New research infrastructure for high-performance computing
The new computer infrastructure for artificial intelligence (AI) research at the Heinrich Heine University Düsseldorf (HHU) has cost more than 800,000 euros and is now installed and ready for operation. Depending on the application, performance increases 8 to 450 times. This allows new research questions to be addressed in areas such as brain research, pharmaceuticals and computational linguistics.
In 2019, the HHU successfully raised funds from the Federal Government and the State of NRW within the framework of a large-scale equipment application reviewed by the German Research Foundation. The high-performance computers purchased with this funding have now been put into operation at the Center for Information Technology (ZIM).
"In addition to smaller extensions, four new high-performance computing units designed for AI research are starting up," says Philipp Rehs, a member of the ZIM's High-Performance Computing (HPC) department. "We have the latest hardware, which has only been on the market for three months.
Prof. Dr. Axel Bucher, as CIO (Chief Information Officer) responsible for the strategic orientation of IT: "Thanks to the foresight of our HPC team, we are now one of the first in Germany to have by far the most powerful systems for AI applications in the world. This is a significant advance from which many at our university will benefit".
Depending on the field of application, the computing power increases significantly: For so-called "double-precision operations", which achieve very high accuracies, the performance increases eightfold to 310 Tera-FLOPS; one Tera-FLOP corresponds to 1 trillion computing operations per second. For standard AI applications, which usually operate with lower accuracy, the computing power is now almost 10 peta-FLOPS, or 10 quadrillion computing operations. This means that the AI computing power is 450 times higher than that of previous computers at the HHU.
Dr. Stephan Raub, team leader HPC at ZIM: "With the new hardware, users can now tackle research questions that were not possible with the previous equipment at the HHU". For example, the research group of Prof. Dr. Svenja Caspers at the Institute of Anatomy I wants to simulate the folding of the brain. Prof. Dr. Holger Gohlke from the Institute of Pharmaceutical and Medicinal Chemistry wants to carry out drug simulations, including a focus on the treatment of corona viruses. And the computational linguist Prof. Dr. Laura Kallmeyer will develop algorithms for machine text comprehension and automated text translation.
Professor Dr. Martin Mauve, Vice Rector for Digitization and Director of the Heine Center for Artificial Intelligence and Data Science, is also enthusiastic: "This is a very important step for all scientists at the HHU who use artificial intelligence methods for their work. It is particularly important that this technological leap is really available to everyone who is active in this field".
The computers developed for AI applications are based on so-called graphics processors, or GPUs for short, which are particularly powerful for highly parallelizable tasks. The new "Nvidia DGX A100" computers each have eight GPUs of the latest generation, which are networked together very efficiently. These not only have very high computing power, but are also energy efficient, which benefits power consumption and the necessary cooling capacity. Furthermore, they can be easily integrated into the existing supercomputer infrastructure at ZIM, especially into the high-performance memory system that supplies the AI computers with data.
If programs do not require the full computing power, "partitioning" makes it possible to split the system into several computers working in parallel, each running different applications. In this way, the infrastructure can be used to full capacity. In addition, the data security aspect plays a major role in the new hardware, since the human genome is increasingly being analyzed with the systems.