Performance and accuracy of the matrix multiplication routines : CUBLAS on Nvidia Tesla versus MKL and ATLAS on Intel Nehalem

Philippe Estival (1) , Luc Giraud (2)

1 CFD
2 CERFACS - Centre Européen de Recherche et de Formation Avancée en Calcul Scientifique

Scientific computation relies heavily on 64 bits arithmetic. The evolution of the Graphical Processing Units to the status of massively micro-parallel vector units and the improvement of their programmability make them stand as powerfull algebraic coprocessors for many classes of matrix calculus. But on these processors inheriting from architectures dedicated to video processing in the first place, the space for double precision is narrow yet. One building block of dense linear algebra, the GEneralized Matrix Multiply Routine has been considerably accelerated on the GPU. We figure in this paper more details regarding its speed, but first and foremost, accuracy.

Mots clés: GPU; GEMM; Matrix Multiplication Routines; MKL; Atlas; CUBLAS

Domaines: Calcul parallèle, distribué et partagé [cs.DC]; Arithmétique des ordinateurs

Philippe Estival, Luc Giraud. Performance and accuracy of the matrix multiplication routines : CUBLAS on Nvidia Tesla versus MKL and ATLAS on Intel Nehalem. 2010. ⟨hal-00699377⟩