HP and Caltech power-up 256-CPU Exemplar X-class system

Palo Alto, 17 November 97 The Hewlett-Packard 256-CPU Exemplar X-Class system with 64 gigabytes of main memory and 1 terabyte of disk space is up and running at the California Institute of Technology. According to Caltech it is the biggest shared-memory cache-coherent system ever put together. The installation of the 256-CPU Exemplar system at Caltech signals the completion of phase one in a three-phase collaborative project with Caltech. This project, initiated a year ago, was announced officially at Supercomputing '96.

The Exemplar system is providing computing applications for Caltech and affiliated research projects at the Jet Propulsion Laboratory (JPL) and the US National Partnership for Advanced Computational Infrastructure (NPACI). HP's Exemplar system is being used as a computational resource for the scientific, academic and engineering communities.

The system is applicated in such projects as global-climate modeling, the design of new fusion reactors for etching microcircuits in future microprocessors, aeronautics, and space/science data gathering and computing. For example, the system will be used to study the collisions between electrons and molecules that drive the chemistry in plasma reactors used in microelectronics manufacturing.

The Exemplar system also is being used in Caltech's continued collaboration with JPL on large-scale computing. JPL is providing funding for and says it will use 50 percent of the cycles of the system for space- and planetary-science projects, such as building an accurate gravity map of Venus. NPACI funded the deployment of the Exemplar X-Class system at Caltech to provide the academic research community with access to this system.

The 256-CPU system is composed of 16 nodes, with each node having 16 processors. The system supports globally addressable architecture and is cache-coherent using only hardware. The nodes are connected using HP's toroidal interconnect. The resulting system is characterized as a cache-coherent, nonuniform memory-access (ccNUMA) architecture. Additionally, the system supports the usual standard interconnects to the outside world: 100Mb Ethernet, ATM, FDDI and HIPPI.

Sandra Wermer