New PGI Community Edition supports NVIDIA V100 Tensor Cores in CUDA Fortran, the full C++17 language, PCAST CPU/GPU auto-compare directives, OpenACC 2.6 and more.
PGI Compilers & Tools are for scientists and engineers developing high-performance computing (HPC) applications. PGI products deliver world-class multicore CPU performance, an easy on-ramp to GPU computing with OpenACC directives, and performance portability across all major HPC platforms.
New PGI Community Edition Features
- CUDA Fortran – 16-bit REAL(2) data type for V100 tensor core operations, optimized array assignment-based data movement
- OpenACC Auto-compare – Detect diverging results between CPU & GPU
- PGI on AWS – Develop, test, benchmark, deploy on V100s for as little as $3/hour
- Additional Fortran 2008 features – Block construct, g0 edit descriptor, multiple sourced allocation, vector norm2 and several other features
- C++ – Now interoperable with GNU releases through GCC 9.1
- LLVM 8.0 – The default LLVM back-end for Linux on x86-64 and OpenPOWER is updated from LLVM 7.0 to 8.0