Moving away from the graphics processor, GPU supercomputer cluster visualization to break the bottleneck

Dedicated graphics processing chip in after the birth, in some high demand ahead of the visual field is usually used for multi-core supercomputer to achieve 3D. However,architectural visualization the development of super computers of today, the speed of the machine processing of data has long been over they can the I / O transfer speed. GPU-based clusters are becoming obsolete.
No GPU, can be visualized?
United States and other parts of the Argonne Laboratory researchers are trying to find a solution - they are doing programming work can make thousands of processor supercomputer visualization, rather than large amounts of data are transferred a dedicated graphics processor to render farms.
Argonne two computer scientists Tom Peterka, Rob Ross, and the University of California, Davis, two researchers at the peak (transliteration), Ma Kuang six (Davis, Department of Computer Science to increase Chinese professor) has Intrepid (an IBM Blue Gene supercomputer) completed the preparation of a software, making the data completely bypass the graphics chip into a reality. "It allows us and the data in the vicinity - in fact the same machine, the running visual experiments."architectural rendering Peterka said that the team's solution to avoid the data from the secondary cluster system generated when the transfer delay.
Peterka from North Carolina State University, John Blondin and Oak Ridge National Laboratory Anthony Mezzacappa obtained experimental data, representing the 30 steps of simulated stellar explosion, which is such as Argonne supercomputer to simulate the common data.
Typically, Intrepid generated visualization and post-processor in the 557 Teraflops of data around, which is the world's seventh fastest supercomputer, you need a separate graphics processor unit. Range where the Petaflop require more powerful computers, there are still more challenges.
"The greater the cluster system, I / O transfer problem will be more obvious." Peterka said. In a Petaflop supercomputer-level data generated is written to disk when needed unimaginable time. The reason is simple: When not wait to run to the next generation of supercomputer time, storage capacity and memory bandwidth does not keep pace.Free internet home
GPU drag petaflop costs?