August 15, 2006
[ return to the news list ]
HPC Wire (05/26/06) Vol. 15, No. 21,
Researchers at the Lawrence Berkeley National Laboratory tested the STI Cell processor in several scientific applications, comparing its performance with other processor architectures. They presented their paper, "The Potential of the Cell Processor in Scientific Computing," at the recent ACM International Conference on Computing Frontiers. "Overall results demonstrate the tremendous potential of the Cell architecture for scientific computations in terms of both raw performance and power efficiency," the researchers wrote. "We also conclude that Cell's heterogeneous multi-core implementation is inherently better suited to the HPC environment than homogeneous commodity multi-core processors." Cell, the product of a partnership of Sony, Toshiba, and IBM, combines substantial floating point resources with software-controlled memory hierarchy to process complex numerical algorithms. Unlike traditional multi-core designs, Cell uses a standard high-performance PowerPC core that runs eight SIMD cores called synergistic processing elements, each of which contains a local memory, a memory flow controller, and a synergistic processing unit. Though its architecture is a marked departure from traditional designs, Cell is especially intriguing because it can be mass-produced at a competitive price. The researchers evaluated Cell's potential to serve as the engine of high-end parallel systems by testing its performance in numerous scientific computing environments, including dense matrix supply and sparse matrix vector multiply. Cell has a three-level memory architecture that offers more predictable performance and greater memory bandwidth for long block transfers. On average, Cell tested eight times more efficient and powerful than Itanium and Opteron processors.
For the complete article, see http://www.hpcwire.com/hpc/671376.html
Posted by rab5 at 09:37 PM
[ return to the news list ]
