Innovatives Supercomputing in Deutschland
inSiDE • Vol. 8 No. 1 • Spring 2010
current edition
about inSiDE
index  index prev  prev next  next


QPACE (Quantum Chromodynamics Parallel Computing on the Cell) is a massively parallel and scalable computer architecture optimized for Lattice Quantum ChromoDynamics (LQCD). It has been developed in cooperation between several academic institutions (SFB TR 55 under the leadership of the University of Regensburg) and the IBM Deutschland Research and Development GmbH.

At JSC, a 4-rack QPACE system is installed. The building block is a node card comprising an IBM PowerXCell 8i processor and a custom FPGA-based network processor. 32 node cards are mounted on a single backplane and eight backplanes are arranged inside one rack, hosting a total of 256 node cards. The closed node card housing, which is connected to a liquid-cooled cold plate, acts as a heat conductor thus making QPACE a highly energy-efficient system.

To remove the generated heat a
cost-efficient liquid cooling system
has been developed, which enables high packaging densities. The maximum power consumption of one QPACE rack is about 32 kW. The 3D-torus network interconnects the node cards with nearest-neighbor communication links driven by a lean custom protocol optimized for low latencies. For the physical layer of the links
10 Gigabit Ethernet PHYs are used providing a bandwidth of 1 GB/s per link and direction.

In the context of PRACE work-package 8, promising technologies for future supercomputers have been evaluated and QPACE was identified as one of the advanced prototypes. In order to extend the range of applications for QPACE beyond LQCD, an implementation of the high-performance LINPACK (HPL) benchmark was a first step. Here, different from QCD applications, the transfer of large messages and collective operations must be supported efficiently. This required extensions of both network communication and the software stack.

A special FPGA bit-stream and a
library of MPI functions for message passing between compute nodes has been developed at JSC in co-operation with IBM. The HPL benchmark produced excellent results (43.01 TFlop/s on 512 nodes) and the 4 rack system in Jülich corresponds to an aggregate peak performance of more than
100 TFlop/s.

Providing 723 MFlop/s/W QPACE was recognized as the most energy-efficient supercomputer worldwide and ranked top in the Green500 list at the Supercomputing Conference 2009 in Portland, Oregon.


• Willi Homberg
Jülich Supercomputing Centre, Forschungszentrum Jülich

top  top