Today's Debian technical committee resignation: Ian Jackson

Just to comment, as I'm someone who's been working in HPC for a while.

You are both right - there are indeed distributed memory systems out there as dlang mentions (the now "traditional" 20 year old Beowulf clusters) and also large Single-System-Image systems that bind lots of NUMA nodes together into a single machine that Kamilion mentions (think SGI's UltraViolet and earlier Altix systems for example).

Of course building a really massive Single-System-Image is very challenging as I believe the Linux kernel is limited to just 64TB of RAM so you can only get larger than that with distributed memory systems. :-(

All the best,
Chris