Gaming technology makes new telescopes flexible, cheap and fast
17 September 2015
Computer gamers will be familiar with the technology of Graphics Processing Units (GPUs): they power the ultra-fast, ultra-realistic graphics for modern computer games. Out of this mass market for incredibly high performance rendering of graphics in gaming, GPUs have emerged as a very serious technology for a wide range of complex high performance computing problems in science and engineering.
CAASTRO Associate Investigator Dr Stephen Ord (Curtin University) and the Murchison Widefield Array (MWA) team have placed GPUs at the heart of the SKA precursor. Just as a GPU has to rearrange, process and output data in real time for computer games, the MWA system uses GPUs to correlate the signals from 128 antennas scattered over several kilometres in real time, then send the visibilities 800km to the Pawsey Supercomputing Centre in Perth for storage and further (offline) processing. Other processing stages are being performed by bespoke hardware based on Field Programmable Gate Arrays (FPGAs), both of which are housed in general purpose rack mounted servers. This hybrid approach to the correlation task is required to handle approximately 8 TFLOPS (Tera FLoating point Operations Per Second). The MWA correlator currently generates 8.3 TB of data per day. Its design, signal path and processing elements are being described in a recent publication.
Projects such as the MWA and the SKA will be more flexible, cheaper and faster to develop because of these technologies. The Australian Government is funding Dr Ord to help design the SKA during its pre-construction phase, to scale the MWA GPU solution up by a factor of over 100 and meet the SKA data processing challenge. He works closely with global technology companies NVIDIA (GPUs), IBM (computing) and CISCO (networking), as well as international partners in Canada, Europe, the UK and India.