Media: Steve Conway 612/683-7133 Financial: Brad Allen 612/683-7395 1024-PROCESSOR CRAY T3D SYSTEM SETS NEW MPP PERFORMANCE, SCALABILITY RECORDS ON NAS PARALLEL BENCHMARK TESTS Cray Research Widens Lead In Race To Tackle "Grand Challenge" Problems Of Science And Industry EAGAN, Minn., Nov. 7, 1994 -- A 1024-processor CRAY T3D system has set new performance and scalability records on the full set of NAS Parallel Benchmarks, the most widely accepted independent comparison tests for parallel computer systems, Cray Research (NYSE:CYR) announced today. On the three Class A tests considered closest to real-world problems, called simulated applications, the CRAY T3D system was 3.6 to 8.7 times faster than the next-fastest MPP system, according to Steve Oberlin, the company's MPP program director. "On the BT simulated application, for example, the CRAY T3D system ran at 39.8 billion calculations per second (gigaflops), versus the next-best reported figures of 4.5 billion for IBM's SP2 system, 3.8 billion for Thinking Machines' CM-5E system and 1.9 billion for Intel's Paragon system." Cray is the first to report results for a massively parallel processing (MPP) system with more than 1,000 processors on the complete suite of tests, Oberlin said. Before this, the largest MPP system for which complete results were reported was a 256-processor CRAY T3D system, he said. Oberlin said the NAS Parallel Benchmark results are important because: - The 'Holy Grail' for MPP is scalability--the ability to increase performance efficiently as more processors are applied to a problem. The CRAY T3D system delivers very efficient (near-linear) performance increases all the way to the top-of-the-line, 1024-processor version of the product. Demonstrating good performance only on small MPP systems is not enough to address the most challenging, economically important scientific and industrial problems. - Achieving leading results on all eight tests, each representing a different computing challenge, confirms that the CRAY T3D system is a well-balanced product that can efficiently tackle a wide range of highly parallel problems. Vendors typically report results only on tests where their systems do well. The CRAY T3D system achieved industry- leading MPP performance on all of the original "class A" NAS Parallel Benchmarks and all of the newer "class B" tests, consisting of larger versions of the "class A" problems. - Cray has announced plans to deliver an MPP system able to operate at a sustained rate of one trillion calculations per second (one teraflops) in this decade. These test results show that Cray is on track for achieving that performance level, which the government's High Performance Computing and Communications (HPCC) program and National Information Infrastructure (NII) initiative have identified as needed for solving "grand challenge" problems-- enormously large, complex problems critical for advancing science and boosting industrial competitiveness. Oberlin noted that on seven of the eight class A tests, the 16- processor CRAY C90 parallel vector supercomputer system outperformed all MPP products except for the CRAY T3D system; on two of these tests, the C916 system outperformed every system, including the CRAY T3D. "We expect continuing strong demand for both types of systems. The choice depends on the nature of the customers' work." On the separate Linpack independent benchmark test, the 1024- processor CRAY T3D system achieved 100.5 billion calculations per second--another industry record for a system of this size, Oberlin said. "Although customer applications are the ultimate performance tests, independent benchmarks are strong performance indicators." Cray Research expects to be the MPP revenue-leader at the end of 1994, the company's first full year in the MPP market, he said. "The CRAY T3D system is being used for advanced scientific research, high-profile technology transfer programs, and industrial computing." The NAS Parallel Benchmarks were developed by the NASA Ames Research Center to measure the performance and scalability of parallel computers. The LINPACK benchmark results are periodically reported in "Performance of Various Computers Using Standard Linear Software," by Jack J. Dongarra. Cray Research provides the leading supercomputing tools and services to help solve customers' most challenging problems. ### Editor's note: Table of NAS Parallel Benchmark simulated application results follows on next page. CLASS A NAS PARALLEL BENCHMARKS -- SIMULATED APPLICATIONS HIGHEST-REPORTED SCALABLE PERFORMANCE COMPARED TO ONE PROCESSOR OF A CRAY Y-MP MULTIPROCESSOR SYSTEM TEST SYSTEM # PROCESSORS Y-MP/1 RATIO LU CRAY T3D 1024 47.0 CRAY C 16 18.9 IBM SP2 64 13.4 THINKING MACHINES CM-5E 128 5.1 INTEL PARAGON 256 4.4 KENDALL SQUARE KSR2 64 3.3 SP CRAY T3D 1024 87.2 CRAY C90 16 36.1 IBM SP2 64 13.6 THINKING MACHINES CM5E 128 7.7 INTEL PARAGON 324 5.3 KENDALL SQUARE KSR2 64 3.6 BT CRAY T3D 1024 173.8 CRAY C90 16 27.9 IBM SP2 64 19.9 THINKING MACHINES CM5E 128 16.5 INTEL PARAGON 512 8.3 KENDALL SQUARE KSR2 64 6.1 CLASS B NAS PARALLEL BENCHMARKS -- SIMULATED APPLICATIONS HIGHEST-REPORTED SCALABLE PERFORMANCE COMPARED TO ONE PROCESSOR OF A CRAY C90 MULTIPROCESSOR SYSTEM TEST SYSTEM # PROCESSORS C90/1 RATIO LU CRAY T3D 1024 31.2 IBM SP2 128 13.0 CRAY C90 16 12.6 INTEL PARAGON 512 3.7 THINKING MACHINES CM5E 128 2.0 KENDALL SQUARE KSR2 64 0.3 SP CRAY T3D 1024 28.3 IBM SP2 128 11.2 CRAY C90 16 8.9 INTEL PARAGON 484 3.4 THINKING MACHINES CM5E 128 2.2 KENDALL SQUARE KSR2 64 0.2 BT CRAY T3D 1024 61.7 IBM SP2 128 16.7 CRAY C90 16 13.1 INTEL PARAGON 306 5.9 THINKING MACHINES CM5E 128 5.0 KENDALL SQUARE KSR2 64 0.2