Unified Memory simply isn't viable at the scale that a full Nvidia cluster operates at, and it would hamstring the parallelism that gives these clusters an edge. The current solution of Mellanox-style interconnect is probably still going to be the status quo in 5 or 10 years.