Red Storm rising: Sandia/Cray team to create 40 teraOPS supercomputer

An announcement at Sandia — long awaited by the nation’s defense labs and computer industry — confirmed last week that Sandia has teamed with Cray Inc. to develop and deliver Red Storm, a massively parallel supercomputer theoretically capable of reaching a peak performance of 40 trillion calculations (teraOPS) per second.

The agreement was formally announced at a Sandia-hosted news conference Oct. 21 at the new International Programs Building east of the Eubank gate. Speaking were Sen. Pete Domenici, R-N.M.; Linton Brooks, acting administrator of the National Nuclear Security Administration; Bill Reed, acting director of NNSA’s Advanced Simulation and Computing Program; Jim Rottsolk, president and CEO of Cray Inc.; and Tom Hunter, Senior VP of defense programs at Sandia. Sandia President C. Paul Robinson made the opening remarks and introductions.

Said Tom, "It’s wonderful to see a vision become reality."

Sandia reported in June that Cray had been selected for the award, subject to successful contract negotiations.

The machine’s speed is based partly on its expected ability to deliver two calculations per clock cycle rather than one, which would total to 20 teraOPS. The new Red Storm architecture also relies upon a very high performance, specially designed 3-D mesh interconnect and Advanced Micro Devices’ (AMD) Opteron processors.

Custom aspects of the machine are a departure from recent trends in supercomputing architectures, which have made a point of relying on off-the-shelf parts. Nevertheless, the machine is said to have an excellent price/performance ratio.

The computer, expected to be operational at Sandia in the summer 2004, will be approximately seven times more powerful than ASCI Red, Sandia’s fastest computer now. It is also expected to have the capability to achieve 100 teraOPS with added hardware.

Nuclear weapon engineering simulations are the major driver of the computer, although it will also serve a broad spectrum of scientific and engineering applications.

For this reason, the installation at Sandia will operate in a dual-network configuration — classified (Red) and unclassified (Black). The machine can be rapidly reconfigured to make all the compute nodes classified, all the compute nodes unclassified, or, in normal operations — three quarters of the compute nodes available to either of the two networks and one quarter of the machine available to the other.

Currently, the world’s fastest supercomputer is NEC’s Earth Simulator (35.86 teraflops) in Japan, followed by ASCI White (7.22) at Lawrence Livermore National Laboratory.