ASSUMING that they use quad channel interface for each chip with the fastest ddr3 in existence and its 96GB not 72GB, that would make the total memory bandwidth of the rig 1TB/s which is 4 times faster than a RX580 or around 120MH/s. That would be the worst case scenario.
Thats incorrect, if they are using quad channel spread to 8 16bit 512MB DDR3 chips for each core thats ~14mh per core (assuming around a 1GHZ ram speed, some DDR3 can go up to 1600mhz. 14x 6 cores x 3 boards ~ 250 MH.
How did you calculate 14 MH/s per core? 14 MH/s would require roughly 140 GB/s bandwidth which seems unrealistic even for a quad channel DDR3. A quad channel 1866MHz DDR3 can reach 59.7 GB/s (source https://www.microway.com/knowledge-center-articles/performance-characteristics-of-common-transports-buses/). Therefore one core hashrate should be close to 6 MH/s. 6 MH/s * 6 * 3 = 108 MH/s.
933 mhz x 2 (DDR) x 4 (quad channel) x 128 bit wide bus (8 x16 bit DDR3 chips) / 8bits = 120 GB/s
Remember stock ram for motherboard is 64bit wide no reason why you can't go higher, if they are really using something between 72-96GB per board and I'm assuming article is wrong and its 72GB which would make sense for 8 x 512mb chips per core for 128bit wide bus.
Sure, a 128 bit quad channel bus is effectively a 512 bit bus. AMD and Nvidia have used 512 bit buses in the past, however they both have moved away from those due to high cost and high power consumption. If Bitmain's ASIC really does have a separate 512 bit bus for each core (so 18x 512 bit bus in total), then 200+ MH/s is possible. However, in that case the retail cost and the power consumption will be pretty high. I'm sure they wouldn't release anything that's worse than a comparable GPU rig but most likely it's not going to drive GPU rigs out of the market either.