​One promising source of gravitational waves, not yet detected, is rapidly rotating neutron stars. Neutron stars are hyperdense leftovers from stellar evolution, formed from the core of stars of a certain weight class (not too light, not too heavy). Instead of collapsing all the way to a black hole, they stop just short, ultimately packing the mass of the Sun into a ball about 10 kilometers across. Neutron stars are known to spin rapidly, up to hundreds of revolutions per second, and they are so fantastically dense that even a small (millimeters high!) mountain will emit continuous gravitational waves (CWs) that are potentially detectable by LIGO.

However, detecting these gravitational waves is no mean feat. Although they are continuously emitted (as opposed to gravitational waves from merging neutron stars and black holes, which last no longer than a few minutes), they are very quiet, and digging these signals out of the noise is very challenging. The task is complicated by the fact that we often have to search over a wide range of gravitational wave frequencies and sky locations, since we do not know where a gravitational wave-emitting neutron star might be in the sky, or how fast it might be spinning. All of these facts combine to create a computational challenge which is formidable – many searches for these continuous gravitational waves are limited by the available computing power.

This motivates us to make these searches as computationally efficient as possible, and to take advantage of all resources available. One important resource which has so far been under-utilised in CW searches is graphics processing units (GPUs). Although initially designed, as their name suggests, for crunching numbers in service of producing 3D graphics, over the last twenty years they have proven themselves to be equally useful in many scientific applications, often providing significant speedups over CPUs. Most supercomputing clusters are now equipped with some number of high-powered GPUs for exactly this reason.

Our recent paper [1] presents the implementation of one very common method used in CW searches, the “F-statistic”, on GPUs. We show that, using our implementation, one GPU can do the work of 10–100 CPU cores, unlocking a significant new source of computational power to be used in analyses using the F-statistic. We also show that achieving these speeds does not require sacrificing sensitivity, which is extremely important given the faintness of the signal we’re looking for. Finally, as a demonstration of the utility of this new implementation in a real-world context we run a small search for continuous gravitational waves from four recently discovered neutron stars spinning between 200 and 400 times per second. The search consumes 17 hours of GPU time, in contrast to the 1000 hours of CPU time which would have been required to run the equivalent search.

This work will allow more CW searches to take advantage of the computing power offered by GPUs in the future and continue to push towards the first detection of continuous gravitational waves.

[1] https://dx.doi.org/10.1088/1361-6382/ac4616

Written by OzGrav PhD student Liam Dunn, the University of Melbourne.