CRAY RESEARCH UNVEILS NEW TOP-OF-THE-LINE SUPERCOMPUTER FOR THE 90s, WORLD'S MOST POWERFUL
CRAY RESEARCH UNVEILS NEW TOP-OF-THE-LINE SUPERCOMPUTER FOR THE 90s, WORLD'S MOST POWERFUL EAGAN, Minn., Nov. 19 /PRNewswire/ -- Cray Research, Inc. (NYSE: CYR) introduced today the world's most powerful parallel vector computer system, the CRAY Y-MP C90 supercomputer. The new system has 16 central processing units (CPUs), and on actual customer problems, operates four times the speed of Cray Research's previous fastest supercomputers. The CRAY Y-MP C90 system, announced earlier today at a news conference in Tokyo and unveiled at the Institute of Electrical and Electronic Engineers Supercomputing '91 conference in Albuquerque, is the most powerful general-purpose supercomputer ever offered for scientific and engineering applications. The CRAY Y-MP C90 system, fully binary-compatible with the company's recently broadened CRAY Y-MP product line, features an all-new CPU with a peak performance of one billion floating-point operations per second (Gflop/s).
With 16 of these powerful CPUs and 256 megawords (2 gigabytes) of central memory, the CRAY Y-MP C90 system has a peak performance of 16 Gflop/s.
"This powerful system will help change the world," said John Rollwagen, Cray Research chairman and chief executive officer. "The
CRAY Y-MP C90 will enable scientists and engineers to solve some of the world's most complex and critical problems (Grand Challenges) that, until now, couldn't be solved with available computational tools. In addition, this system will open exciting new computational avenues for solving critical commercial challenges."
"The new supercomputer has achieved exceptional, sustained performance rates on a broad-range of user application codes. We have run initial benchmarks on the most popular third-party software codes and the results have been outstanding. In fact, one large weather model from the European Centre for Medium-Range Weather Forecasts (ECMWF) ran on one C90 CPU two and one-half times faster than on one CRAY Y-MP CPU. We anticipate this model to perform on the full system at a sustained rate of four to five Gflop/s, compared to the previous high speed of less than one Gflop/s on our eight-processor CRAY Y-MP system." Rollwagen said that the most important measurement of system performance is its ability to do real work. The CRAY Y-MP C90 system is able to achieve and maintain exceptional speeds on popular software codes that today are applied to a broad range of sciences. In addition, the CRAY Y-MP C90 system delivers its speed with the production-tested UNICOS family of compilers, tools, libraries, and applications, allowing existing programs to be used unchanged. Scientists and engineers of the new system can immediately apply its performance to their tasks without re-engineering their software. This will enable these users to make new breakthroughs, enhance their insights and knowledge, investigate problems that couldn't be considered in the past, and shorten time-to-market for new products. "Today also marks a milestone for our network supercomputing strategy," Rollwagen said. "We have fulfilled our promise to deliver, by the end of 1991, a full spectrum of high-performance computing systems that support industry standards and can be integrated easily into customers' heterogeneous computing environments. Our full spectrum of products is topped by the CRAY Y-MP C90 system, the most powerful, general-purpose supercomputer today." Rollwagen noted that network supercomputing increases productivity by optimizing workload distribution in customers' distributed processing networks. With Cray Research's expanded product line, which now includes the CRAY Y-MP EL system (an entry-level system announced last month) the company implemented a new performance measurement called Relative Processor Performance (RPP), based on the performance of a single CRAY Y-MP CPU. Today's announcement extends the performance scalability of binary-compatible, UNIX-based CRAY Y-MP platforms to more than two orders of magnitude of performance, ranging from the CRAY Y-MP EL system providing .2 to .9 RPP, to the CRAY Y-MP C90 system delivering from 32 to 40 RPPs. Customer Interest Cray Research already has received seven orders and one letter-of- intent for the powerful CRAY Y-MP C90 system, and demand for the new system has filled the production schedule for the next 12 months. The full 16-CPU system has been running at Cray Research since October, and customer shipments are expected to begin in January 1992. The first international installation of the CRAY Y-MP C90 system will be at ECMWF in Reading, England. "We placed our order for the CRAY Y-MP C90 system some time ago and have been looking forward to taking advantage of the additional power and capabilities of the new system to improve the accuracy of our medium-range (10-day) weather forecast," said Dr. David Burridge, director of ECMWF. "We are very excited about the five-fold speedup on the