A perspective trend in modern high performace computing is the application of graphic processor units (GPUs) to solve computational tasks previously solved only on CPUs. The GPU architecture initially targeted at 3-D graphics acceleration is now seen to perform well in solving scientific problems dealing with parallel processing of uniform data objects. The simulation tasks capable of obtaining gain from GPU-computing are abundant in the world of molecular simulations, one of such examples being molecular dynamics simulations.
Considerable progress in the feasibility of using GPU-computing was attained during the last several years with the introduction of appropriate programming languages (CUDA C, OpenCL) and marketing of specially designed GPUs (NVIDA TESLA).
However, the expense of getting the GPU power still exists in software modification and rethinking of software algorithms.
We are now conducting efforts on testing and adapting of molecular simulations software for using GPUs. The main point of interest being the use of heterogeneous GPU-clusters for MD simulations where two levels of paralellism have to be utilized: at the level of the GPU and at the level of the cluster nodes.
The work is funded under the project of Federal agency for research and innovation "Distributed computational systems for solving the problems of molecular bioengineering".
The actual information on our activities as well as the overview of the state-of-the-art GPU-accelerated technologies for molecular simulations is presented in the project blog http://molsim.org/gpublog (the blog is maintained in russian).