GPU DNS with IBM & RHEA
UPC’s Multiscale Fluid Mechanics Lab reported full GPU‑accelerated Direct Numerical Simulation of aerodynamic profiles using an Immersed Boundary Method implemented in their RHEA solver—demonstrating DNS fidelity on GPU hardware for high‑speed aerodynamic modeling. That kind of GPU DNS pushes turbulence‑resolved studies into more accessible turnaround times for design studies. (x.com) (x.com)
RHEA is an open‑source flow solver developed and maintained by Lluís Jofre’s Multiscale Fluid Mechanics Lab at UPC and documented in a Journal of Open Source Software entry and public GitLab repository. (zenodo.org)) A GPU port of RHEA using MPI+OpenACC was presented by Ahmed Abdellatif, with the non‑managed data approach delivering an overall ~6× speedup versus the CPU build and kernel speedups up to ≈26× (inviscid flux) and ≈16× (viscous flux) in published materials and the talk recording. (openacc.org)) The RHEA development and validation targeted heterogeneous HPC hardware, including a hybrid test node with an AMD Ryzen 9 3900XT plus 2 NVIDIA Quadro RTX 4000 GPUs and earlier large runs on the BSC cluster composed of IBM POWER9 nodes with 4 NVIDIA Volta GPUs per node. (upcommons.upc.edu)) “IBM” in the card context refers to the Immersed Boundary Method (IBM) used to represent complex bodies on Cartesian grids, a family of techniques reviewed in recent literature as suitable for high‑fidelity DNS around arbitrary geometries. (mdpi.com)) The RHEA paper states DNS computational cost scales roughly as Re^(37/14); combining that scaling with the reported ~6× wall‑time reduction implies, by inference, the same wall‑time budget could target a Reynolds‑number increase of about 6^(14/37) ≈ 1.97 (i.e., roughly double Re for equivalent turnaround). (zenodo.org)) The solver’s codebase, MPI+OpenACC acceleration notes, and the recorded GPU‑acceleration presentation are all publicly accessible for replication and benchmarking, with the project framing RHEA as an exascale‑aware hybrid CPU+GPU research platform. (gitlab.com))
