Well, that article is talking about moving data around on a multi-chip Xeon based systems with Tesla or Quadro-class GPUs. Those cards are very different beasts from the consumer based graphics cards we have been discussing here. Very different.
I would believe that having a pair of 980Ti's in SLI on one Xeon, and another pair of 980Ti's in SLI on a second Xeon, even on the same motherboard, would slow things down. Especially when trying to generate 4K resolution graphics.
There is always latency when jumping from one massive CPU to the next. And I do not believe that will be eliminated until we get optical channels that run from the CPU through a motherboard and to a second CPU. And we are years away from that still. My guess is probably 5 to 10 years, at a minimum.