Intel True Scale Fabric Architecture for Enhanced HPC Performance Improved interconnect operability increases scalable performance for today’s HPC clusters There are two types of InfiniBand architectures available today in the marketplace, the first being the traditional InfiniBand design, created as a channel interconnect for the data center. The latest InfiniBand architecture was built with HPC in mind. This enhanced HPC fabric offering is optimized for key interconnect performance factors, featuring MPI message rating, end-to-end latency and collective performance, resulting in increased HPC application performance. Enhanced Intel True Scale Fabric Architecture – Offers 3x to 17x the MPI (Message Passing Interface) message throughput of the other InfiniBand architecture. For many MPI applications, small message rate throughput is an important factor that contributes to overall performance and scalability. Intel tested a number of MPI applications and found that they performed up to 11 percent better on the cluster based Intel True Scale Fabric QDR-40 (dual-channel) than the traditional InfiniBand-based architecture running at FDR (56 Gbps)   Improved end-to-end Latency – End-to-end latency is another key determinant of an MPI application’s performance and ability to scale. The Intel True Scale Fabric end-to-end latency is 50 percent to 90 percent lower at 16 nodes than the traditional InfiniBand offering available today. Increased Collective Performance – Critical for an MPI application’s performance and ability to scale. Intel True Scale architecture makes it possible to achieve significant collective performance at scale, without hardware based collective acceleration, resulting in 30 percent to 80 percent better collective performance for the three major collectives: All reduce, Barrier, and Broadcast. Faster Application Performance – Intel tested a number of MPI applications and found that they performed up to 11 percent better on the cluster based Intel True Scale Fabric QDR-40 than the traditional InfiniBand-based architecture running at FDR (56 Gbps). Conclusion The interconnect architecture has a cluster and the applications running on the cluster. Intel True Scale Fabric host and switch technologies provide an interconnect infrastructure that maximizes an HPC cluster’s overall performance. The Intel True Scale Fabric Architecture, with its onload protocol processing engine, connectionless implementation, and lightweight semantic-based PSM interface, provides an optimized environment that maximizes MPI application performance. With the use and size of HPC clusters expanding at a rapid architecture and technology extracts the most out of your investment in compute resources by eliminating adapter and switch bottlenecks. Read the full report :https://www.dropbox.com/s/7tygtfrup0eau7r/Intel-TrueScale_WP_CompleteArchitecturewithPSM.pdf