The Machine Learning Computing Center of the German Research Center for Artificial Intelligence (DFKI) is increasing its computing power from the current 20 petaFLOPS to 45 petaFLOPS. At the end of July, the research center put the first Nvidia DGX A100, a new high-performance computer for AI algorithms, into operation. Four more are to follow, making DFKI a leading provider of ML with the new DGX A100 systems. Artificial intelligence is a megatrend in the industry.

The third generation of the Nvidia DGX system offers 5 PetaFLOPS of performance with eight Nvidia A100 Tensor Core computing accelerators each. One PetaFLOP corresponds to 1 quadrillion computing operations per second. By way of comparison, if every person in the world were given a pocket calculator and each person were to perform 125,000 calculations within one second, the computing power would correspond to around one PetaFLOP.

At the same time, the energy requirements of the data center are being further optimized. While previous systems required around 5 kW per PetaFLOP, the DGX A100 requires around 1.2 kW per PetaFLOP.

DFKI's AI infrastructure uses Nvidia's Mellanox InfiniBand network to connect the DGX systems to an ultra-fast, low-latency fabric, enabling multi-system AI training and providing the fastest time to solve computational problems.

Making complex AI algorithms available for industry

DFKI is one of the world's first users of the new systems. This will further accelerate AI research into learning systems and their explainability and make complex AI algorithms available for practical use in industry.

"High-performance hardware is a central foundation for data-rich and computationally intensive AI methods," says Prof. Andreas Dengel, Managing Director and Head of the Smart Data & Knowledge Services research area in Kaiserslautern. "Due to the immense increase in data volumes in the most diverse fields of application, many of our project questions and also the market demand the optimal combination of high-performance AI computing systems and sophisticated algorithms."