I'm not sure what they mean by on-chip noise reduction. My understanding of dark frame subtraction is that an exposure is taken with the shutter closed, to account for thermal noise, hot pixels etc. The D70 can do this, but of course it requires a second exposure after the 'true' one. By default this feature is off.
They might indeed do some pixel averaging, maybe only in dark areas. I guess this is the sort of thing they want to keep secret, and rely on marketing to pursuade is that it is the best on the market.
Out of interest I do wonder if the RAW file is an exact representation of the image i.e. the Bayer grid readings, or whether the on board chip processes the Bayer readings to create the pixel grid we normally associate with an image e.g. 6MP with each pixel have colour and brightness. If the RAW file is not a direct representation, than I could see that on chip processing might be worth having, as it can work with the true raw data. In other words, Neat Image etc work with processed data.