Tom Henderson, Mark Govett, Jacques Middlecoff, Paul Madden, Jim Rosinski, Craig Tierney
As the clock frequencies of Central Processing Units (CPU) stall, chip designers are turning to fine-grained parallelism to increase performance. Graphics Processing Units (GPU) already provide massive fine-grained parallelism and have recently been generalized to support scientific computation. Like most emerging technologies, GPUs are not yet easy to use. Significant costs must be incurred to re-code existing software. New compilers from several vendors now target GPUs and promise to reduce these costs. We report on our experience using commercial and open-source compilers to port a prototype NWP dynamical core to NVIDIA GPUs.
Tom Henderson has spent most of the past 20 years providing high performance computing and general software engineering support to scientists developing atmospheric and oceanic prediction models. Beginning in 1991 he worked on a team that transitioned NOAA's Forecast Systems Laboratory (now Global Systems Division) from VAX VMS to Intel Paragon MPP. He served on the MPI Forum during creation of the MPI-1 standard and contributed to CCSM, WRF, RUC, FIM, and several other models while working at NOAA and NCAR. Recent interests include application of computational accelerators (GPU and soon MIC) to NWP codes.
| Attachment | Size |
|---|---|
 Henderson_GPUFortranCompilers_SEA12.ppt | 1.05 MB |
