Infiniband Interconnect Preferred by 47% of Top500 HPC – InfiniBand Trade Association …

The InfiniBand Trade Association (IBTA) highlighted the latest TOP500 list results, revealing 235 of the world’s supercomputers rely on its standard interconnect technology, up from 225 a year ago.

This number represents 47% of the TOP500 list, ranking IB above Ethernet and proprietary solutions. The dominance of IB in the world’s most powerful supercomputers is due to the standard’s superior performance, efficiency and scalability.

IB connects 45% of the 73 most powerful petascale systems. Additionally, almost half of the systems on the list are used for research or in academic institutions, industries that require high performance and high efficiency. IB’s high bandwidth, low latency and CPU offloads make it the preferred solution for these systems, accounting for 61% of the research and 80% of the academic systems ranked on the TOP500.

“IB leads the industry in system efficiency, which is a key factor in controlling HPC costs,” said Bil Lee, IBTA marketing working group co-chair. “While some 10GbE-based cloud systems are breaking into the TOP500, they must do so with a massive amount of servers since these systems only achieve 20 to 40% compute efficiency. IB is designed to produce leading performance and up to 99.8% efficiency to fully utilize data center infrastructures at any scale-resulting in lower CapEx and OpEx, and higher data center ROI.”

Published twice a year and publicly available, the TOP500 list ranks the world’s most powerful computer systems according to the LINPAC benchmark rating system.

Other TOP500 takeaways include:

• Enhanced Data Rate (EDR 100Gb/s) IB, IBTA’s latest standard, is utilized in four systems
• 161 systems deploy Fourteen Data Rate (FDR 56Gb/s) IB, a 16% year over year increase

IB’s ability to carry multiple traffic types over a single connection makes it for clustering, communications, storage and management. As a result, the interconnect technology is used in thousands of data centers, HPC clusters, storage and embedded applications that scale from two nodes to a single cluster of tens-of-thousands of nodes.