Originally Posted by nrdavis
The latter. On previous cameras they did the full, normal demosaic process (using all pixels), then downsampled to the lower resolution (still using all pixels). I would guess they are doing the same on the 7D.
Originally Posted by nrdavis
With downsampling, the noise can never be better, but with the ideal downsampling filter, the noise will be the same. Many slightly imperfect filters, like all the ones Canon had before, increase the noise slightly. It also causes an increase in aliasing.
However, downsampling isn't the only option. It would be possible for Canon to use on-sensor binning (connecting multiple pixels together electronically), and that would theoretically reduce read noise (shadow noise). However, it may require additional transistors on the pixel, and they would have to solve some other difficult problems such as aliasing (much worse than aliasing from downsampling).
Most people find sRAW to be perfectly acceptable, I think you will too.