Wolfram Library Archive

Courseware Demos MathSource Technical Notes
All Collections Articles Books Conference Proceedings
Title Downloads

Optimization and Application of GPU Calculations in Material Science

Grzegorz Korpala
Organization: Universitšt Bergakademie Freiberg

Wolfram Technology Conference 2015
Conference location

Champaign, Illinois USA

Modern Graphic Processing Units (GPU) provide in combination with a very fast Video Random Access Memory (VRAM) very high computational, outrunning the conventional combination of a Central Processing Unit (CPU) and Random Access Memory (RAM) in terms of parallel computing and calculation. Since the complex structure and programming code for such parallelization of computational tasks, only a minor number of commercially software is already using the benefits of a shared CPU/GPU approach. But since the performance of modern graphic cards is still increasing with new chips, compromising billions of transistors, improved calculation architecture and algorithms and large amount of fast VRAM, GPU computation is predestined to accelerate numerical computations by using its parallelization capabilities. The computer algebra system (CAS) Wolfram Mathematica is used for numerical calculation of a large Finite Difference Model (FDM). The CUDA-link feature of Mathematica was used to a chieve a parallel working environment with a parallelized computation on available CPUs with a also parallelization of calculations of Nvidia GPUs at the same time. It will be shown, that the calculation time can be reduced by using shared-memory and an optimization of the used block and/or register size to minimize data communication between GPU and VRAM. Results for diffusion, stress field and deformation field for a deformation sample will be shown, numerically calculated from crystal plasticity, with over four million of FDM nodes being calculated by each of the four used graphic cards.


1444771976.pdf (1.1 MB) - PDF Document
1445541120.zip (13.4 MB) - ZIP archive