Understanding G80 behavior and performances

Cyril Crassin 1 Fabrice Neyret 1
1 ARTIS - Acquisition, representation and transformations for image synthesis
Inria Grenoble - Rhône-Alpes, LJK - Laboratoire Jean Kuntzmann, INPG - Institut National Polytechnique de Grenoble
Abstract : Criticality of optimization on GPU is very different than for CPU due to the extremely high performance contrasts. GPUs have "fast paths": Unintuitive depressing 1000x slowdown as compared to expectations or after a small change are often met when programming. Conversely, knowing these fast paths can lead to more than 1000x speedup. G80 series can be even less intuitive regarding this issue since it can very efficiently hide some unoptimalities (via threading) up to the point where suddenly it can no longer. To find and remain in these fast paths, it is important to understand how the GPUs work and behave. In particular, we were interested in how fragments are scheduled among G80 stream processors.
Type de document :
Rapport
[Research Report] LJK-INRIA. 2008
Liste complète des métadonnées

https://hal.inria.fr/hal-01643579
Contributeur : Fabrice Neyret <>
Soumis le : mardi 21 novembre 2017 - 15:02:44
Dernière modification le : mercredi 27 juin 2018 - 15:46:11

Fichier

Understanding G80 behavior.pdf
Fichiers produits par l'(les) auteur(s)

Identifiants

  • HAL Id : hal-01643579, version 1

Collections

Citation

Cyril Crassin, Fabrice Neyret. Understanding G80 behavior and performances. [Research Report] LJK-INRIA. 2008. 〈hal-01643579〉

Partager

Métriques

Consultations de la notice

219

Téléchargements de fichiers

45