Enabling Next-Generation Real-Time Pulsar Astronomy via Accelerated Processing on Tensor Cores

Abstract

Pulsars are neutron stars that can be observed in the form of periodical radio signals. Observations of pulsars provide valuable insights for a variety of astronomical research domains. These range from investigations on the nature of the neutron stars themselves, over investigations on the evolution of binary star systems, to tests of general relativity. Future radio telescopes will, due to their dramatic increase in observational capabilities, open up numerous possibilities for these research domains. However, this increase in capabilities comes at the cost of an immense increase in computational cost of the processing of the signal that is required to remove the interstellar propagation effects. This thesis investigates the possibility to utilize 16-bit precision Fast Fourier Transformations (FFTs) to accelerate the processing step of coherent dedispersion, which is the most computationally intensive processing step typically present in the processing pipelines of pulsar astronomy. To this end we have found that the standard library for the computation of FFTs on GPUs cuFFT is not suitable for this purpose, but that specific minor changes to an implementation can remedy the encountered issues of accuracy and overflow. Furthermore, this thesis seeks to explore the possibility to utilize tensor cores to accelerate the computation of 16-bit FFTs. The implementation developed for this purpose incorporated the changes mentioned that are required to retain suitable accuracy and the achieved accuracy demonstrates the general usability of 16-bit FFTs for this application. However, cuFFT's higher precision FFTs provide better accuracy than our 16-bit implementation, while also having a better runtime performance. An advantage of this implementation over cuFFT is the reduced memory requirement, as cuFFT at 16-bit precision is not suitable in this context. The more general question whether the overheads associated with the usage of tensor cores for the purpose of the computation of FFTs outweighs their benefits in terms of throughput, remains inconclusive and would require additional research. This is due to the state of optimization of the implementation, which does not allow for an accurate estimation of the respective performance ceiling. However, as the general usability of 16-bit precision FFTs for this application has been shown further optimizations of this implementation or the development of a new implementation with the goal to surpass cuFFTs performance at 32-bit might proof worthwhile.