initial benchmark of our 10-day weather forecast model. And our scientists and analysts are enthusiastic about installing our new CRAY Y-MP C90 system and achieving comparable, possibly even higher, sustained speeds on our operational forecasting model." Fleet Numerical Oceanography Center (FNOC), the U.S. Navy's primary numerical processing center for global atmospheric and oceanographic predictions, will install a CRAY Y-MP C90 system in fourth quarter 1992 at its computer center in Monterey, Calif. The system will be the principal computational resource for the Primary Oceanographic Prediction System (POPS), a Navy program designed to provide air-ocean prediction capabilities of greater resolution and accuracy than that possible with previous computer resources. "The power, memory, and throughput capabilities of the CRAY Y- MP C90 system will allow us to run the large and very complex scientific models that couple the ocean and atmosphere," said Randy Nottenkamper, POPS site manager for FNOC. "With today's computer resources, we model the atmosphere and ocean separately. This new class of machine will enable us to model the air and ocean as the single continuous medium it is in nature, and thereby produce more accurate environmental products for the U.S. Navy and other U.S. DOD forces worldwide." The Pittsburgh Supercomputing Center (PSC), one of four National Science Foundation (NSF) Centers, has also placed an order for the new system. PSC will be the first non-government lab in the U.S. to receive the CRAY Y-MP C90 system. "We at the Pittsburgh Supercomputing Center are taking another step forward in technology leadership with the acquisition of Cray Research's Y-MP C90 system," said Michael Levine, Ph.D., scientific director of PSC. "We were the first non-government lab to receive the company's previous generation Y-MP8 and to make its capabilities available to the national research community. We are excited to be able to repeat this achievement once again with the CRAY Y-MP C90 system to bring this most powerful scientific computing system to American researchers. "Working with our users, our goal is to solve significant scientific and engineering problems. We are confident that the power and applicability of the C90 system will enable our users to master important challenges essential to our national scientific leadership and competitiveness." Parallel Power The CRAY Y-MP C90 system uses a balanced parallel, vector-scalar architecture to maximize sustained performance. "The C90 system provides the highest level of memory and aggregate I/O bandwidth available in any computer system in the world today," said Les Davis, chief technical officer for Cray Research. One of the most significant architectural features of the CRAY Y-MP C90 system is a dual-vector pipeline that allows each of the system's 16 CPUs to deliver two vector results per functional unit every clock period, said Davis. With its 64-way parallelism and efficient multiprocessing capabilities, the CRAY Y-MP C90 system can deliver a total of 64 vector results per clock period, or four times that of Cray Research's previous top-end systems, he added. "The CRAY Y-MP C90 system has real and usable 'parallel power,'" said Davis. "The system's multiprocessor hardware capabilities, dual- vector pipeline feature, and proven, production-tested multitasking software, make it easy to apply the parallel power of this system to today's most widely used application codes. The CRAY Y-MP C90 efficiently handles scalar, short vector, long vector, parallel, and highly parallel problems." The CRAY Y-MP C90 system, like all CRAY Y-MP products, operates Cray Research's feature-rich UNICOS operating system and the CF77 Fortran, Cray Standard C, Cray Ada, and Pascal compilers. The new system supports the complete CRAY Y-MP library of more than 1,000 application codes covering nearly every scientific and engineering discipline. Cray Research, Inc. designs, manufactures, markets, and supports high-performance computer systems for scientific and engineering applications. Editor's Note: Attached are more specific details of the CRAY Y-MP C90 system features and a glossary of computer terms. The CRAY Y-MP C90 model is on display at the Cray Research booth at the Supercomputing '91 conference at the Hyatt Regency in Albuquerque from Nov. 19 - 22. CRAY Y-MP C90 SYSTEM FEATURES -- Autotasking, a Cray Research software product that portions a problem and distributes it across the system's multiple CPUs, provides automatic parallel processing. -- Full complement of production system software and industry standard applications available today. -- Four double-width memory ports per CPU and more than 250 gigabytes/sec. of memory bandwidth help deliver more sustained computing power to the user. -- With up to 16 I/O clusters, aggregate I/O bandwidth of over 13 gigabytes/sec., and the ability to directly attach more than four terabytes of disk storage, the CRAY Y-MP C90 is the most powerful and versatile computer