Mixed Precision LU Factorization on GPU Tensor Cores: Reducing Data Movement and Memory Footprint - Archive ouverte HAL Accéder directement au contenu
Article Dans Une Revue International Journal of High Performance Computing Applications Année : 2022

Mixed Precision LU Factorization on GPU Tensor Cores: Reducing Data Movement and Memory Footprint

Résumé

Modern GPUs equipped with mixed precision tensor core units present great potential to accelerate dense linear algebra operations such as LU factorization. However, previous works have focused solely on improving speed, neglecting memory consumption. Indeed, state-of-the-art mixed half/single precision LU factorization algorithms all require the matrix to be stored in single precision. This is explained by the fact that simply switching the storage precision from single to half leads to significant loss of accuracy, forfeiting all accuracy benefits from using tensor core technology. In this article, we propose a new factorization algorithm that is able to store the matrix in half precision without incurring any significant loss of accuracy. Our approach is based on a left-looking scheme employing single precision buffers of controlled size and a mixed precision doubly partitioned algorithm exploiting tensor cores in the panel factorizations. Our numerical results show that compared with the state of the art, the proposed approach is of similar accuracy, up to twice faster, and with only half the data movement and memory footprint.
Fichier principal
Vignette du fichier
paper.pdf (544.17 Ko) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-02937325 , version 1 (13-09-2020)
hal-02937325 , version 2 (28-09-2022)

Identifiants

  • HAL Id : hal-02937325 , version 2

Citer

Florent Lopez, Théo Mary. Mixed Precision LU Factorization on GPU Tensor Cores: Reducing Data Movement and Memory Footprint. International Journal of High Performance Computing Applications, In press. ⟨hal-02937325v2⟩
345 Consultations
225 Téléchargements

Partager

Gmail Facebook X LinkedIn More