Human vision can distinguish approximately 100 shades of tone in each colour on average---a bit less (60 - 80) in the violet-blue and in the deep-red ranges (i. e. at the ends of the visible spectrum), and a bit more (120 - 130) in the yellowish-green range (i. e. at the middle of the visible spectrum). So under ideal viewing conditions we may just barely see a difference between 6 bits and 7 bits. The usual 8 bits already provide us with considerable headroom by a factor of 2 - 3.
12 bits provide up to 4,096 shades of tone per colour which is 40× more than we can ever hope to be able to see with our eyes. More bits mean more native dynamic range in the A-to-D converter which basically is a good thing ... however we may just as well spread the range by multiplying as we already have so many shades. Not-too-small photosites, cleverly designed photosite read-out, A-to-D conversion, noise reduction, raw conversion, and image processing are way more significant to the final image quality than the A-to-D converter's width being 10, 12, or 14 bits.
As Olaf's analysis demonstrates, the increased number of levels provided by increasing the bit depth from 12 to 14 is of limited or no value, since the human eye can not distinguish these additional levels. Moreover, the number of actual levels with current cameras is limited by noise. Even with perfect sensors and analog to digital converters, photon sampling noise (shot noise) will limit the number of discrete levels that can be distinguished in the image. Emil Martinec gives a good analysis in a [a href=\"http://forums.dpreview.com/forums/read.asp?forum=1021&message=27646854]thread[/url] in the DPReview Nikon D3 forum.
Bit depth does impose an absolute limit on dynamic range, which is limited to 1 f/stop per bit as explained by Sean McHugh on his web site
. However, this maximal bit depth is not achieved with current cameras because of noise. Also, this maximal DR has only one level in the darkest f/stop. In practice, you would most likely want more levels.
The ideal SNR of an ADC equals 6.02N+1.76 dB, where N is the number of bits (explanation
). A 12 bit ADC has an ideal SNR of 74 dB and a 14 bit ADC has an ideal SNR of 86 dB (6 dB represents a doubling of SNR). As Roger Clark explains in his digital sensor analysis, current 14 bit ADCs do not attain the ideal, since it is difficult to design an ADC approaching ideal at the high bit rates required for cameras like the Nikon D3 or Canon EOS 1D MIII.
In summary, while a 14 bit ADC has theoretical advantages, these are not attained in practice with current ADCs.