Virginia Tech's G5-based System X announced this week that thanks to a recent acquisition, they are now operating at 12.25 teraflops, sealing their already leading position in academic supercomputer leadership.
12.5 teraflops is more than many of us can imagine putting to work, but the commercial installations are pushing the boundaries much farther. SGI laid claim earlier this week to the current and fleeting number one supercomputer spot with an announcement of sustained performance of 42.7 teraflops (per second) by itssupercomputer built of Intel chips. I say fleeting, because it's only been a few weeks since a prototype Blue Gene/L nudged outNEC's Earth Simulator (final score for thatmatch was Blue Gene/L 36.01, NEC 35.9). Striking while the iron is hot, NEC immediately countered by announcing the availability of their newest 65-teraflop model.
Also this week, an article about Cray's new parallel super machine. It doesn't mention a current ranking, but says that the AMD-based computer is capable of a potential peak performance of up to 144 teraflops, if it were running all of the30,000+ chips it's designed to hold.
I love that so many supercomputers these days are based on desktop chips -- although I have to admit that, in the case of SGI and Cray, it does feel odd for them not to be usingtheir own chips anymore.
Even odder than that, though -- well, to me -- probably not for actual supercomputer experts -- is IBM's choice of embedded PowerPC 440 processors for Blue Gene/L. I think it is the unexpected choice of the embedded processors that makes Blue Gene/L my favorite in this 500+-horse race. That, and -- probably a bit of a bias, being as I'm affiliated with IBM, of course =)
But also, Blue Gene/L is a very safe horse to bet on, as it is designed to run at 360 teraflops once it gets out of prototype and into real "production" in 2005.