Assessing the relevance of APU for high performance scientific computing - Archive ouverte HAL Accéder directement au contenu
Communication Dans Un Congrès Année : 2012

Assessing the relevance of APU for high performance scientific computing

Résumé

Fusion APUs eliminate the PCI Express bus which bottlenecks many GPU applications. However, integrated GPUs in APUs are less powerful than discrete GPUs, and while the upcoming APUs will rely on a unified memory, the first APUs still have a distinct GPU memory partition. Hence, it is worthwhile to investigate the applications, as well as problem sizes, for which the GPU part of an APU may outperform a discrete GPU. In this talk we assess the relevance of Llano and Trinity APUs for scientific HPC via hardware and applicative OpenCL micro-benchmarks. We present detailed measurements of CPU-GPU data transfers and performance tests of highly optimized 3D finite difference stencils (a synthetic one and a more realistic and complex one for wave propagation simulations) depending on the frequency of snapshot retrieval. Our results show that integrated GPUs of upcoming APUs can outperform discrete high-end GPUs for medium sized problems with high communication requirements.
Fichier non déposé

Dates et versions

hal-02535670 , version 1 (07-04-2020)

Identifiants

  • HAL Id : hal-02535670 , version 1

Citer

Henri Calandra, Romain Dolbeau, Pierre Fortin, Jean-Luc Lamotte, Issam Said. Assessing the relevance of APU for high performance scientific computing. AMD Fusion Developer Summit (AFDS), Jun 2012, Bellevue, WA, United States. ⟨hal-02535670⟩
34 Consultations
0 Téléchargements

Partager

Gmail Facebook X LinkedIn More