the problem is that DSP bins pixels on-the-fly, so there is no information in memory about all 38 million of pixels... 808 receives only 5 MP (or whatever resolution you have chosen) from DSP
nice article if you have the time to read it
http://www.imageval.com/public/Papers/20091118 PixelBinningSPIE.pdf
Thanks for the link. This paper raises many interesting questions. I learned a lot and it turned out to be a good starting point for further investigation.
This has been a difficult subject to discuss because pixel binning, resampling, and resizing are frequently used interchangeably, and in such cases when differentiation is attempted, they tend to still not be very well defined. But this paper clearly describes the two methods under study: pre-ADC binning vs. digital resizing. Going by its definitions, I can imagine how the 808 might use pre-ADC binning to reduce the amount of data coming out of the sensor when doing resampling. This would speed things up and could possibly have a noticeable advantage when it comes to noise reduction in low light environments. It would also imply the limitation you mentioned: the impossibility of saving both full-res and resampled images for a single shot. Note that this is pre-ADC binning, which implies that it's not the DSP doing the binning as you say. Wouldn't the noise reduction advantage of binning be lost if you did it post-ADC?
In any case, the paper raises a critical point:
In conditions when there is no significant ADC rate or channel bandwidth limitation, it is advantageous to read the entire high resolution image. In this case it will be possible to resize the image later to produce an image that is very similar to that obtained by binning. A resized image can be further blurred to match a binned image with respect to vSNR and MTF50. There is an image quality tradeoff between noise and blur, and depending on the application users may prefer one error over another. This question is open to experimental investigation through user-preference studies.
Now, Nokia's proprietary implementation and unique set of conditions may somehow exempt them from this assertion, but there's no evidence to support that. It sounds to me like there's probably no quality reason to use pre-ADC binning... only a performance reason. And with the data rates involved, It seems likely that the 1020 is doing this for video. They probably just managed to avoid it for still images this time around, and that's a good thing. It doesn't seem like it would necessitate any sacrifice in quality compared to what they did with the 808.
Finally, this paper led me to an interesting GSMArena interview with the head of the team that developed the 808. He pretty much comes right out and says there's no advantage to the PureView resampling compared to Photoshop, other than his strange view that being stuck with Nokia's preferred tradeoff between noise and sharpness is somehow an advantage (seems quite the opposite to me):
Nokia 808 PureView in focus: Interview with D. Dinning - GSMArena.com
The interviewer tried to help him out by pointing out that saving only the phone-resampled images saves space on your phone while shooting, but he didn't seem too interested in that.