I read Prof Jack's comments somewhere today. He commented how the China computer did not depend on Infiniband interconnects ... suggesting that his in ORNL does. Infiniband is a decade old interconnect standard. The operative words are "decade" and "standard" == not nearly the optimal for what could be custom designed with today's tools.
The China system uses a custom interconnection system AND that is not so expensive even if completely built in the US! (And I will bet anything that the interconnect system is designed and built in the US.) Anyway that suggests & translates to the various individual systems of the China system can communicate each other much faster than the ORNL system.
Jack also comments about how the h/w, s/w, data tables etc. must all be managed properly. Not sure what he was actually thinking about, but I it is clear that as the individual systems are able to communicate with each other at higher speeds, that the method of distributing computing can also become more sophisticated. Just think about the efficiency difference of single core CPU systems transferring data across an Ethernet network versus a multi-core CPU performing the identical computation ... for a lame analogy.
Also, the China system is massively larger than the ORNL system. When finesse is not enough, then use brute force ..
more can still be better!
Last but not the least, the China budget was bigger.
