JCuda vectorized and parallelized computation strategy for solving integral equations in electromagnetism on a standard personal computer

Christophe Rubeck (1) , Bertrand Bannwarth (1) , Olivier Chadebec (1) , Benoît Delinchant (1) , Jean-Paul Yonnet (1) , Jean-Louis Coulomb (1)

1 G2ELab - Laboratoire de Génie Electrique de Grenoble

The paper presents a computation strategy for solving integral equations in electromagnetism. Nowadays, powerful programmable Graphic Processing Units (GPU) can be found in any standard computer. The paper investigates the benefits of the use of GPUs in addition to the CPU one in order to improve computation speed by using integral methods. Java language and the JCuda library, not often used in speed calculation by the computing community, has been used here. A 100 time speed-up is reported in matrix assembly between an optimized traditional CPU computation and a CPU+GPU one. FMM… Index Terms-Fast Multipole Method on GPU, JCuda computing, Pure Java ...

Domaines: Energie électrique

Christophe Rubeck, Bertrand Bannwarth, Olivier Chadebec, Benoît Delinchant, Jean-Paul Yonnet, et al.. JCuda vectorized and parallelized computation strategy for solving integral equations in electromagnetism on a standard personal computer. COMPUMAG 2011, Jul 2011, Sydney, Australia. ⟨hal-00610893⟩