Performance Evaluation of Adaptive-Precision SpMV with Reduced-Precision Formats

Stef Graillat (1) , Fabienne Jézéquel (1, 2) , Théo Mary (1) , Roméo Molina (1, 3) , Daichi Mukunoki (4)

1 PEQUAN - Performance et Qualité des Algorithmes Numériques
2 Université Paris-Panthéon-Assas
3 IJCLab - Laboratoire de Physique des 2 Infinis Irène Joliot-Curie
4 RIKEN CCS - RIKEN Center for Computational Science [Kobe]

This study explores the potential for performance improvement in the adaptive-precision sparse matrix-vector product (SpMV) (Graillat et al. 2023) by using reduced-precision formats other than the IEEE standard ones. In addition to FP32 and FP64, we consider to utilize reduced-precision formats of 8/16/24/40/48/56-bit length. Our evaluation compares the performance of four- and nine-precision versions with that of the existing two-precision only version using FP32 and FP64 and demonstrates the effectiveness.

Mots clés: sparse matrix-vector product (SpMV); mixed precision; adaptive precision; reduced precision

Domaines: Informatique [cs]

Stef Graillat, Fabienne Jézéquel, Théo Mary, Roméo Molina, Daichi Mukunoki. Performance Evaluation of Adaptive-Precision SpMV with Reduced-Precision Formats. 2023. ⟨hal-04261073⟩