^^ The possibility of the package size and same limited coms setup makes for an interesting thought: Perhaps aside from some changes to the support chips to handle higher throoughput, the new chips could basically use the existing hashboards or at least most of their already validated portions perhaps? Hmmm..
That would certainly speed deployment in BitFury's almost mobile mini-pharms
http://bitfury.com/products#container-datacenter and the new DC... Once more than just engineering samples are produced, tweaked hashboards tested and the foundry is given the the green light to start a full scale run(s) it's then just a matter of changing tankfulls of boards one at a time. If the data tanks are removable as a unit, all the better. prep/change over tanks to the new 16nm chips & boards off line and swap out the old ones with the new. Rinse & repeat as desired & production (and money) allows for.
Should give their selected integrators a leg up as well I'd think.
A last thought on why deploy 28nm in the new DC: Time to produce enough boards/chips to fill it. The building & all infra itself was completed how long ago? (seriously, when?) With 28nm chips in production at the time or left over from the final runs, why not? Is a good CYA to start filling/using the DC until the 16nm chips start getting made en-mass which is, um, now if I've read things right.
Query to BitFury: If I may ask, how many boards & chips are in each tank? Based on the stated 100GH/s a 12Ph mini-farm takes 120k chips. What size wafer is used? What is the actual silicon die size (along with full package size?