Media: Steve Conway 612-683-7133 Financial: Bill Gacki 612-683-7372 CRAY RESEARCH ACHIEVES HIGHEST MPP-SYSTEM PERFORMANCE EVER REPORTED ON COMPLETE NAS PARALLEL BENCHMARKS Moving To 256 Processors, CRAY T3D System Widens Performance and Scalability Lead EAGAN, Minn. Mar. 16, 1994 -- Five months after unveiling the CRAY T3D system, and just six weeks after building the first 256-processor version, Cray Research, Inc. (NYSE:CYR) has achieved the highest performance ever reported for a massively parallel processing (MPP) system on the complete set of eight NAS Parallel Benchmarks tests, the company said today. On every test, the 256-processor CRAY T3D system posted the highest MPP-system performance to date, according to Cray Research MPP program director Steve Oberlin. "On some tests, the CRAY T3D system was as much as three times faster than the performance reported for the next-fastest 256-processor system." The NAS Parallel Benchmarks were developed by the NASA Ames Research Center to measure the performance and scalability of parallel computers. "This news is important for several reasons," said Oberlin. "First, achieving industry-leading results on all eight tests confirms that the CRAY T3D system is not only the highest- performing, but also the most balanced MPP system in the world -- the NAS Parallel Benchmarks represent a broad range of MPP challenges with varying computation and communication demands. Second, getting complete 256- processor results this soon shows that the CRAY T3D system is easier to work with than other MPP systems. Third, moving up to 256 processors with near-linear scalability is a strong sign that the CRAY T3D system will efficiently scale performance to even larger versions of the product." The Cray system can be built with up to 2048 Alpha RISC microprocessors from Digital Equipment Corp., he said. "Because of its industry-leading high bandwidth and low latency, the CRAY T3D system does not run into the performance-degrading data traffic jams other MPP systems encounter when moving to larger sizes." Oberlin said some MPP vendors have had 256-processor and larger configurations of their systems available for as long as two years. "They haven't reported full NAS Parallel Benchmarks results for these larger configurations yet, however." The 256-processor CRAY T3D system is also the first MPP system to approach the performance of the company's 16- processor CRAY C916 parallel vector supercomputer system on several of the NAS Parallel Benchmarks tests, he said. "People sometimes forget that the CRAY C916 system is also a powerful parallel computer. Until now, no MPP system has reported results that approach the CRAY C916's published performance on any of the eight benchmarks except for the so- called Embarassingly Parallel test, where MPP systems are expected to excel." Oberlin said that when the CRAY T3D system was announced in September 1993, the company reported the highest NAS Parallel Benchmarks results for any MPP system with up to 128 processors -- the largest CRAY T3D system manufactured at that time. "The CRAY T3D system is still the best all- around performer on these tests at 128 processors, and now our early 256-processor results have set another new standard." Cray Research said it may have even faster results in the future, after having time to fine-tune the system for these tests as others have done. Oberlin said the 256-processor CRAY T3D system achieved industry-leading performance on both the original "class A" NAS Parallel Benchmark suite and the newer "class B" suite, which includes larger versions of the "class A" problem types. The company has announced 15 orders to date for the CRAY T3D system, and has said it expects to be the leading MPP vendor by the end of 1994. Cray Research creates the most powerful, highest-quality computational tools for solving the world's most challenging scientific and industrial problems. ### Editor's note: NAS Parallel Benchmark results for the CRAY T3D system, and for the 16-processor CRAY C916 system, are attached. For historical reasons, NASA Ames Research Center uses the performance of a one-processor CRAY Y-MP supercomputer system as the basis for measuring all parallel systems. The CRAY Y-MP system is Cray Research's previous- generation parallel vector system. Its successor, our current parallel vector system, is the the CRAY C90 system. PEs stands for "processing elements," or processors. Note near- linear scalability of the CRAY T3D system as it moves to larger sizes (numbers of processors). NAS PARALLEL BENCHMARK RESULTS Speed in units of one CRAY Y-MP processor System PEs EP MG CG FT IS LU SP BT ---------------------------------------------------------------------------- C916 16 81.7 31.3 34.6 36.1 43.1 18.9 36.1 27.9 T3D 256 96.8 25.8 5.4 33.5 12.4 11.9 18.4 23.1 T3D 128 48.5 14.2 3.5 17.2 6.6 6.5 9.5 12.0 T3D 64 24.2 7.3 2.1 8.8 3.4 3.4 4.8 5.9 T3D 32 12.1 3.0 1.1 4.4 1.6 1.9 2.5 3.0