News Briefs - Hardware:

TOTS: Teraflops Off-The-Shelf To Be Demonstrated At SC03

TOTS (TeraFlops Off-The-Shelf) is enabled by the significant performance gains delivered by the marriage of Intel standard servers and InfiniBandSM Technology. As demonstrated with existing InfiniBand clusters it is now simple and easy to deploy a cluster of 128 or more InfiniBand connected servers that deliver over 1 TeraFlop of performance. TOTS is a play on the term: COTS or Commercial-Off-The-Shelf product. This term is widely used in the embedded and military markets to identify industry standard products that greatly reduce the cost to customers by embracing industry standard components and moving away from proprietary solutions. InfiniBand is a COTS technology that is enabling the clustering of industry standard servers into powerful compute systems.

Thanks to the InfiniBand architecture and fast Intel processors it is now possible for almost any university, lab, or commercial research facility in the world to deploy Teraflop levels of computing power to solve truly complex issues. This powerful combination creates an inflection point in computing. The high bandwidth, low latency 10 Gb/sec InfiniBand interconnect enables today's industry standard servers to be clustered together, with open software, to create massive amounts of compute power quickly and inexpensively. And it can be done quickly, in a matter of weeks, as all the components are readily available in the market.

It is estimated that a One TeraFlop cluster can be deployed today for well less than $1M. The most recent example is Virginia Tech who recently deployed a 10+ Teraflop InfiniBand cluster from industry standard servers for only $5.2M. That represents less than $600K per Teraflop.

A demonstration of TOTS will be provided at SC 2003 in the Mellanox booth (#1011). The demonstration consists of 192 Intel based servers clustered with high performance InfiniBand interconnect, housed in a total of only three racks. The cluster demonstrates just how easily this level of performance can be delivered anywhere, anytime. It is a multi-vendor collaboration with equipment and software provided by: ATI Technologies, Callident, Ciara, InfiniCon Systems, Intel Corporation, Mellanox Technologies, Ohio State University, Rackable, Sandia National Labs, The SourceForge InfiniBand Open Source Project, Topspin Communication, and Voltaire.

In addition to the SC2003 TOTS demo and the Virginia Tech cluster, Mississippi State University and the Sandia 128-node Intel Xeon cluster are deploying InfiniBand clusters that clearly demonstrate the ability to deploy enormous computational power, simply, easily and affordably.

At 1 trillion mathematical operations per second, a TeraFlop was once the Holy Grail of supercomputing. This milestone was first achieved only five years ago by Cray Computer at a cost of more than $80 Million. Even just two years ago there were only 12 computers in the whole world that ranked over 1 Teraflop and these systems cost on average greater than $20M per teraflop. Now, thanks to Intel and InfiniBand this level of performance can be delivered anywhere, anytime for less than 1/10 the cost.