link to homepage

Institute of Energy and Climate Research (IEK)

Navigation and service

Parallel computing with CUDA

The programming package CUDA (Compute Unified Device Architecture) is an interface developed by Nvidia, which allows to outsource large parts of demanding computations from the CPU to the GPU, utilizing the specific architecture of graphic cards as an additional advantage. It leads to significant improvements of simulation runtime even when only consumer graphic cards are used. Comparing the graphical and central processing units one can see a major increase of ALUs (Arithmetic Logic Unit) in the GPUs while the control and cache structure is reduced. Using these ALUs in parallel increases the code performance as a result. CUDA is based on the programming language C but has been extended further since 2006 to support a wide range of languages, like C++, Fortran or Python. Therefore, developing new programs or implementing codes, which have been written for serial computations, is straightforward.
In our research we use GPU computing in phase field simulations, in which we investigating the bainite formation in alloys.