ENGLISH
La vitrine de diffusion des publications et contributions des chercheurs de l'ÉTS
RECHERCHER

Leveraging consumer-grade GPUs for efficient high-order computational fluid dynamics

Vermeire, Brian et Saghir, Muhammad. 2025. « Leveraging consumer-grade GPUs for efficient high-order computational fluid dynamics ». Communication lors de la conférence : CSME-CFDSC-CSR 2025 International Congress (Montreal, QC, Canada, May 25-28, 2025).

[thumbnail of 404 - Leveraging consumer-grade GPUs for.pdf]
Prévisualisation
PDF
404 - Leveraging consumer-grade GPUs for.pdf - Version publiée
Licence d'utilisation : Tous les droits réservés aux détenteurs du droit d'auteur.

Télécharger (46kB) | Prévisualisation

Résumé

Modern many-core hardware architectures, specifically Graphical Processing Units (GPUs), have advanced rapidly over the past decade. More recently, a clear distinction has emerged between consumer/workstation and data center GPUs. The former are characterized by low double-precision compute, low memory availability, and low cost. In contrast, the latter are characterized by high double-precision compute, high memory availability, and an order of magnitude higher cost. However, despite these differences, both consumer/workstation and data center GPUs have comparable single-precision compute and power consumption characteristics.In this talk, we will explore implementation details for a high-order unstructured solver (HORUS) that allow it to exploit the high single-precision compute of consumer/workstation GPUs while minimizing memory footprint using strategies including array aliasing, batched computing, and time-stepping optimizations. Results from scale-resolving simulations of turbulent flows will demonstrate the ability to run simulations with up to 250 million unknowns on a single GPU, comparable accuracy to double-precision data center GPUs for turbulent flows, and significant reduction in hardware cost. Ultimately, these results demonstrate that scale-resolving simulations of turbulent flows can be performed by leveraging only the single-precision compute capability of consumer/workstation grade GPUs, at an approximately 20-40 times reduction in hardware cost.

Type de document: Communication (Communication)
Informations complémentaires: Progress in Canadian Mechanical Engineering, Volume 8. Co-chairs: Lucas A. Hof, Giuseppe Di Labbio, Antoine Tahan, Marlène Sanjosé, Sébastien Lalonde and Nicole R. Demarquette.
Date de dépôt: 18 déc. 2025 14:38
Dernière modification: 18 déc. 2025 14:38
URI: https://espace2.etsmtl.ca/id/eprint/32155

Actions (Authentification requise)

Dernière vérification avant le dépôt Dernière vérification avant le dépôt