The QCDOC project at BNL
QCDOC (Quantum Chromo-Dynamics On-a-Chip) is an international collaboration that was formed
to design a massively parallel supercomputer architecture tailored to the needs of
of Lattice Quantum Chromo-Dynamics (LQCD) simulations. The project has been centered at
Columbia University with contributions from the
UKQCD collaboration, the
RIKEN BNL Research Center (RBRC),
the IBM Thomas J. Watson Research Center and
Brookhaven National Laboratory.
Brookhaven National Laboratory (BNL)
currently hosts two large (12,288 nodes, 10 TFlops each) QCDOC machines:
one for the RBRC community funded
by RIKEN (Japan's Institute of Physical
and Chemical Research) and the other for the
US Lattice Gauge Theory community funded by the
US Department of Energy. A third large QCDOC machine for the
UKQCD community, funded by PPARC
(UK's Particle Physics and Astronomy Research Council), is hosted in Edinburgh, Scotland.
The QCDOC architecture has been designed to provide a highly cost-effective,
massively parallel computer capable of focusing significant computing resources
on relatively small but extremely demanding problems. The individual processing nodes
are PowerPC-based and interconnected in a six-dimensional, low-latency mesh network
with the topology of a torus. Each node, designed by our collaboration and built by IBM,
includes a single custom ASIC plus DDR SDRAM. It has a peak speed of 1 Gigaflops.
More information about this architecture can be found on the
QCDOC architecture
and publication web pages.
The QCDOC design is a natural evolution of that used in our earlier
QCDSP machines
(Quantum Chromodynamics on Digital Signal Processors). QCDSP incorporated a low-latency
four-dimensional mesh network to realize peak speeds of 1 Teraflops with 20,000 nodes.
QCDSP won the Gordon Bell
prize at Supercomputing 98 and was acknowledged as the world's
fastest non-commercial supercomputer. A separate evolution of the regular mesh architecture
of QCDSP is represented by IBM's Blue Gene/L supercomputer, which uses a three-dimensional
mesh, will incorporate up to 64,000 processing nodes and has a peak speed of 360 Teraflops.
|