Supercomputing is the Formula 1 of computing, where speed takes precedence over everything else, including cost. Supercomputers are not just larger, faster mainframes, any more than Formula 1 race cars are just faster sports cars. There technologies like Infiniband rule, and ultra-low latency is a hot issue. 

So when the latest FDR Infiniband technology from Mellanox loses a head-to-head bake off against last year’s technology from rival QLogic, the shock is palpable. And that, reports Wikibon HTC Analyst David Floyer, is exactly what happened at the 2011 International Supercomputer Conference (ISC’11), the showcase for new supercomputer technologies, June 19-23 in Hamburg, Germany.

What happened is that supercomputer builder Appro won a competitive award for the Tri-Lab Linux Capacity Cluster 2 (TLCC2) to support the Lawrence Livermore, Los Alamos, and Sandia National Nuclear Security Administration weapons labs, which need the power for advanced work on fusion power. Mellanox bid against QLogic for the subcontract to provide the interconnect part of the cluster, but in a head-to-head test, the new Mellanox FDR Infiniband Technology was out-performed by Qlogic’s QDR InfiniBand, which is a last generation product. Qlogic is still developing its FDR-generation technology.

Floyer blames the investment Mellanox has received from Oracle for creating a slower Infiniband system. In supercomputing Infiniband is the technology of choice for high-performance systems, while Ethernet is becoming the interconnect of choice for lower performance, lower cost systems. Mellanox appears to have compromised the underlying latency of its FDR 56 Gb/s technology to boost Ethernet latency, influenced by Oracle’s focus on Ethernet.

Floyer’s article discusses this and its implications along with the other big news from the supercomputing conference.