I think the interesting point that hasn't been raised yet (as far as I've seen in this epic thread) is how KNC have binned their chips, if at all. BFL have 2 chips per Jalapeno and each can be (and usually seems to be) of a different quality in terms of hash rate and they must manually balance which chips get put on which board to achieve the advertised hashrate. Bigger boards have way more chips and it then becomes a classic partitioning problem.
Very good point, AFAIK, the chips are currently not binned at all. Not even tested for functioning (neither wafer probing or final chip testing) before being soldered on the PCB. Thats a huge mistake IMO, one that was pointed out months ago. They could be testing/binning the completed modules, but Ive not heard a word on that.
I just installed "bertmod 0.2.1 by uski" on my bad performing Jupiter (only 380 gh/s). And shows two Asci-chips (2 and 3) with lots of disabled cores:
ASIC_0
Die_0: 48 cores on, 0 cores off
Die_1: 48 cores on, 0 cores off
Die_2: 46 on, 2 off
Die_3: 46 on, 2 off >> 98 %
ASIC_1
Die_0: 48 on, 0 off
Die_1: 48 on, 0 off
Die_2: 48 on, 0 off
Die_3: 48 on, 0 off >> 100%
ASIC_2
Die_0: 47 on, 1 off
Die_1: 47 on, 1 off
Die_2: 19 on, 29 off
Die_3: 47 on, 1 off >> 83%
ASIC_2
Die_0: 48 on, 0 off
Die_1: 48 on, 0 off
Die_2: 3 on, 45 off
Die_3: 48 on, 0 off >> 76%
In total there are 81 cores off. That's about 10.5 % not working.... This while I'm only hashing 412 Ghs at the moment. As 412 / 550 is about 25% less performance, I guess that there is quite some overhead in the FGPA management and Firmware to "steamline" a stable system (about 15% additional loss)....