GRAPE PROJECT - AN OVERVIEW

We axe developing a series of special-purpose computers, GRAPE (GRAvit y PipE), for the simulation of N-body systems, such as proto-planetary systems, globular clusters, galaxies, and clusters of galaxies. In si mulations of N-body systems, almost all computing time is consumed in calculating the gravitational force between particles. GRAPE calculate s the forces at high speed using hard-wired pipelines. The host comput er, which is connected to GRAPE, sends the positions of particles to G RAPE. Then, GRAPE calculates the force exerted on a particle and sends this value back to the host computer. Using the force calculated by G RAPE, the host computer then integrates the orbits of the particles. W e have already developed six different machines (GRAPE-1, GRAPE-1A, GR APE-2, GRAPE-2A, GRAPE-3, and GRAPE-3A), which are divided into low- a nd high-accuracy types. Those machines with odd numbers (GRAPE-1, GRAP E-1A, GRAPE-3, and GRAPE-3A) are among the low-accuracy type. They wer e designed for simulations of collisionless systems, such as galaxies, in which only the mean potential plays an important role. Simulations of such systems do not require high accuracy in the force calculation . GRAPE-1 and GRAPE-1A are machines made by wire-wrapping. GRAPE-3 is a highly parallel system with 48 full-custom LSI chips (GRAPE chip). E ach LSI chip has one GRAPE pipeline. The sustained speed of GRAPE-3 is 10 Gflops for a 200,000 particle simulation. The machines with even n umbers (GRAPE-2 and GRAPE-2A) are among the high-accuracy type. They w ere designed for collisional systems, such as globular clusters and pr oto-planetary systems, in which close encounters play an important rol e. In simulations of such collisional systems, we must calculate the f orce accurately. GRAPE-2 was the first machine of the high-accuracy ty pe. GRAPE-2A was designed for applications involving molecular-dynamic s simulations, as well as gravitational N-body simulations. GRAPE-2A c an calculate the forces of an arbitrary functional form using interpol ation tables. The computational speed of GRAPE-2A is 180 Mflops. We ar e now developing a highly parallel machine, GRAPE-4, in which many GRA PE pipelines (about 1,600) will work in parallel.