The SC2004 Conference on supercomputing heated up Pittsburgh as the weather got a little chilly last month in the Iron City. It was definitely the show for big iron, although high-performance computing has moved to racks of blade servers connected by high-speed switch fabrics.
Tom Dunning started off the show with a keynote speech entitled "What is the Future of Computational Science and High Performance Computing?" Based on the presentation and the show, it's obvious that things are getting bigger and better. High-performance computing is letting people tackle jobs that were impossible in the past. Benefits are rubbing off on the general computing world as well.
The show was full of academic presentations, from processor and cluster architectures to programming. It covered topics such as the Message Passing Interface and Unified Parallel C. The show also hosted the Fifth IEEE/ACM International Workshop on Grid Computing. Sessions touched upon high-performance computing for applications such as physics, biological sciences, and medicine plus data modeling and visualization.
The conference showed off new hardware and software as well. PathScale introduced a new HTX-based InfiniBand adapter (see "HTX InfiniBand Adapter Cuts Latency For Low-Cost, High-Performance Supercomputers," below), along with its OptiPath MPI Acceleration Tools and a new set of 64-bit compilers. Appro International introduced its next-generation blade server based on AMD's Opteron. New Appro Storage Systems supported the clusters, which can run Linux or Windows.
Meanwhile, Hewlett-Packard revealed a portfolio of hardware, software, and services for the high-performance computing (HPC) clusters market. Two entry-level Unified Cluster Portfolio systems are based on 1.6-GHz Intel Itanium 2 processors (see the figure). Two ProLiant systems are based on Intel Xeon processors with Extended Memory 64-bit Technology (EM64T). HP can deliver systems with more than 1000 processors. A 16-node HP Cluster Platform starts at $62,000.