NVIDIA DGX-2 at IT4Innovations

4.3.2019

On Monday, March 4, we launched the first NVIDIA DGX-2 AI supercomputer in Central and Eastern Europe. The system was purchased by the national supercomputing centre IT4Innovations to tackle the most demanding AI tasks, achieving a peak performance of 2 PetaFLOPs. The launch ceremony was attended by the Governor of the Moravian-Silesian Region Ivo Vondrák, the Mayor of Ostrava Tomáš Macura and the Rector of the Technical University of Ostrava Václav Snášel.

Project for IT4I

The National Supercomputing Center will integrate the NVIDIA DGX-2 system into its infrastructure, which will be made available to a broad base of users from academic, research and commercial institutions in the Czech Republic and abroad. “Our main mission is to give Czech scientists access to cutting-edge technologies. This new system will enable them to keep up with the world in the field of artificial intelligence, which, together with digitalization, we consider crucial for the development of society. That’s why we decided to acquire this unique device,” says Vít Vondrák, director of IT4Innovations.

“NVIDIA DGX-2 is the world’s most powerful AI computing computer and IT4Innovations is the first institution in Eastern Europe where we have installed this system,” said Serge Palaric, Vice President at NVIDIA.

Installation

The actual installation went smoothly. The NVIDIA DGX-2 weighs 180 kg and required a server lift to install it into the rack. After a few hours of updating all system components and performing stress tests, they were able to start running calculations in Ostrava. The DXG-2 comes with the Docker environment pre-installed and the most commonly used AI frameworks and HPC applications tuned for this system can be easily downloaded from the NVIDIA GPU Cloud.

More information about NVIDIA DGX systems is available at https://mcomputers.cz/nvidia-dgx-systemy/.

CT report

report here

NVIDIA DGX-2

NVIDIA DGX-2 is equipped with sixteen of the most powerful accelerators for data centers, NVIDIA V100 Tensor Core GPUs. These accelerators are interconnected using revolutionary NVSwitches with a total throughput of 2.4 TB/s. Total HBM2 memory is 512 GB. NVIDIA DGX-2 also offers 30 TB of internal capacity on fast NVMe SSDs. Connectivity to the surrounding infrastructure is provided by eight 100 Gbps Infiniband/Ethernet adapters.

One NVIDIA DGX-2 can replace three hundred dual-socket servers with Intel Xeon Gold processors for training deep neural networks (ResNet-50). The NVIDIA DGX-2 comes with a pre-built and fine-tuned software environment to run the most widely used machine and deep learning frameworks with maximum performance. NVIDIA-DGX2 can also be used for traditional High Performance Computing (HPC), where it achieves a theoretical compute performance of 130 TFLOPS.