Development of Real-time In-situ Visualization Methods for AMR Fluid Simulations on GH200 Tightly Coupled Architecture
Takashi Shimokawabe(Information Technology Center, The University of Tokyo)

This project develops real-time in-situ visualization methods for Adaptive Mesh Refinement (AMR) fluid simulations on the Miyabi-G supercomputer, utilizing the NVIDIA GH200 tightly coupled architecture. To tackle I/O bottlenecks, the research leverages unified memory and NVLink-C2C to offload visualization processing to Grace CPUs while Hopper GPUs focus on simulation, eliminating PCIe transfer overheads. By comparing this approach with PCIe-based in-transit transfers on TSUBAME4.0, the project aims to establish efficient visualization frameworks applicable to next-generation systems like FugakuNEXT.

研究紹介ポスター／最終報告書