They are for sure not single image machines. Each node is running its own kernel, sees only its own resources and is scheduled by the system-wide batch scheduler node. Running across nodes requires MPI to tie things together (tenuously) at the application level.

If they are impressive, it's only because of physical size and aggregate resources available, not because there's the least bit of magic involved. It's the usual Intel+-NVidia CPUs/chipsets/etc. At the current time, it's mostly their legacy proprietary interconnect.