Supercomputing continues to get faster, cheaper, and move available. Costs are dropping rapidly partially because of graphics processing units (GPUs) and their highly parallel architecture. NVidia continues to deliver high-performance Tesla GPU computing platforms for the masses with its Tesla Preconfigured Cluster.
The Tesla Preconfigured Cluster is designed as a scalable solution for midrange supercomputing applications. A typical four-node system starts with head node that’s the cluster controller. It is connected via a 24-port InfiniBand switch to the nodes that pair a 1U computer system with a four-GPU Tesla S1070 1U system. The S1070 connection is via x16 PCI Express links.
Pricing for this type of system starts around $50,000. It comes installed with Linux cluster software and CUDA cluster support. CUDA is NVidia’s programming environment for its GPUs (see “Go With The Flow—Dataflow, That Is”).
The Tesla S1070’s four 1.44-GHz GPUs house a total of 960 processing cores that have a peak double-precision floating-point performance of 345 GFLOPS. For single-precision work, the system delivers 4.14 TFLOPS. The four-node cluster delivers four times this performance. For many applications, the GPU approach offers significant benefits including lower power requirements, lower cost, and less space sometimes by more than a factor of 30.
Linux cluster software is available now, and Windows Server 2003/2008 will be available later in the year. Larger configurations are available.