An Early Performance Evaluation of the SiCortex SC648

We report the results of an early performance evaluation of the SiCortex SC648 that contains 648 processor-cores. Performance was measured using microbenchmarks to evaluate node and network performance, and with real-world parallel scientific applications. The SiCortex machines are notable for a number of architectural innovations, including the use of a custom, very low power, six-core processor; a state-of-the-art network interconnect; and the use of a degree-three (three input and three output channel per node) Kautz graph interconnect topology, together enabling high compute density, low power relative to potential compute capability, and fault-tolerant, relatively low contention communication. The purposes of this performance evaluation are two-fold: first to measure absolute performance of the SiCortex SC648 at both node level and with wellcharacterized parallel applications; and second, to obtain an indication of how the systems compare to nominally competing processors—current generation quad-cores from AMD and Intel—on scientific applications of interest.