Navigation and service

Supercomputer for Highly Complex Calculations and AI

It is Europe’s fastest supercomputer: Jülich’s JUWELS made huge gains in computing performance thanks to a new booster module. Researchers from all over Europe have been adapting and developing programs over recent months so that they run massively parallel on the more than 3,700 graphics processors (GPUs) in the booster.

“The enormous power of the new booster allows us to solve much bigger problems and perform much more demanding simulations than ever before,” says Dorian Krause, head of the High-Performance Computing Systems department at the Jülich Supercomputing Centre (JSC). The booster module also opens up new opportunities to use artificial intelligence (AI). 180 groups from different research fields use the upgraded modular JUWELS system. Their research evolves around fundamental questions relating to, for example, climate science, the environment, renewable energy, material properties or brain research.

Supercomputer im Forschungszentrum JülichCopyright: Forschungszentrum Jülich / TRICKLABOR

Drug discovery

“A very recent example in the current COVID-19 crisis is supporting simulations for drug development," said Prof. Thomas Lippert, head of the Jülich Supercomputing Centre (JSC). "Only the computing power of the booster enables our researchers to simulate the processes before, during and after a potential drug meets a receptor or protein realistically enough."

Normally to do this, scientists use programs that allow them to find out whether a potential active substance spatially and chemically matches particular binding sites on a receptor or another protein. However, the effect also depends on how long a substance is retained on the receptor. That is why researchers now want to use JUWELS to accurately simulate the entire dynamical process, in order to estimate the retention time and efficacy of a candidate drug. Molecule dynamics simulations of this kind pose huge demands in computing power and are now made possible by the JUWELS booster.

Simulation of groundwater flows

Another program, ParFlow, simulates surface, earth, and groundwater flows, all the while taking human factors such as groundwater pumping and irrigation into account. With the booster, the researchers will be able for the first time to carry out ParFlow simulations for Germany and Europe with the fine resolution that is required for elements such as individual slopes and river corridors. This is one of the requirements that needs to be met to produce forecasts of the water resources for each individual plot of agricultural land across Germany.

Investigating the effect of trace gases on the climate

One more example is MPTRAC. The program simulates the paths of trace gases in the upper atmosphere. One such trace gas is sulfur dioxide, which is released by sources such as industrial plants and volcanoes. Researchers want to investigate its effect on the climate. Satellite measurements do not completely reflect the distribution of the gases around the world. Thus, the data are cleverly combined with MPTRAC simulation data. Experts call this hugely compute-intensive procedure “inverse modelling”. Thanks to the JUWELS booster, the researchers can now take satellite measurements over ten years with around 100 volcanic eruptions into account.

Intelligently using resources

JUWELS is based on a highly flexible modular architecture developed by Forschungszentrum Jülich together with European and international partners like ParTec, Supercomputing Centres like Leibniz-Rechenzentrum in Garching as well as the University of Barcelona, now comprising a booster module closely interconnected with the cluster module. One example of an application that benefits from this architecture in a particular way, is the Terrestrial System Modelling Platform (TSMP), with which researchers predict such scenarios as over what length of time and with what volume of precipitation will the water resources of a region recover from periods of drought. The TSMP is made up of several computer models coupled together, including ParFlow, which in particular benefits from the booster. By contrast, the atmospheric model that is used in the TSMP is not optimized for the graphics card-based architecture of the booster. Therefore, the current version of the model will also be computed on the JUWELS cluster with its conventional processors (CPUs).

Strongest platform in Europe for artificial intelligence (AI)

The Jülich supercomputer also meets the special demands that learning software – artificial intelligence (AI) – puts on computers. Thanks to its new booster module, JUWELS offers the strongest platform in Europe for the use of artificial intelligence (AI). The booster unites around 12 million CUDA cores (FP64) on its more than 3,700 graphics processors based on NVIDIA's new Ampere architecture.

Also, the connected data storage systems and the input-output capabilities of the system are optimized for AI applications. While classic computer simulations generate a lot of data from a small input, it is often the other way around in the case of machine learning. At first huge amounts of data are read in and analysed to train an algorithm, while much less data is read out as the result.


The DeepACF project provides one example. Its purpose is to improve predictions of extreme and localized weather events such as storms and heavy rains using deep learning. To do this, it relies on training millions of model parameters using enormous volumes of meteorological data. The algorithm used is based on applications that originated in the field of video analysis. The approach is based on extending patterns from the past into the future. By contrast, conventional weather forecast models solve equations that are based on physical laws such as the conservation of energy, mass, and momentum.

Early Access Applications

Computing time on the JUWELS booster is precious. Already during the first test phase in late summer and autumn of this year, the first applications ran on the system. The so-called "Early Access Applications" provide useful information for code optimization and first scientific results.


Frank Frick