According to Thom Hogan ( http://www.bythom.com/nikond3xreview.htm ) the D3X shoots 5 FPS at 12 bits but only 1.8 FPS at 14 bits. A suggestion was on this forum that 14 bit readout is achieved by combining two 12 bit readouts with different preamplifiaction before the column DA converters.
Another proposal (also made for the 14-bit mode of the D700) is that the same photosite charge is run through the sensor's ADC four times, each with the 12-bit ADCs of the Sony sensor, and the digital levels are summed. With the 12-bit ASC levels being integers from 0 to 4095, the sum is an integer from 0 to 16380, nicely fitting the 14-bit range from 0 to 16383.
That would average over time-random variation in ADC output for the same input. Four reads would only be expected to improve S/N by a factor of sqrt(4)=2, or a "one bit improvement", but if that is enough to go beyond the S/N ratio that 12-bits can record, the normal jump in bit depth is to 14 rather than 12, especially as that allows simply adding the four twelve-bit levels.
Some might raise the objection that A/D conversion destroys the charge in the sense capacitor at the bottom of the column. But that charge gets there from the photosite using charge gain amplification, not actual transfer of electrons, so the photosite-to-sense-capacitor transfer could be repeated. Doing that would have the advantage that the averaging also applies to the noise in the charge amplification process, which I have reason to believe is a significant source of noise in CMOS sensors.