I haven't gone digging on this really, but part of me wonders if there's room in this communication link for packet (or, transaction) fragmentation. Say, we know the maximum size of each transaction can be 20 bits. Maybe it can work like some sort of packed VLIW system where individual commands can be packed together to make more use of the bus? Of course, that also assumes that there will be the other cases where you can't pack enough together and thus you get transactions that use all 20 bits.
That seems fairly unlikely, to be honest. I mean, 20 bits isn't much data... But I still can't figure out any real reason they've gone from measuring in bytes to transactions.