Abstract:
A promising approach to the creation of high-performance computing systems with performance capability exceeding foreign supercomputers is considered. Manufactured basic modules and systems on theirs basis are discussed. These systems provide high specific performance due to a new organization of parallel calculations. A single-board basic module contains 320 elementary processors with 50 GFlops board performance. A system with four basic modules has an overall performance of 200 GFlops and occupies a volume of 40 dm$^3$. Hardware and software packages of new multiprocessor systems provide the linear performance growth for different application areas with a real performance above 50% of the peak performance. The paper was prepared on the basis of the authors' report at the International Conference on Parallel Computing Technologies (PaVT-2007; http://agora.guru.ru/pavt).