high performance computing on graphics processing units: hgpu.org

In the paper the result of application of artificial neural networks (ANN) for radionuclides transport modelling with surface runoff is presented. ANN with supervised training based on back propagation algorithm was used to predict radionuclides transport in the soil and on its surface. Application of ANN for substances migration modelling is worth using, because it works like a "black box" and can catch patterns, which are hard to formalize in mathematical form or can’t even be introduced in the form of equations. Factors, which influence mass transfer process in different types of soil and on the soil surface, were defined for micro and meso/macro scales. They were introduced in a numeric form, normalized and used as input layer neurons of ANN. The obtained radionuclides concentration was used as neuron of an output layer. ANN is an example of empirical modelling approach and requires computing operations of the same type with a large number of field measured datasets. That is why the ANN training process took a lot of time. This problem was solved using parallel computing, which had a sufficient accelerating effect due to the nature of the training process. One of the technologies for parallel computing is CUDA from Nvidia, which uses a video card graphical processing unit. CUDA-technology was applied to accelerate ANN training and computing based on back propagation algorithm, which runs on a CUDA-device up to 5 times faster than the same operations on a CPU (central processing unit).