Rodrigo Bartolomeu, Rene Halver, Jan Meinke and Godehard Sutmann have received the PPAM Best Paper Award, in the category of PPAM 2026 Award for Workshops, for their work “Speeding up N-Body Simulation Using Low-Precision Data Types” during the 16th International Conference on Parallel Processing & Applied Mathematics (PPAM 2026), organized in Poznań (Poland) from 30 August 30 to 2 September 2026.
Abstract:
A large part of the performance improvements in recent Nvidia GPUs has focused on low precision floating point formats. In addition to half precision (FP16), which was already available for Volta GPUs, other floating point formats such as TensorFloat-32 (TF32) and bfloat16 (BF16) have received hardware support starting with Nvidia’s A100 card. These data types are sufficient for training neural networks and increase peak performance compared to single precision (FP32) by a factor of up to 8 or 16 for TF32 and FP16 or BF16 respectively. Most scientific applications, however, use double precision (FP64) to avoid problems due to the limited precision offered by FP32 or even lower precision formats. This prevents them from taking advantage of the increasing performance of lower precision data types. In this paper, we explore the usage of different lower precision formats for the N-Body problem to prepare for the use of tensor cores.
Further information available here
