Speaker
Dr
Jongsoo Kim
(University of Cambridge)
Description
Phase I of the SKA will comprise of about 10% of the final
collecting area. It is likely that a software correlator will
be used for early science. Assuming 300 antennas, 4 GHz
bandwidth, and two polarizations, the required performance of
the software correlator for auto and cross correlations is 5.76
Peta FLOPS (Floating Operations Per Second), which is just
about three times larger than that of the current fastest
supercomputer in the world. Considering the near future
technology developments, we designed the software correlator
based on a cluster with multi-core processors and possibly
graphic processing units (GPUs). Since the data rate from an
antenna with a single pixel feed is 64 Gbit/sec (two
polarizations, 4bit Nyquist sampling of 4 GHz bandwidth), a
simplest possible network topology between antennas and
computer nodes is point-to-point connections with 100 Gbit/sec
ethernet. Due to the increase of number of bit for a sample
after its fourier transform, the bandwidth of interconnections
of computer nodes might be 4 or 8 times larger than the input
data rate to each node. The cluster is composed of 300 computer
nodes, and the performance of each node should be better than
about 20 Tera FLOPS. We also estimated the expected
performance, cost, and power usage of a 300 node GPU cluster
with current technology.
Primary author
Dr
Jongsoo Kim
(University of Cambridge)
Co-authors
Dr
Andrew Faulkner
(University of Cambridge)
Dr
Paul Alexander
(University of Cambridge)
