Guild Companies

The Defense Advanced Research Projects Agency of the U.S. military has ponied up $146 million in the second phase of to push the envelope on supercomputing technology so it can meet its petaflops-scale computing needs at the end of this decade. DARPA provided a few million bucks each to Cray, SGI, IBM, Sun Microsystems, and Hewlett-Packard this time last year to come up with innovative proposals and concept studies, and now the organization is paying real money for a 36 month research and development phase. Cray, IBM, and Sun made the phase two cut, while HP and SGI did not.

The High Productivity Computing Systems (HPCS) initiative managed by DARPA is akin to the Accelerated Strategic Computing Initiative (ASCI) undertaken by the US Department of Energy to drive supercomputing technology to from the hundreds of gigaflops to the hundreds of teraflops level in the late 1990s and early 2000s. HPCS is shooting for creating supercomputers that can deliver petaflops--a million teraflops--of computing capacity in the 2009 to 2010 timeframe. Such big computers present enormous challenges, and IT vendors love to accept those challenges because it is like having your research and development for products that will eventually make it into the commercial computing market funded by Uncle Sam. Nothing quite like having fun on the taxpayer's dollar.

Sometime in the middle of 2006, two vendors will be chosen for phase three, which will involve what is presumably a much larger sum of money to engineer and manufacture working systems that are expected to be delivered by 2010. That gives Cray, IBM and Sun three years to prototype their machines and five more years to actually build them.

Under the contract awards, Cray and a wholly owned subsidiary called New Technology Endeavors will be working on a concept machine dubbed "Cascade." Cray was awarded a $43.1 million contract to develop Cascade, which the company says will include new processor designs that make better use of memory hierarchies--the whole L1-L2-L3 cache scheme is running out of gas. Cray will be working on what it is calling processor-in-memory technologies that it hopes to demonstrate will deliver bigger sustainable memory bandwidths and lower latencies, which is the name of the game when building a giant cluster of computers. The Cascade machine will also try to include shared memory and distributed memory programming models while trying to allow existing applications to port to the new Cascade architecture. Cray's Cascade effort will be headed up by Burton Smith, the chief technologist at Cray; Thomas Sterling of the NASA Jet Propulsion Lab at Caltech, William Dally of Stanford University, and Peter Kogge of Notre Dame will be working with Cray on the Cascade project.

IBM was awarded a $53.3 million contract to fund research into a future Power-based supercomputer called PERCS, which is short for Productive, Easy-to-use, Reliable Computing System. (And in tongue and cheek fashion, the moniker also suggests that one of the perquisites of being the largest IT supplier in the world that will soon be able to demonstrate machines in the 100 teraflops range after a decade of hard work is being a shoe-in for this DARPA award.) IBM isn't saying exactly what PERCS is, except that it is based on its Power architecture and that it will have a new hardware-software design that will enable IBM to hit the multi-petaflops performance level while maintaining compatibility with existing HPC applications. PERCS will be designed and prototyped at IBM's research labs in Austin, Texas, where the "Squadron" Power5 servers were designed. IBM's Microelectronics, Enterprise Systems, Software, and Research divisions will all have a hand in make PERCS. IBM will collaborate with Los Alamos National Lab and a dozen academic institutions with HPC expertise as part of the PERCS project as well.

Considering the huge amount of credibility that both HP and SGI have in the HPC space and their desires to build systems in the petaflops class, it may seem odd that Sun was one of the phase two finalists with its $49.7 million award under the HPCS program from DARPA. But Sun appears to have taken a slightly different tack from trying to show it could build a monstrous supercomputer. Sun is intentionally vague about its "Hero" project, which DARPA is now funding, but it seems to be a little bit different from the Cray and IBM proposals, and what we presume the HP and SGI proposals looked like. What Sun will attempt with Hero is to take a machine with hundreds or thousands of processors and create development tools that will mask the underlying parallelism of the machine and make it look, at least to applications, like a single big processor even though it is not. This essentially what modern operating systems and symmetric multiprocessing do for commercial applications, but spanning 2 to 100 processors is different from spanning thousands, especially where HPC code is concerned. All Sun will say about Hero until this September, when it will raise the curtain a little, is that Hero will allow supercomputer programmers to be more productive, have greater numerical precision and better security than existing machines, and support legacy applications. Sun will also be developing the benchmarks that gauge the programmer productivity and performance of the DARPA HPSC prototypes.

MIT's Lincoln Laboratory in Lexington, Massachusetts will be managing the HPSC project in phase two on behalf of DARPA.