i agree with testerx... for units that are designed to sell to customers, liquid cooling in the box is probably the best option there is. for units designed to be placed in a data center... more interesting 'mass cooling' systems are available... including rack based liquid cooling... (the pumps and liquid are cooled at a facilities level and the boxes have in and out ports with quick disconnect tubes)... and the ultimate... is the two phase immersion that 3m invented and asicminer is using supplied by allied control.... which would allow you to cool lots of chips simultaneously (and cool the other components like the dc/dc converters)
i think it'll be an interesting experiment to watch bitmine's effort.. which is the opposite of knc/hashfast/cointerras approach. bitmine has used the same 28nm process but with a much smaller chip (25 gh) that they think they can aircool easily with just a small heatsink on top. this could potentially be the cheapest to cool, if it works out and can scale... but theyve got 1/20th of the performance for each chip, which means 20x the packages, and maybe 10-20x the board space (the chip is smaller), 20x the dc/dc converter components etc. and a lot more, smaller, dc/dc converters etc. i suspect having lots of small chips, costs you more in packaging, and board and box space, and dc/dc converters... but costs you less in not requiring exotic liquid cooling. i would also wonder whether having lots of smaller dc/dc converters is less efficient than having bigger, more powerful ones... in the same way that bigger power supplies tend to be more efficient than smaller ones (especially if you use them under thir max load)
im eager to see the results of all of these asic mining options/experiments and see if one particular way becomes the more cost effective long-term to go.
i should add, that the small 'ant farm' type of cool chips approach has been done to death... by avalon, asicminer, bit fury etc... and has shown it works but has reliability problems... especially when all the chips are wired in a daisy chain together. one failure of a chip or board or minor component tends to create mass havok and take whole boards down. more 'parts' means more points of failure.
as gmax says.. cointerras is a very dense package.. so when youre paying for space in a data center, having more TH's in a 4U box is a cost advantage because you pay for both the power and space that you consume.
dont forget that each of knc, hashfast and cointerra... are using 4 dies in each of their packages. it wouldnt take much at all for them to use exactly the same dies... but re-package their 4 dies into 4 separate smaller packages and redesign the board to take 4x the smaller chips if that ever became a better way to go. but i suspect theyve already considered it. Although, in ct and hf's case, each die is still quite powerful and is not going to be as easy to air-cool as the weedy bit mine or bit fury chips (each hf/ct die is 100+ gh, thus 60-100 watts).
and then what they all do for the next generation will undoubtedly be better than what they did this time around, because theyve all learned a lot.. and will put the experience to good practice. the 2nd gen chip from each of them will be better than the 1st gen chip, in a more meaningful way than just the process shrink would give them because of architectural improvements.