CLOUD

By Uwe Harms, Harms Supercomputing Consulting -- In late February NEC announced in a press tour its new departmental SX-6i vector supercomputer with a peak performance of 8 GFlop/s, up to 8 GByte memory, DDR-SDRAM, and 32 GByte/s read/write performance. The list price is 200,000 EURO, but as users in the supercomputer arena know, the price is a matter of negotiations. Additionally NEC disclosed the actual status of DKRZ, Hamburg, Germany, and the Earth Simulator in Japan. NEC starts to sell the deskside vector supercomputer NEC SX-6i with its proprietary fastest one-chip-vector processor with a peak performance of 8 GFlop/s. Its size of 45x73x70 cm is comparable to a PC in a tower. Additionally the NEC-SX6i is available in a rack version with one and two processors. The processor consists of 8-way vector pipes running at 500 MHz. There are vector add/shift, vector multiply, vector logical and vector division. The peak performance is calculated only for add and multiply. If a vector division is running in parallel to vadd and vmult, the SX-6 exceeds its peak performance. The memory module is up to 8 GByte/s with a memory bandwidth of 32 GByte/s. This compares in the Stream Triad benchmark to 31.982 GByte/s for the 1-processor SX-6. The IBM p690 turbo (32 CPUs) reaches 25.501 GByte/s, the p690 (16 CPUs) 20.051 GByte/s. As other examples named NEC the Itanium 800 MHz with 1.4 GByte/s and .891 GByte/s for the UltraSparc 750 MHz. The SX-6 with 8 CPUs has a Stream Triad bandwidth of 213 GByte/s. The SX-6i is delivered with the standard Unix operating system SUPER-UX with extensions, e.g. compilers that produce vector code automatically and other tools. NEC sees users in departments and those who have access to big machines and can prepare and optimize their codes on the minor machines. Application Areas for Vector Supercomputers NEC recognizes the usage of these machines in areas, where a high memory bandwidth is needed. Thus there is the automotive industry, e.g. Volkswagen, DaimlerChrysler, Renault, the aerospace industry, e.g. DLR (German Aerospace Research), NLR (Dutch), ONERA, EADS, and academic research, e.g. HLRS, Stuttgart, Max-Planck, Munich, CSCS, Manno Switzerland, and IDRIS, France. An example is a climate code, where the relationship of peak and application performance of the NEC SC-6 ( 8 GFlop/s) sums up to more than 180 GFlop/s compared to the peak of 512 GFlop/s. The IBM p690 reaches about 50 GFlop/s out of 512, this means less than 10% efficiency. NEC SX-series Sales Worldwide As of the end of December 2001 there are a total of 377 units ordered. The worldwide distribution is quite interesting: Country --- Units Japan ----- 212 Europe ----- 109 NA ----- 34 Asia ----- 12 Oceania ----- 6 SA ----- 4 German Climate Research Center (DKRZ) and the Earth Simulator NEC installed at DKRZ, Hamburg, 8 SX-6 nodes, and another as a hot standby. DKRZ is just running the 30 days acceptance test. The next step starts in the beginning of 2003, 192 processors with 1.5 TByte data. In Japan NEC installed all 640 nodes, 5120 processors, at the Earth Simulator. Now the multi-node application is tested. The operations start in the beginning of March. NEC expects to become No. 1 in the June 2002 Top500 list, published in Heidelberg at Hans Meuer’s Supercomputer Conference.