cancel
Showing results for 
Search instead for 
Did you mean: 

PC Drivers & Software

gcc and -fopenmp on EPYC 9654 and EPYC 9754

We did a benchmark of our software on the mentioned processor. There i noted that compiling with different versions of gcc resulted in different openmp performance (numbers given are min in time to complete the code):

  AMD (128 cores) 
 gfortran 105.8 
 gfortran 118.7 
 gfortran 1211.3 
 gfortran 1312.5 

 

The strange thing is that gfortran 10 is performing better than the later versions.

Is there a problem with recognizing the CPU? The code does not support GPU.

Can i add an additional flag in the later gfortran versions to improve the performance in the later gfortran versions?

In terms of scaleup the gfortran 12 and 10 are similar on AMD, in terms of speedup there is a big difference:

openmpi-amd-gfortran.png

 

0 Likes
1 Solution

If no answer please open this same thread that deals with AMD EPYC Processors and it hardware and software: https://community.amd.com/t5/server-gurus/ct-p/amd-server-gurus

View solution in original post

3 Replies

If no answer please open this same thread that deals with AMD EPYC Processors and it hardware and software: https://community.amd.com/t5/server-gurus/ct-p/amd-server-gurus

Apparently, adding the additional flags:
-march=native -mtune=native
fixes the problem (tested gfortran 12). But shouln't this be the default to use for gfortran?