Getting high utilization of peak GFLOPS in real applications in the Cray X1

This paper will show the advanced characteristics of the Cray X1 and discuss how they are used to achieve high-utilized GFLOPS on many real world applications. On most MPP systems, other than the Earth Simulator and the Cray Inc. X1, advanced scientific applications do not obtain a high percentage of peak performance, in some cases less than 2%. When this small percentage of peak is attained on the processor, one needs to have more individual processors to achieve a TFLOP of sustained performance. Larger numbers of processors result in a tremendous burden on the interconnect. Here again MPPs other than the Earth Simulator and the X1 do not have the interconnect to support the increased number of processors. Combining low processor performance with insufficient scaling results in less than desired performance for many applications. (author)