The Murchison Widefield Array (MWA) is a radically new type of radio telescope, with no moving parts, and dependent on computer power to create exquisite real-time wide-field images of the radio sky.
The expected data rates from this telescope is prodigious, of the order of 19 gigabytes per second. The telescope is in a remote location in Western Australia with no data connectivity or power lines. Given the vast data rates it is impractical to store all the data (a petabyte a day). Harvard researchers and colleagues from around the world are designing a Real Time System (RTS) that can consume the vast amount of data and reduce it down to manageable proportions. The telescope design demands that the RTS be capable of about 18TFlops and consistently generate a reduced data set every 8 seconds. The remote location of the facility also imposes a limit of the cooling and electrical power available to the RTS. This translates to a requirement to sustain 18TFlops on 20kW of diesel generator power.
To solve this problem Harvard scientists used general purpose graphics processing units (GPGPUs). Traditional CPU based computers were unable to meet the 8s deadline. Switching to a GPGPU pipeline resulted in a 18x speedup which brought the pipeline within the 8s requirement. In addition the GPGPU pipeline resulted in a 10x improvement of Flops/W which allows the system to fit into the 20kW power budget.
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License. Permissions beyond the scope of this license may be available at Attribution.