The Petaflop Race is On: NSF Contract Goes to IBM and NCSA

There are a multitude of teams racing to build the world's fastest supercomputer, probably by 2011. Yesterday, the US government's star contract for a DARPA computer, complete with $208 million in funding from the National Science Foundation, went to the National Center for Supercomputing Applications (NCSA) at the University of Illinois at Urbana-Champaign.

Local TV newscasts all over America (judging from their Web sites) ran with the story that IBM was awarded the US government contract to build "the world's fastest supercomputer" at the NCSA site (where, incidentally, the first Mosaic Web browsers were developed).

While "Blue Waters," as the new cluster will be dubbed, will likely smash the coveted one petaflop (one thousand trillion instructions per second) milestone, whether it ends up being the fastest on the Top 500 list has yet to be determined. There is plenty of competition.

Last September, for instance, the US Department of Energy awarded a contract jointly to IBM and AMD to build a hybrid supercomputer, made up of AMD Opteron and IBM Cell BE processor clusters. That machine has already been nicknamed "Roadrunner," and its target speed has already been set at 1.6 petaflops.

By comparison, the current fastest recorded supercomputer on the June 2007 Top 500 list - the DOE's and Lawrence Livermore Labs' BlueGene/L - has held on to the #1 spot for the past three years at 367 teraflops - about 23% the speed of the machine they're building to succeed it. The DOE isn't planning to give up pole position without a fight.

Then just last June, Sun Microsystems - the dark horse in this race by several paces - set a goal for itself of 1.7 petaflops, with its "Constellation" system being built for the Texas Advanced Computing Center.

Yesterday, the TACC team won part of a three-way, "Track 2" NSF fund worth $65 million, to build something that the NSF described as "toward the petascale." The Texas group might appreciate the help, but not that term "toward." The trick to its lower-cost design is a single InfiniBand switch that clusters processors together, replacing the 300-plus switches on a typical supercomputer.

A Knoxville, Tennessee newspaper yesterday lamented what it characterized as the University of Tennessee's "second place" award, for being one of the other partners receiving the NSF's "Track 2" chunk...while the DOE gets the "Track 1" prize plus another piece of "Track 2." The newspaper's information came from The New York Times, which raised a cloud of political suspicion over the reasons why UT and its Oak Ridge National Laboratories partner, among others, didn't fare as well as a government department.

Of course, had anyone thought to use Google, they might have discovered a simpler answer to all this than a government conspiracy: Oak Ridge already won a contract from Cray to build a competing petascale supercomputer in 2006. What's more, its completion date remains 2008, which means it could beat everyone else to that petaflop milestone. Besides, the characterization that the NSF favored the DOE at Tennessee's expense is inaccurate, since Oak Ridge National Laboratories is owned and operated by the DOE.

The Times article (which also ran in the International Herald-Tribune) also raised suspicion over why Livermore Labs didn't win the NSF contract, characterizing the loss as a shift of power away from New Mexico and California and towards the Midwest.

Of course, the answer to that question should now be obvious: LLNL decided as far back as 2005 to succeed BlueGene/L with a Roadrunner architecture cluster, which we now know will be an IBM/AMD hybrid.

So to bring you back up to speed, there are (at least) four major supercomputer projects in the works, any of which stands a chance to burst through the petaflop tape first: NSF awarded its contract to NCSA yesterday, where its IBM team will build "Blue Waters" for the University of Illinois. The DOE and LLNL are already at work building "Roadrunner," which will contain IBM and AMD chips. The TACC team in Texas has mounted a Seabiscuit-like effort to build a petascale supercomputer called "Constellation" with Sun chips and a single InfiniBand switch. And Oak Ridge Labs is already at work with Cray to build the "Baker" cluster, and may have had a year-plus head start on everyone else.

No conspiracies here, just a real horse race. Journalists still having difficulty interpreting it are advised to buy themselves a computer.