And what defines whether the cards run well vs turn into melted goo like seen in some of the other photos
Well, I suspect someone experimented with stuff that even bitfury himself said isn't worth the try. The issue is how to provide the 0.6-0.9V VDD for the chip. On the melted boards I don't see a power regulator (the thing that most people used to "pencil mod").
The chip was designed to support "chained VDD" configuration - in this case if let's say you have 5V power supply and you want to feed that into 8 chips you can chain them one after another and presuming everything works fine each of the chips will get 0.625V (or 5V/8chips). It's the same as wiring a bunch of christmas lights in series.
Everything works fine until one of them shorts. Then instead of each chip having 0.625V they'll get 0.71V (or 5V/7) which will make them produce more heat. When the next one fails the voltage goes to 0.83V (5V/6 chips), and with the next one it goes to 1V per chip. Etc, etc. etc.
So basically things turn from bad to worse very very quickly.
If you're interested in the technical details - that's a good place to start:
Single chips is quick & dirty. Yes I want more - ideally I would like to see board that is powered with 12 V strings and have no external components (costs) except chips and passive components.
But that won't be simple to get. But that's what I was aiming to actually blow off any other component vendors from bill of materials and do not make bottlenecks with turn-around-times and such with inductors, many power regulators and such.
But this is what again - likely can't be done quicky, only if very lucky and there should be no complex filtering/anti-resonance issue between chips in a string (you see - we now connect CMMINUS, CMQ, CMPLUS to GND).
