Another way to put it is that specialized supercomputers no longer exist, and the top500 is now populated with super-branded racks of PCs running Linux with souped up network cards.
For the young ones in the audience, the supercomputer term has historically been used to describe machines with largely custom hardware from the CPUs up - or at least from the CPUs onwards. Cray vector machines are the canonical example. These were shared-memory / single-system-image machines, software really saw it as one computer instead of a networked cluster.
I don't think this description is entirely fair. What is expensive with these computers is the fast interconnection buses between processors and modules, which is needed for many scientific problems that can't easily be divided into independent subproblems. Just having a large cluseter of computers connected with normal Ethernet would not cut it. OK, you say souped out network cards, and if you had said souped up switches as well, I could have given you some points, but this is the major cost of modern supercomputers. You can get a non souped-up cluster with the same number of processors for a one digit percentage of the price of the supercomputer.
Pretty sure everything in top500 is using infiniband by now.
I was briefly involved with the supercomputing world about four years ago. "Should we use Infiniband?" was no longer considered a non-rhetorical question even then.
It's either Infiniband or some other high speed interconnect. It's not exclusively Infiniband, for instance the Chinese developed their own, but it's usually something similar.
> Another way to put it is that specialized supercomputers no longer exist, and the top500 is now populated with super-branded racks of PCs running Linux with souped up network cards.
While there's a grain of truth in this, in reality, a supercomputer still looks like a room full of refrigerators and not at all like a typical computer or even a server farm.
Some of the computer hardware in there is just commodity hardware but for example the interconnects can't really be described as "souped up network cards", they're in a category of their own.
Yes, the era of the Cray boxes is gone but a supercomputer is still a supercomputer.
In addition to just the networking hardware, the networking stacks and topologies are very different. I would consider something a supercomputer if it was using a high-dimension torus topology, and some MPI variant, and a different internet / transport protocol (I forget the name, but as I understand it's a very different approach where there's much less decision making in the routers). On the other hand, what I would consider to be a cluster is almost going to use off-the-shelf networking hardware (as you mention), TCP/IP, and a simpler hierarchical topology with a switch or two per rack.
They're not exactly "super-branded" racks of "PC"s the way we tend to think about them, a bunch of white/beige boxes. They are server-grade computers, with all that comes there (special interconnect features, ECC memory and all that), but there is indeed no question as to their distributed nature.
This is probably a technological response to the rapid pace of development. It's more efficient this way, instead of developing custom chips and boards for two years and deploying them when the high pace of advance makes them half-obsolete.
For the young ones in the audience, the supercomputer term has historically been used to describe machines with largely custom hardware from the CPUs up - or at least from the CPUs onwards. Cray vector machines are the canonical example. These were shared-memory / single-system-image machines, software really saw it as one computer instead of a networked cluster.