APPLICATIONS

Microway will show a cluster running their new NumberSmasher-DC in their booth at Supercomputing 2005. The latest Microway 1U computational engine incorporates the powerful Dual-Core Intel Server Board SE7520JR2. Number Smasher-DC is the ideal platform to consider for applications that are floating point bound, that can take advantage of the large cache and floating point performance of the Dual-Core Intel Xeon processor 2.80 GHz. In addition to high end numeric and memory performance, the motherboard also comes standard with a PCI Express back end. Current functionality features of the Intel platform include hot swap memory, hot plug PCI-X, ROMB, local control panel, and Hyper-Threading Technology with the addition of dual core capabilities that can increase performance by as much as 40%. When this cluster is connected with a Microway FasTree InfiniBand switch, it becomes possible for MPI applications to hit interprocessor bandwidths of 900 MB/sec while taking advantage of the 3.8 microsecond latency of the latest Mellanox HCA's. The demo running on this cluster will be MMDS: Microway's MPI Diagnostic Suite used for testing the performance of an MPI cluster. The elements of this software include MPI Link-Checker which measures the bandwidth and latency between the nodes and displays performance graphs, and MPI Fast-Check, which can analyze the network of a very large cluster in seconds and point out faulty or underperforming nodes. Microway's CTO, Stephen Fried, commented, "The latest Intel architecture is especially interesting for applications that can take advantage of the SIMD floating point units found in Dual-Core Intel Xeon processors. These make it possible to execute up to four floating point adds per cycle per core, a real plus for users who are performing DSP computations like radix 8 FFT's that are bound by the speed of the floating point adder. Using the vector feature of the Intel compilers, it becomes possible for this dual processor and dual-core combination to crank out 16 adds per cycle, which at 2.80 GHz turns out to be 45 single precision gigaflops! The large caches also make it possible to store 64K complex data sets in core, dramatically reducing the time it takes to perform large FFT's. The dual-core Intel processors really shine in situations where code inefficiency is encountered due to the nature of the problems being solved or the software being run. Applications which do not take advantage of hand-tuned kernels will end up speeding up by 50% or more, because they do not run memory bandwidth limited. Problems where addressing has to pass through filters before fetches can be made, such as sparse matrix solvers, also show very large speed ups." "Intel is pleased to be working with Microway to increase the performance of their 64-bit computing platforms with Dual-Core Intel Xeon processors," said Jerry Braun, Product Marketing Manager at Intel Corporation. "Together we are committed to providing Microway customers with the best choices to help reduce overall total cost of ownership for today's HPC and data center needs." For more information Microway's new NumberSmasher DC products please visit the company's web site.