大規模アプリケーションの高性能な実用的アクセラレータ対応手法
下川辺　隆史(東京大学　情報基盤センター)

This project aims to establish practical methodologies for porting CPU applications to accelerator-equipped supercomputers, focusing on the NVIDIA GH200-based “Miyabi-G” system. Through collaboration between domain and HPC experts, it addresses the trade-offs between performance, portability, and maintainability. The research includes five themes: optimizing unified memory for OpenSWPC, developing a lambda-based custom wrapper for N-body codes, improving Kokkos performance for fluid dynamics, introducing a Kokkos-Python interface via MLIR, and enhancing the Kokkos ecosystem. These efforts utilize the GH200’s tightly coupled architecture to advance scientific applications.

研究紹介ポスター／最終報告書