The .8Tflop project at Columbia
Paper: 434
Session: F (talk)
Speaker: Gara, Alan, Columbia University, Irvington
Keywords: massive parallel systems
The .8Tflop project at Columbia
Columbia University
Igor Arsenin, Dong Chen, Ping Chen,
Norman H. Christ, George Fleming, Alan Gara (Nevis),
Chulwoo Jung, Adrian Kaehler, Steve Kasow,
Roy Luo, Catalin Malureanu, Robert D Mawhinney,
John Parsons (Nevis), ChenZhong Sui, Pavlos Vranas
SCRI at Florida State University
Robert Edwards, Tony Kennedy
Trinty College, Dublin
Jim Sexton
Fermilab
Sten Hanson
Ohio State University
Greg Kilcup
Abstract
The Columbia .8 Tflop project has designed and is nearing
completion of the assembly of a custom MIMD machine. This machine is
optimized for lattice problems such as QCD (Quantum Chomodynamics)
which require an enormous amount of floating point power.
The computer is based on 16384 processing nodes connected in a
4-dimensional torus. Each node is comprised of a DSP (digital signal
processor), a custom gate array, and conventional DRAM. The nodes are
extremely cost effective allowing for a large, scalable machine at a
cost far less than anything commercially available.
The simplicity of the hardware architecture enables a
reasonably simple software enviornment in which to develop massively
parallel applications. A commercial assembler and a C compliler for
the DSP (Texas Instruments) as well as a C++ compliler all enhance the
software environment.
A detailed description of the architecture as well as the
current status and measured performance will be given.