Computing on Multiple Graphic Cards


The research on parallelization of existing simulation codes to run on machines with multiple graphics processing units (GPUs) is led by Professor Michael Griebel, director of the INS and director of the Fraunhofer Institute SCAI.

"Our vision is to develop a massively parallel, completely multi-GPU based high performance molecular dynamics software package, as well as a massively parallel, completely multi-GPU based high performance fluid dynamics code," says Griebel. "Our customers from industry and research institutes will profit from our ability to solve general challenges of high-performance computing in this way."

Today, numerical simulations are indispensable in industrial production. Examples are the creation of new materials, the modeling of manufacturing process chains, and the simulation of material strength and fluid dynamics. However, these simulations require computing times from hours to days – even on high performance computers. This is why industry and science are very interested to shorten processing times.

Computing on multiple graphics cards promises an enormous acceleration of these simulations. NVIDIA’s CUDA parallel computing architecture, enables a dramatic increase in computing performance by harnessing the tremendous power of the GPU. Especially for software that is well suited for parallel computing, the graphics processor is faster than conventional CPUs by orders of magnitude. For example, the INS successfully ran the fluid solver package NaSt3DGPF on eight traditional processors coupled with eight graphics processors. Performing a benchmark study showed that the multiple GPU configuration was even slightly faster than a system using 256 conventional processors.

The researchers from INS and SCAI hope to gain similar effects from adapting the software package Tremolo-X for use on multiple graphics cards. Tremolo-X is used for the molecular dynamics of atoms or molecules. This software simulates materials at the nano scale, and therefore makes it possible to efficiently design new and innovative materials.

Computing on graphics cards not only promises an enormous acceleration of numerical simulations. The GPUs also require much less electricity, delivering a much higher performance per watt benefit. A particular computing task on a conventional parallel computer with 256 processors uses up to 70 kilowatts, compared to only 3 kilowatts on the machine with multiple GPUs. Furthermore, companies profit from GPU computing because the hardware is cheaper.; Source: Fraunhofer Institute for Algorithms and Scientific Computing SCAI