Threadripper performance tweaks for pgi fortran?


I am using a threadripper 1950X to run my simulation code. It is compiled with pgf90 with optimization turned on.

Execution times appear to be on the order of 5 to 10% longer on the 1950X compared to an i7-4770 with the same pgi fortran installed with the same compiler options (basically just -O).

pgf90 -V gives the following output

pgf90 17.10-0 64-bit target on x86-64 Linux -tp piledriver
PGI Compilers and Tools
Copyright © 2017, NVIDIA CORPORATION. All rights reserved.

… which identifies the system as piledriver which is not accurate but may still be the best choice (?) for the current pgi compilers?

Since Ryzen CPUs have been out almost a year now … I was wondering if there is a better choice for the -tp option for using pgi fortran for compiling on AMD targets. I was also wondering if there are other options that would give better single thread performance.

Based on available benchmarking data the 1950X should be slightly ahead of the i7-4770 for single thread performance.

I see 18.1 has a new cpu type - zen for AMD epyc and ryzen.
So we are beginning to support the newer AMD chips.


Do you know when 18.1 (or a version with explicit Zen support) will be released? Only 17.10 is currently available for download.

The release 18.1 is imminent, and later than we wanted. Correctness and performance have to both get better. But we are getting there.