Last modified
December 6, 2002

  Seminar Abstract
Center for Data Intensive Computing



BlueGene/L - The next generation of scalable computing platform

BlueGene/L is a next-generation massively-parallel computing system designed for research and development into computational science targeted at selected applications of interest to the ASCI Tri-Lab community and its University Alliance partners. A select but broad set of science application areas have been identified as an initial focus for execution on BlueGene/L. BlueGene/L is a scalable ultra-computer targeted for 216 = 65,536 compute nodes. Each node consists of a single compute ASIC (application-specific integrated circuit) and SDRAM-DDR memory chips. The nodes are interconnected through multiple networks, one of which connects nearest neighbors, mapping the nodes into a 64x32x32 three-dimensional torus. The ASIC that powers the nodes is based on IBM's system-on-a-chip technology giving a very compact, low-power building block that is being used to create an extremely high compute-density system with very attractive cost performance. BlueGene/L is a cellular architecture in that the 512-node basic building block of the system can replicated in a regular pattern, with no introduction of bottlenecks as the system is scaled up. In many cases, one CPU of each dual-CPU node will be devoted to communications processing, yielding a nominal peak performance of more than 2.8 teraFLOP/s in a single cabinet, or more than 180 teraFLOP/s for the full system. For select applications that can effectively utilize both CPUs, the peak system performance will be over 360 teraFLOP/s. A teraFLOP/s is a trillion floating-point operations per second. At 360 teraFLOP/s, every man, woman and child on the face of the earth would need to perform about 60,000 calculations per second to equal this prodigious computational rate. Each set of 64 computing nodes attaches to a dedicated dual-processor I/O node for handling of communications. Using Gigabit Ethernet, a full system of 1,024 I/O nodes has the capability to drive up to 1 terabit/s of bi-directional bandwidth to disk or other resources. At 1 terabit/s, it would be possible to transmit the equivalent of the book collection of the Library of Congress in about three minutes.

URL links:
LLNL public web site for BlueGene/L


Top of Page



Copyright © 1999 Brookhaven National Laboratory ALL RIGHTS RESERVED
Comments/Sugestions about this site contact: Webmaster