As per the constant need to solve larger and larger numerical problems, it is not possible to neglect the opportunity that comes from the close adaptation of computational algorithms and their implementations for particular features of computing devices, i.e. the characteristics and performance of available workstations and servers. In the last decade, the advances in hardware manufacturing, the decreasing cost and the spread of GPUs have attracted the attention of researchers for numerical simulations, given that for some problems, GPU-based simulations can significantly outperform the ones based on CPUs. The objective of this book is first to present how to design in a context of GPGPU numerical methods in order to obtain the highest efficiency. A second objective of this book is to propose new auto-tuning techniques to optimize access on GPU. A third objective of this book is to propose new preconditioning techniques for GPGPU. Finally, an original energy consumption model is proposed, leading to a robust and accurate energy consumption prediction model.

• Presents step-by-step patterns for parallel programming on GPU
• Helps to implement efficient linear algebra operations on GPU
• Helps to implement efficient iterative methods on GPU
• Proposes new techniques to speed-up algorithms through auto-tuning on GPU
• Proposes new preconditioning techniques on GPU
• Proposed new approach to measure and to predict energy consumption of a scientific application on GPU