Benchmark Tests on a Silicon Graphics R8000-Based Workstation

This note describes a performance evaluation of a workstation based on the Silicon Graphics R8000 microprocessor running at 75 MHz. The evaluation is based on a variety of Fortran77 benchmark codes that have a range of computational characteristics and that represent Los Alamos National Laboratory applications. Benchmark data from the R8000 system are compared with those from an IBM RISC System/6000-590 and with two older Cray Research vector processors. Only single-processor results are considered. The results of the study echo that of previous microprocessor/vector processor comparisons: for non-vectorizable problems microprocessors can run at favorable speeds relative to the vector processors (although on these codes the RS/6000 is generally faster than the R8000). However, because of relatively poor data cache reuse, on vectorizable codes the vector processors still maintain a significant advantage over the microprocessors. The R8000's large secondary data cache does provide an important benefit on codes whose working set size fits within it; however, on codes that exceed the size of the secondary cache, the IBM RS/6000 runs faster, because it has higher bandwidth to memory than the R8000.