With a Stouffer 13 1/3 stop test wedge, I could clearly see more steps (about 2 stops worth) on a Canon 1D MK III than a Canon 20D.
But why? You'd see them on a mk2, also, because the read noise is the same, relative to max signal, at ISO 100. The pixel read noise is lower in the mk3 than the 20D at all ISOs, plus it has more pixels, so if you achieve the same FOV, you will have even less subject read noise.
I've quantized ISO 100 RAWs from the mk3 to 12 bits, and they are indistinguishable from the full 14 bits after RGB interpolation, even zooming deep into the shadows, as long as the 12 bit version is promoted back to 14 bits with zeros in the 2 LSBs.
So, more bits might help for indirect reasons, and not necessarily because of what they contain. In my experience, working bit depth in conversion is generally more important than the significant bit depth of the RAW data. That's because there is so much noise, and the noise gets softened in conversion, and the greatest dangers of posterization occur in the rendered product.