system in the industry. Powerful I/O software further enhances the system's I/O capabilities. -- An optional Solid-state Storage Device (SSD) provides very high- speed secondary memory with up to 2,048 megawords (16 gigabytes) of storage capacity. With a bandwidth of 7,200 megabytes/sec to the mainframe, the SSD allows users to solve large problems in less time. GLOSSARY OF TERMS Autotasking -- A Cray Research software product that portions a problem and distributes it among the system's processors for more parallel problem solving. Binary Compatible -- Refers to compatibility of computers; it's when no software changes must be applied to an actual application in order to run the application on another machine. This is in contrast with "source compatibility," where programs must be re-compiled. Binary compatibility enables all software to be accepted and run across all computer systems in the full CRAY Y-MP product line. Central Processing Units (CPU) -- Computer hardware where the actual computation takes place. CPUs originally were referred to as the element that both issued and executed computer instructions. The interpretation has changed, especially with respect to Single Instruction, Multiple Data (SIMD) parallel machines, where the "CPUs" do not actually issue an instruction. Clock Period -- The time in which a computer runs through one complete cycle. Usually measured in nanoseconds (one billionth of a second). Some computers have a "major" and "minor" clock period where some special instructions cycle in a shorter minor clock period, but the shortest time in which the entire machine cycles is the major clock period. Dual-Vector Pipeline -- A functional unit that produces two results per clock period. A single vector pipeline generally produces one result per clock period. Floating-Point Operations Per Second -- The rate of computations for decimal arithmetic computations. The technical abbreviation is flop/s, but it is often incorrectly used as "FLOPS." General-Purpose -- A computer designed to operate on a program of instructions for the purpose of solving a wide variety of data processing problems, rather than being designed to fulfill a single function or type of function. General-purpose computers use general- purpose operating systems and languages. They do not rely on proprietary languages and unique extensions. Gigaflop/s or Gflop/s -- One billion floating-point operations per second. A measure of computational speed. Highly Parallel -- Program characteristics with an unusually large number of independent (parallel) tasks. Generally refers to one thousand or more parallel tasks. Long Vector -- Program characteristic with an extremely long vector. Generally refers to vectors in length of more than 1,000. Megaword -- Measure of computer memory capacity. One megaword is equivalent to eight million bytes. Multitasking -- A methodology in multi-user systems that incorporates multiple interconnected CPUs. It lets users run their programs simultaneously (in parallel) because each user has a CPU, and shares resources such as memory, storage devices, and printers. This word can often be used interchangeably with "parallel processing." Network Supercomputing -- The use of computational elements, distributed on a computer network, to solve a single problem. Parallel -- The internal handling of instructions and/or data in groups, all elements of a group being handled simultaneously. When the word parallel is used to describe a computer this means a computer in which the computation is handled concurrently by separate units of the computer. The units may be interconnected in different ways as determined by the computation. Relative Processor Performance (RPP) -- Performance measure based on the relative power of computers. The Cray Research RPP is based on the computation power of one CRAY Y-MP CPU. Scalar -- Non-vector. "Scalar" often refers to floating point or integer computation that is not vectorized, but more generally it also refers to logical and conditional (jump) computation. Vectorized code that does not amortize start-up time for computation. That is, scalar refers to computation not involving a specific vectorizable independence. Short Vector -- "Short vector" typically refers to vector lengths less than 50 or 100. Vector -- Construct in applications have a specific parallel and allow computation by vector processors. Vector processing is a special case of parallel processing and is similar to some forms of (SIMD) parallelism. Vector-Scalar -- Calculations involving both vector and scalar computation. -0- 11/18/91 /CONTACT: Kate Neessen-Haugh of Cray Research, Inc., 612-683-7132/ (CYR) CO: Cray Research Inc. ST: Minnesota IN: CPR SU: PDT AL -- MN003 -- 4790 11/19/91 09:02 EST
|Printer friendly Cite/link Email Feedback|
|Date:||Nov 19, 1991|
|Previous Article:||AUSTRALIAN GOVERNMENT AWARDS MAJOR TELECOMMUNICATIONS LICENSE TO BELLSOUTH CONSORTIUM|
|Next Article:||DOSKOCIL REPORTS IMPROVED THIRD QUARTER OPERATING RESULTS|