The Cray XT3 supercomputer's architecture, co-designed with Sandia as part of the $90 million "Red Storm" system contract, delivers superior scalable application performance and value across a range of configurations from 200 to 30,000 processors, with peak performance of up to 144 teraflops (trillions of calculations per second).
Cray has shipped a 10-teraflop portion of the "Red Storm" system
to Sandia. When fully installed, "Red Storm" will have over 40 peak
teraflops of performance, more than 11,000 AMD Opteron(TM) processors, and 240 terabytes of disk storage. The system is expected to be at least seven times more powerful than Sandia's current ASCI Red supercomputer on real-world applications.
That is just the beginning. "Today's Cray XT3 is the first in a
series of increasingly powerful scalable Cray products that exploit
the Red Storm architecture. The architecture will allow capability to
be increased with a simple processor upgrade," said Sandia's Bill
Camp, Director of Computers, Computation, Information and Mathematics.
The U.S. Department of Energy's Oak Ridge National Laboratory
(ORNL), which selected the Cray XT3 system for their National
Leadership Class Facility computing initiative, is slated to receive a 20-teraflop Cray XT3 supercomputer, along with a 20-teraflop Cray
X1E(TM) vector MPP supercomputer, in 2005. In May 2004, the DOE chose
ORNL "to lead a partnership with a goal of building the world's most
powerful supercomputer by 2007," according to Energy Secretary Spencer Abraham. ORNL plans to expand to a 100-teraflop Cray system at Oak Ridge in 2006, and to move in 2007 to a system with over 250 peak teraflops and up to 100 sustained teraflops on real-world problems.
The Pittsburgh Supercomputing Center (PSC) has signed a contract
for a 10-teraflop Cray XT3 supercomputer that can be expanded over
time. In a joint statement, PSC scientific directors Michael Levine
and Ralph Roskies said, "We are very enthusiastic about making this
new and powerful scientific instrument available to National Science
Foundation researchers. Remotely using part of the XT3 system which
will soon be shipped from Cray to PSC, we have successfully run the
Advanced Regional Prediction System (ARPS) from the Center for
Analysis and Prediction of Storms, led by Kelvin Droegemeier. Our
ability to execute this full application, a comprehensive regional to
storm-scale atmospheric modeling/prediction system, as well as our
successes with segments of other simulation programs, strongly
indicates that it will be a highly productive computational resource."
According to Rich Partridge, Enterprise Systems analyst with D.H.
Brown Associates, "Thanks to its blend of high performance computation and robust communication, Cray's T3E was the leading MPP system for years. The Cray XT3 becomes the logical successor. Employing updated processor and interconnect technology, Cray again offers a highly scalable MPP design, with a balance of computation and communication capabilities that promises to deliver superior performance on real-world problems."
"The Cray XT3 supercomputer, designed in partnership with Sandia,
advances the achievements of the Cray T3D and Cray T3E, widely
recognized as the gold standard for MPP systems, as well as the ASCI
Red supercomputer. The Cray XT3 sets a new standard for the efficient
scalability and reliability of standard microprocessor-based system
designs," said Peter Ungaro, Cray senior vice president for sales,
marketing and services. "On real problems, the Cray XT3 system's
balanced design will enable it to outperform large-scale clusters with substantially greater theoretical peak or Linpack (Top500) speed."
The Cray XT3 Supercomputer - Scalable By Design
The Cray XT3 supercomputer's high-bandwidth, low-latency design --
purpose-built for high performance computing (HPC) applications --
delivers a much higher percentage of peak performance in practice than HPC clusters and other alternatives.
The Cray XT3 supercomputer uses Advanced Micro Devices Inc. HyperTransport(TM) technology and Opteron(TM) processors connected via a Cray low-latency, high-bandwidth, three-dimensional torus
- A 3D torus direct connected processor (DCP) architecture tightly links processors to nearest neighbors.
- A highly scalable compute system incorporates high performance AMD Opteron(TM) processors, high speed memory, an MPP optimized operating system, and a standards-based programming environment to deliver unmatched sustained application performance in configurations from 200 to 30,000 processors.
- A high bandwidth, low latency interconnect with embedded communications processing and routing offers total interconnect bandwidth of more than 100 terabytes/second.
- High speed, global I/O is scalable to over 100 gigabytes/second of I/O (input/output) bandwidth and hundreds of terabytes of parallel disk storage.
- A tightly integrated management and operating system provides high reliability and is designed to run full-system applications to completion.