Searching for binary pulsars is a computationally costly process that next generation radio telescopes will have to perform in real time, as data volumes are too large to store. In this talk we will explore how we have reduced the energy cost of an SKA-like Fourier Domain Acceleration Search (FDAS) implementation in AstroAccelerate, utilising a combination of mixed precision and dynamic frequency scaling on NVIDIA GPUs.

The goal of this talk will be to encourage you to consider exactly how much precision you need, as we have managed to save 50% of the execution time by using mixed precision with a completely tolerable (<3%) sacrifice in numerical sensitivity. With dynamic frequency scaling, the overall energy cost of FDAS was reduced by over 70%.