Radio-Astronomical Imaging Acceleration for Energy Efficiency
DescriptionThe Square Kilometre Array (SKA) Telescope project, a highly complex initiative requiring four supercomputer facilities to process radio-astronomical software pipelines, presents a significant challenge in energy consumption and environmental issues' impact. Modern High-Performance Computing (HPC) clusters have introduced accelerators such as Graphics Processing Units (GPUs) and Field-Programmable Gate Arrays (FPGAs) that significantly improve energy efficiency compared to traditional CPUs by parallelizing scientific software. Imaging is a critical component of the SKA Science Data Processors (SDPs) and performs computations like Gridding and Degridding, which underperform on CPUs. We explore the possibility of using custom FPGA hardware to improve energy efficiency compared to CPUs and GPUs. Additionally, we employ the technique of reduced precision to enhance performance and efficiency. Despite the potential of FPGAs, GPUs remain the preferred option for this type of computation. As GPU architectures continually evolve to adapt to new AI and HPC software, it is essential to evaluate the performance of various hardware options before procuring new computing systems. In this study, we also optimize the performance of imaging computing motifs on several GPU architectures with interesting results. Optimizing GPU performance is critical in reducing energy consumption and minimizing the environmental impact of the SKA Telescope project.
TimeMonday, June 2615:30 - 16:00 CEST
Climate, Weather and Earth Sciences