Teraflop in a box
By relling on Apr 20, 2007
About 20 years ago, I was working on a project to create a massively parallel processor capable of delivering 1 Teraflop (1 Trillion floating point operations (per second)). This was a grand challenge for the time and technology. Almost all of the projects trying to tackle this sort of problem were massively parallel architectures, and many of the companies formed to build such machines are long gone.
Today, it is quite feasible to put a bunch of machines in a (relatively) small room (or Black Box) and achieve 1 Teraflop. But the problem with these sorts of architectures (grid is the current buzzword) is that the programming and debugging environment is difficult. Lots of work has been done on that front, too, but it is still far easier to develop and debug programs in a single OS model with a single shared memory space.
I am very happy to say that we now have a single machine, running a single OS instance, with a single address space, which is capable of 1 Teraflop! The Sun SPARC Enterprise M9000 server is the fastest single machine on the planet!