The node interconnect is crucial. However, things like V100s (and presumably the CORAL2 GPUS) are clearly built for high performance computation, if not necessarily "HPC" in this sense; arguably also the nodes used by Sierra/Summit, though IBM seem just to say "AI". K computer was all built for the task, and Fugaku may be similarly impressive. See also the Chinese systems, though it's arguable how successful those designs are. There's also a fair amount of more or less hush-hush work in Europe on HPC custom hardware, though it's not clear that's paying off.

The majority of HPC systems probably use fat tree topologies (including Sierra and Summit as far as I remember). Something to hand which compares fat tree, hypercube, and the dragonfly topology discussed in the article is http://www.hpcadvisorycouncil.com/events/2015/swiss-workshop...

I recently did an analysis and it looks like for most supercomputers, only 15% of the total price is interconnect. I was surprised- I expected it to be more. Also, most supercomputers aren't hypercubes, they're using things like folded butterfly.

I'm really curious what process you were able to use to separate the component costs on systems like these. The manufacturers go out of their way to obfuscate that information.