Top placement for the Göttingen supercomputer: In the most recent issue of the Green500 list of the world‘s most energy-efficient supercomputers the new system “Grete“, which has just been taken into operation, achieved rank 1 in Germany and rank 12 worldwide. This placement is based on a benchmark run that was optimized for energy efficiency and conducted by the system’s provider, MEGWARE Computer Vertrieb und Service GmbH based in Chemnitz. On the TOP500 list of the world‘s fastest supercomputers Grete obtained a compute performance of 1.83 PFLOPs in the Linpack benchmark and achieved rank 470. “Grete“, named after Grete Hermann (1901-1984) who as a doctoral student of Emmy Noether made contributions fundamental for computer algebra, is a High Performance Computing cluster based on graphics processors (GPU) and extends the HPC system “Emmy“, which GWDG is operating for the University of Göttingen since 2018, by an especially powerful and energy-efficient component.
The GPU cluster is technically a new partition of the existing NHR system “Emmy“ whose operation for the North German Supercomputing Alliance (HLRN) began in 2018 with the installation of a first phase based on Intel “Skylake“ CPUs. In 2020 this system was extended with a second phase based on the “Cascade Lake“ generation. As NHR@Göttingen the University of Göttingen / GWDG are a member of the National High-Performance Computing (NHR) alliance and since the beginning of the NHR funding in 2021 using these system is possible across Germany for scientific research at Universities. Getting access is easy by applying for a personal account, for higher demands of compute time project proposals can be handed in, which are being reviewed quarterly.
“We always strive for offering the best possible service for our users. Grete complements the existing CPU partition with an energy-efficient GPU system and thus allows opening up to further application domains and run deep learning applications, for example, “ [translated from German] comments Prof. Dr. Julian Kunkel, deputy head of GWDG for High Performance Computing.
In Göttingen a strategy of maximum energy efficiency for running all HPC systems is being pursued. For designing the entire system the direct liquid cooling (DLC) concept, which has already proved itself for the NHR system “Emmy“, is a critical factor in minimizing the power demand of the cooling equipment. According to the Green500 benchmark which gave a result of 32.149 GFLOPs/W the new GPU system is the second most efficient one among the listed systems based on NVIDIA A100 40 GB GPUs. Among all listed NVIDIA-based systems it occupies rank 5.
About the top position of “Grete“ as the most energy-efficient supercomputer in Germany Axel Auweter, member of the management at MEGWARE, says: “The demand for energy-efficient systems is particularly high in Germany, due to the relatively high energy prices. Insofar we‘re happy to have succeeded in setting new standards in this regard. “ [translated from German]
Technical details about “Grete“
At its core the new system features 36 nodes, each equipped with two AMD Epyc 7513 CPUs and therefore 64 “Milan“ cores per node, 512 GB of DDR4 memory, two 1 TB NVMe SSDs and four NVIDIA A100 GPUs. Each of these GPUs offers 6,912 CUDA cores and 432 tensor cores as well as 40 GB of HBM2 memory. Within each node the GPUs are connected by NVlink on a shared NVIDIA HGX “Redstone“ board via SXM4 sockets for faster GPU-to-GPU communication. Finally, the GPU nodes are connected among each other by an InfiniBand HDR fabric with 2x 200 GBit/s per node, as well as to a local flash-based storage solution based on two DDN ES400NVX applicances with a total net capacity of approx. 130 TiB and finally also connected to the existing 8.5 PiB storage of the “Emmy“ system.
Prof. Dr. Julian Kunkel
Gesellschaft für wissenschaftliche
Datenverarbeitung mbH Göttingen
Phone: +49 551 39-30144