographic equipment
The larger sized sensors will never get the optimum quality optical designs to exploit the area advantages (this was already true in the past when medium format lenses lagged behind the 35mm design in optical quality)