5D3 raw histogram analysis

I've put some 5D3 raws through dcraw & rawhistogram.exe and made the observation that ISO6400 and ISO12800 are now "real" values, as opposed to the 5D2 where ISO3200 is the maximum sensor sensitivity and anything above is pushed by the firmware.

Now what's strange is that in the histogram, the odd values at ISO25600 and 3 of 4 values at 51200 are not zero, but some kind of low level background noise. What do you think this is?