PGI Accelerator Compilers Enhanced for NVIDIA GPUs Based on New Fermi Architecture

The latest version of PGI Accelerator compilers provide full support for CUDA Fortran on the latest NVIDIA GPU platforms and add support for allocatable device arrays within Fortran modules along with several API enhancements. CUDA Fortran, co-defined by NVIDIA and The Portland Group, is an extended version of the Fortran 2003 programming language that gives software developers direct control over all aspects of GPU programming. The PGI 10.4 release also enhances support for the PGI Accelerator directives-based programming model on Fermi platforms. The PGI Accelerator directives make GPU-software development easily approachable by application domain experts. Rather than porting and parallelizing entire programs or functions for the GPU, the PGI Accelerator directives allow incremental porting and parallelization of individual compute-intensive loops and code segments using standard-compliant and portable Fortran or C.

The PGI 10.4 release adds several ease-of-use features, including the use of PGI Unified Binary technology to build one version of an application that will run on any CUDA-enabled GPU. With PGI 10.4 compilers, programmers can automatically generate code that works and is optimized for both a Tesla C1060 GPU or the new Tesla C2050 GPU. In addition, they can take advantage of new GPU features including faster double-precision arithmetic, larger and configurable fast shared memory, and increased number of cores. Support for new NVIDIA GPU platforms in PGI 10.4 extends across Linux, Windows and MacOS, and within Microsoft Visual Studio via PGI Visual Fortran.

“With PGI 10.4, HPC users can create highly optimized heterogeneous multi-core applications for the latest CPUs from Intel and AMD in combination with the latest generation of GPUs from NVIDIA,” said Douglas Miles, director The Portland Group. “Efficiently using all available host cores for certain parts of an application while accelerating other portions on GPUs is the key to squeezing maximum performance out of today’s GPU-enabled workstations and cluster nodes. With Fermi’s improvement in double-precision performance, we expect a big increase in the number and type of applications that benefit from GPU acceleration.”

“A large part of the success of Tesla GPUs in the HPC space can be attributed to the quality of the development tools from NVIDIA and its partners,” said Sanford Russell, general manager, GPU Computing at NVIDIA. “This announcement from PGI, building on the tools already in the market, is more evidence of the increasing momentum behind GPU computing in general and our CUDA architecture in particular.”

The Portland Group compilers and tools for Fermi GPUs are part of the PGI 2010 release version 10.4 and are available now.