Most Powerful Graphics Computing

Highlight during SC|05 in Seattle: Fraunhofer Visualization Software on an IBM Cell Cluster Prototyp.

One highlight of the International Conference on High Performance Computing SC|05 in Seattle will be the visualization software PV-4D of the Fraunhofer Institute for Industrial Mathematics ITWM. The software has been installed on a prototype of an IBM cell blade cluster and will be presented to a large public for the first time. A successful test run was performed during the Fraunhofer Annual Meeting in Magdeburg, Germany on October 19 this year.

The visualization and analysis of extremely large amounts of data from the fields of scientific computing, image-rendering methods in medical technology, or oil exploration can be represented by easily understandable realtime animations thanks to PV-4D. The cell processor is now rendering things even faster: it is equipped with nine cores and currently represents the chip with the highest performance available for the mass market.

The professional visualization environment has been developed at the Competence Center High Performance Computing of the ITWM by Dr. Carsten Lojewski and his team and has been awarded the Joseph-von-Fraunhofer-Preis 2005. PV-4D is a pure software solution and explicitly works without the support of graphic cards. Time-critical functions use the vector units and the multi-pipes of modern processors for parallelization. The software works especially effectively if it is running on a cluster system. The network support (Gbit-Ethernet, Myrinet, Infiniband) is directly integrated into the parallel software, thus guaranteeing minimum latencies and large bandwidths.

A further central element of PV-4D is the decentralized software control. A server process is responsible for the computation of images and allows for the integration of simple viewers, which are installed on a laptop within the office network or on front-end computers of a virtual reality environment. The image composi-tion, which is usually done by special hardware, is also completely realized within the software.

The software is the tool with the highest performance available worldwide for a fast and interactive representation of gigantic amounts of data. The performance of the current cluster systems even exceeds the strongest special graphics computers, thus al-lowing for the interactive visualization of more than one terabyte of data. This performance is also convincing industrial partners, such as DaimlerChrysler, Shell AG, or the Berkeley Labs in California.

During the Supercomputing in Seattle (November 12-18), the new PV-4D Ray Tracing Kernel will be shown. It allows for the first time volume rendering of complex seismic data in real time. This opens up a new world of interaction for geophysicists, facilitating the analysis of data for the exploration of new oil wells.

The current implementation on the cell processor exclusively uses the vector units of the individual SPEs for the so-called SIMD-Shaft Ray Tracing; the available PEs are responsible for the parallel or-ganization of data. In such a way, 128 parallel rays can be com-puted at first-order coherence on each cell processor, and at sec-ond-order coherence 32 parallel rays with direct hardware sup-port. At the IBM booth in Seattle, 16 cell processors will be com-bined to form of a high-performance ray-tracing cluster.