Cray is hoping that both machines will help even out its sales as it also promotes the high-end Cray X1 vector supercomputers.
The Pittsburgh Supercomputer Center has convinced the US National Science Foundation to part with $9.7m to build a 10 teraflops supercomputer based on the Red Storm design, which is being productized under the code-name Strider but which is not officially a product yet. PSC will be getting a Red Storm system with 2,000 Opteron processors, and it will be getting it pretty fast too, with installation expected by the end of 2004.
The full Red Storm machine going into Sandia will be rated at 41.5 teraflops, and this marks the first time that Cray has not been vague about exactly what the performance of the Sandia Red Storm box will be. PSC is home to the LeMieux parallel AlphaServer-Tru64 Unix supercomputer rated at 6 teraflops of peak performance; this machine was acquired under the TeraGrid project being jointly funded by the NSF and IT vendors IBM Corp, Sun Microsystems Inc, and Intel Corp. PSC says that the Red Storm machine will succeed LeMieux, which provided 60% of the computing time used by NSF in the past year. LeMieux was great in its time, but the Red Storm machine would fit in a living room while LeMieux occupies a space the size of a basketball court. (Moore’s Law is great, isn’t it?)
Cray says that the Sandia implementation of Red Storm has sophisticated security features (which presumably add to its cost) that the PSC knock-off does not require, which is probably just a way to justify the fact that Sandia is paying $93m, or nearly ten times as much money, to get a machine based on the same architecture that is only four times as powerful. The truth is, Sandia – and therefore US taxpayers – paid for the development of Red Storm, and now other commercial and research institutions are going to benefit from that investment. PSC is one such institution, and Oak Ridge National Lab in Tennessee, which is getting a 20 teraflops Red Storm box in 2005, is another. This is how the very nationalistic supercomputer business works.
Cray also said yesterday that the Saha Institute of Nuclear Physics in Kolkata, India, had acquired one of its XD1 Linux-Opteron boxes. The XD1s were created by an obscure and clever Canadian supercomputer maker called OctigaBay, which designed a sophisticated system that can scale from 12 to 12,000 Opteron processors and which offers very low costs compared to RISC/Unix parallel supercomputers. The Indian order came through Hinditron Group, Cray’s reseller in India that is based in Mumbai. The machine that the institute is buying has 96 Opteron engines and is rated at 422 gigaflops.
