That is a strange form factor.

Also, you can get 0.7W/GH even on a 55nm Bitfury chip (or an sp10 chip for that matter) , so i would say the pricy 20nm isn´t quite justified

Again, power efficiency matters more than process node. In fact, it is better to have lower consumption on a higher node, as you can always do a die shrink to improve even further.

I was also under the assumption that knc spent nearly $5 million in their 28nm NRE as they were paying massively for overtime and such.
I was further under the assumption that they claimed their 20nm NRE was around 3 times as much as their 28nm, so $ 7m might be a conservative estimate.