Home > MSc Dissertation: Arjun Radhakrishnan

MSc Dissertation: Arjun Radhakrishnan


Radhakrishnan, Arjun. Accelerating Pulsar Dedispersion for MeerKAT on Graphics Processing Units. MSc Dissertation. Department of Electrical Engineering, University of Cape Town, 2010.


This thesis aims to investigate the viability of using Graphics Processing Units (GPUs) usually found in consumer graphics cards, but increasingly used for scientific computing, to accelerate computationally intensive radio astronomy signal processing tasks. An eventual aim is to see some of the results of this research being implemented in the signal processing pipeline of the Karoo Array Telescope (MeerKAT).

The specific task chosen was dedispersion of received signals from pulsars – a kind of star that appears to emit pulses of broadband electromagnetic radiation. Dispersion is a common phenomenon affecting extraterrestrial radio signals. It is caused when an electromagnetic wave passes through the ionised Interstellar Medium (ISM), which delays lower frequencies slightly more than higher ones. For faint astronomical sources such as pulsars, this leads to a smearing effect that can cause the signal to be lost beneath the radio frequency interference and receiver noise. To ensure detection, dispersion effects must be removed. The method of doing so being investigated here is called coherent dedispersion.

Coherent dedispersion aims to correct for the effects of the ISM by treating it as a unity gain phase delay filter. The received, dispersed signal is passed through a filter that has the inverse behaviour to the ISM to obtain the original signal. Programmatically that involves sampling the signal at twice the Nyquist rate, performing a Fast Fourier Transform (FFT) on the sampled data, multiplying each sample with the inverse transfer function of the ISM and finally performing an inverse FFT to obtain a time series signal again, with the dispersion effects removed.

This was accomplished on a dual GeForce GTX 285 system. A speedup of 23x was observed for a simulated dual polarisation signal at 1450MHz with a bandwidth of 50MHz compared to an Intel Core i7 920 processor. Speedup as the number of GPUs increased was limited at lower FFT lengths by a large thread initialisation overhead, but beyond a point showed up to 65x better performance. This will allow real-time dedispersion for a relatively low cost and with better power efficiency than a conventional CPU cluster, which is especially applicable to MeerKAT since the site of the radio telescope is in a remote part of the Northern Cape province.

It was concluded that GPUs could greatly help accelerate computation for MeerKAT and should be a part of the processing system to enable better, faster science.