High-performance computing on streaming processors (GPUs)


Faithful computer simulations of dynamic processes in complex systems require immense computing resources. The rapid development of highly parallelised streaming processors, originally driven by the consumer market for graphics processing units (GPUs), has promised a boost in computing resources. Targeting these new architectures, we have developed the software HAL's MD package [1]. It provides an efficient and versatile tool to perform large-scale molecular dynamics simulations for applications in statistical physics. For example, the glassy dynamics of a supercooled binary mixture of 50,000 particles could easily be followed over 5 non-trivial decades in time with excellent conservation of momentum and energy [2], yielding dynamic correlation functions of very small noise level. Currently, a single GPU is able to replace about 64 conventional CPU cores and can accommodate simulations of about 1 million particles. A future development goal for the package is full scalability over many GPU nodes, delivering a unique potential for molecular dynamics simulations of huge systems with complex interactions.

[1]    Highly Accelerated Large-scale Molecular Dynamics package (2008–2015), http://halmd.org.

[2]    P. H. Colberg and F. Höfling, Comput. Phys. Commun. 182, 1120 (2011).