Evaluating Architectures for Application-Specific Parallel Scientific Computing Systems

In this work, we examine the computational efficiency of scientific applications on three high-performancecomputing systems based on processors of varying degrees of specialization: an x86 server processor, the AMD Opteron; a more specialized System-on-Chip solution, the BlueGene/L and BlueGene/P; and a configurable embedded core, the Tensilica Xtensa. We use the atmospheric component of the global Community Atmospheric Model to motivate our study by defining a problem that requires exascale-class computing performance currently beyond the capabilities of existing systems. Significant advances in power-efficiency are necessary to make such a system practical to field.