%0 Journal Article
%T The ¡°Chimera¡±: An Off-The-Shelf CPU/GPGPU/FPGA Hybrid Computing Platform
%A Ra Inta
%A David J. Bowman
%A Susan M. Scott
%J International Journal of Reconfigurable Computing
%D 2012
%I Hindawi Publishing Corporation
%R 10.1155/2012/241439
%X The nature of modern astronomy means that a number of interesting problems exhibit a substantial computational bound and this situation is gradually worsening. Scientists, increasingly fighting for valuable resources on conventional high-performance computing (HPC) facilities¡ªoften with a limited customizable user environment¡ªare increasingly looking to hardware acceleration solutions. We describe here a heterogeneous CPU/GPGPU/FPGA desktop computing system (the ¡°Chimera¡±), built with commercial-off-the-shelf components. We show that this platform may be a viable alternative solution to many common computationally bound problems found in astronomy, however, not without significant challenges. The most significant bottleneck in pipelines involving real data is most likely to be the interconnect (in this case the PCI Express bus residing on the CPU motherboard). Finally, we speculate on the merits of our Chimera system on the entire landscape of parallel computing, through the analysis of representative problems from UC Berkeley¡¯s ¡°Thirteen Dwarves.¡± 1. Computationally Bound Problems in Astronomical Data Analysis Many of the great discoveries in astronomy from the last two decades resulted directly from breakthroughs in the processing of data from observatories. For example, the revelation that the Universe is expanding relied directly upon a newly automated supernova detection pipeline [1], and similar cases apply to the homogeneity of the microwave background [2] and strong evidence for the existence of dark matter and dark energy [3]. Most of these discoveries had a significant computational bound and would not have been possible without a breakthrough in data analysis techniques and/or technology. One is led to wonder the astounding discoveries that could be made without such a computational bound. Many observatories currently have ¡°underanalyzed¡± datasets that await reduction but languish with a prohibitive computational bound. One solution to this issue is to make use of distributed computing, that is, the idle CPUs of networked participants, such as the SETI@HOME project [4]. It is clear that a number of common data analysis techniques are common across disciplines. For example, LIGO¡¯s Einstein@HOME distributed computing project, designed to search gravitational wave data for spinning neutron stars, recently discovered three very unusual binary pulsar systems in Arecibo radio telescope data [5]. These are far from the only ¡°underanalyzed¡± datasets from existing observatories, and this situation is expected to only compound as we look forward to an
%U http://www.hindawi.com/journals/ijrc/2012/241439/