An architecture is described of a specialized PGM (Parallel Grid Multi
processor) supercomputer directed toward efficient solution of large s
cale numerical problems by grid methods for partial differential equat
ions of different types. The project is based on immediate mapping of
the structure of computations and data streams characteristic for grid
methods into the hardware architecture. The organization of two impor
tant PGM subsystems is described, namely, the distributed interconnect
ion network, and the set of parallel functional modules.