Energy Consumption Analysis on Graphics Processing Units

In this paper, we propose to investigate the compromise of the relative gains between the computation time and the energy consumption on Graphics Processing Unit (GPU). The energy consumed by GPU can not be neglected anymore, even if code acceleration is the main target. We aim to propose a prediction model of both time computation and energy consumption of linear algebra operations. Numerical experiments on a worstation equiped with two GPUs GTX275, have been performed on a set of matrices arising from the finite element discretization of the gravity equation with real data issued from the Chicxulub crater. The results exhibit performance, robustness and efficiency of Alinea library, our research group library, compared to Cusp library in terms of speed-up and energy consumption for double precision number arithmetics.