PGI 17.4 Compiler Tests vs. GCC 6.3 vs. LLVM Clang 4.0
When NVIDIA-owned PGI released the PGI 17.4 compiler this week there was interest expressed by some Phoronix readers in seeing comparison benchmarks to GCC and Clang.
For those interested, today are some basic C/C++ benchmarks tested under PGI 17.4 free, LLVM Clang 4.0, and GCC 6.3. Keep in mind that the PGI compiler these days is primarily developed for acceleration on NVIDIA GPUs using OpenACC, OpenMP, and CUDA. PGI is primarily aimed for HPC applications, hence why NVIDIA acquired the company. In this article are just basic C/C++ benchmarks on an Intel CPU and not using any NVIDIA GPU acceleration in order to provide a straight-forward comparison with GCC and Clang. I'll look at some NVIDIA PGI GPGPU benchmarks soon.
Tests were done on the same Ubuntu 17.04 box with Intel Core i7 6800K Broadwell-E CPU.
With the very basic SciMark2 C benchmark, PGI 17.4 Community Edition was performing on par with GCC 6.3. LLVM Clang 4.0 is the fastest in this micro-benchmark/
PGI was coming up short in some of SciMark's sub-tests, but ahead of the GNU Compiler Collection in other tests.
MAFFT was one of the workloads performing abnormally poor on PGI.
With the OpenMP-based GraphicsMagick imaging application, PGI 17.4 was performing on par with LLVM Clang, which was behind GCC.
With the C-Ray multi-threaded ray-tracer, the PGI-generated binary was just behind Clang in terms of speed.
For those curious about the performance of PGI relative to Clang and GCC, hopefully these few benchmarks were of interest, but keep in mind the principal use-case for the PGI compiler is for offloading supported codes to NVIDIA GPUs. Those wishing to learn more about the PGI compiler stack can visit pgroup.com.
For those interested, today are some basic C/C++ benchmarks tested under PGI 17.4 free, LLVM Clang 4.0, and GCC 6.3. Keep in mind that the PGI compiler these days is primarily developed for acceleration on NVIDIA GPUs using OpenACC, OpenMP, and CUDA. PGI is primarily aimed for HPC applications, hence why NVIDIA acquired the company. In this article are just basic C/C++ benchmarks on an Intel CPU and not using any NVIDIA GPU acceleration in order to provide a straight-forward comparison with GCC and Clang. I'll look at some NVIDIA PGI GPGPU benchmarks soon.
Tests were done on the same Ubuntu 17.04 box with Intel Core i7 6800K Broadwell-E CPU.
With the very basic SciMark2 C benchmark, PGI 17.4 Community Edition was performing on par with GCC 6.3. LLVM Clang 4.0 is the fastest in this micro-benchmark/
PGI was coming up short in some of SciMark's sub-tests, but ahead of the GNU Compiler Collection in other tests.
MAFFT was one of the workloads performing abnormally poor on PGI.
With the OpenMP-based GraphicsMagick imaging application, PGI 17.4 was performing on par with LLVM Clang, which was behind GCC.
With the C-Ray multi-threaded ray-tracer, the PGI-generated binary was just behind Clang in terms of speed.
For those curious about the performance of PGI relative to Clang and GCC, hopefully these few benchmarks were of interest, but keep in mind the principal use-case for the PGI compiler is for offloading supported codes to NVIDIA GPUs. Those wishing to learn more about the PGI compiler stack can visit pgroup.com.
2 Comments