SCIENCE

The Frankfurt Institute for Advanced Research (FIAS) at the Goethe University Frankfurt has announced a 285.2 teraflops Linpack score at 736 megaflops per watt at a PUE of only 1.1, a key milestone toward the "Landes Offensive zur Entwicklung Wissenschasftlich okonomischer Exzellenz (LOEWE)."

Professor Lindenstruth, who holds a Helmholtz International Center (HIC) for the international large-scale Facility for Antiproton and Ion Research (FAIR) LOEWE professor position, shares the results as the LOEWE high-performance computing cluster (LOEWE-CSC) shows a performance in the first tuning efforts of 285.2 teraflops Linpack score. This includes single-node Linpack performance of 563.2  gigaflops, multi-node performance of 522 gigaflops and DGEMM performance of 621 gigaflops. This performance result is confirmed by its high rankings in the latest TOP500 and Green500 lists at www.top500.org and www.green500.org, respectively.

The LOEWE cluster development is based on real-life applications performance, utilizing green IT and best TCO directives to support the increasingly-important role played by large-scale numerical simulations in science and engineering, as well as business, financial markets and medicine. One of the key scientific projects is the HIC for FAIR center, lead by Frankfurt.

Performance for Real-Life Applications

With 18,432 AMD Opteron 6172 processor cores operating at 2.1 GHz based on the Supermicro 2U Twin server (SuperServer 2022TG-GIBQRF) with Mellanox ConnectX-2 at 40Gb/s QDR InfiniBand, 768 ATI Radeon HD 5870 GPUs and implementing key BLAS subroutines like DGEMM in assembly language allows for the greatest performance increase, exceeding the base-line performance by more than 360 percent.

"Supermicro is very pleased that this FIAS high-performance computing cluster (HPCC) has achieved outstanding rankings on the prestigious Top500 and Green500 lists," said Hermann von Drateln Business Development Director for Data Center and High Performance computing at Super Micro Computer  "To achieve these ranking using our 2U Twin servers provides further proof that our high-efficiency systems deliver superior performance-per-watt, performance-per-square-foot and performance-per-dollar."

"AMD applauds this outstanding achievement in performance and power efficiency by the new FIAS cluster using AMD Opteron 6172 processors and ATI Radeon HD 5870 GPUs," said John Fruehe, director of server product marketing, AMD.  "AMD has a significant history of HPC leadership, largely based on AMD Opteron processor technology, and is now helping drive a new era of heterogeneous computing in HPC based on world-class CPU and GPU technology."

The target ratio of much less than 500 kW for 285 teraflops for the entire infrastructure including cooling is a first in the industry, as it reduces the cooling power requirements to 10 percent of the computer power (PUE=1.1).  Specially designed datacenter elements allow for top efficiency cooling of the Supermicro 2U Twin GPU servers, which is paramount to the reduction in cooling required for the solution.

FIAS has teamed with HHLR-GU and HIC for FAIR to implement the LOEWE-CSC Cluster, a new HPCC platform, and is pleased to provide a breakthrough in the economics to high performance in real-life applications, green IT and best TCO directives.

This new platform is to support the increasingly important role played by large-scale numerical simulations in science and engineering, as well as business, financial markets and medicine.