Technical Papers Archive

Architecture-Agnostic Exascale Star-by-Star Simulation of Our Galaxy


Authors: Keiya Hirashima (RIKEN Center for Interdisciplinary Theoretical and Mathematical Sciences (iTHEMS), Flatiron Institute); Naoto Harada (The University of Tokyo); Takayuki Saitoh (Kobe University, Japan); Kentaro Nomura (Preferred Networks, Inc.); Masaki Iwasawa (Matsue National College of Technology); Dennis Bollweg (Flatiron Institute); Kohji Yoshikawa (University of Tsukuba); Yohei Miki (The University of Tokyo); Tetsuro Asano (Universitat de Barcelona); Yutaka Hirai (Tohoku University of Community Service and Science); Takashi Okamoto (Hokkaido University); Junichiro Makino (Kobe University, Japan; Preferred Networks, Inc.); Matthew Bettencourt and Laura Morgenstern (NVIDIA Corporation); Mathis Bode (Forschungszentrum Jülich); and Michiko Fujii (The University of Tokyo)

Abstract: Resolving individual stars in simulations of the Milky Way (MW), where we live, is essential to understanding its formation and time evolution. Recent multi-wavelength, multi-modal observations have shown the detailed structures of the MW. Correspondingly, high-resolution simulations that resolve individual stars with multiphysics are required. Leveraging exascale systems, we present PHEASANT: star-by-star MW galaxy simulations including physics across multiple spatiotemporal scales with up to 5.4 trillion particles, ~10,000 times more particles than the state-of-the-art. On Aurora, our code processes 144 billion particles per second using 60,236 GPUs, approximately three times the throughput of a state-of-the-art N-body/smoothed particle hydrodynamics simulation, while its gravitational interaction kernel achieved 976 PFLOP/s in FP32. The code is highly portable, achieving 101 billion and 135 billion particles processed per second with 37,908 GPUs on OLCF Frontier and 20,736 GPUs on JSC JUPITER, respectively, while maintaining high kernel performance across Intel, AMD, and NVIDIA GPUs.



Back to Technical Papers Archive Listing