An Early Look At The GCC 9.0 Performance On AMD EPYC

[phoronix]

Phoronix: An Early Look At The GCC 9.0 Performance On AMD EPYC

While GCC 9 has just been under development for a relatively short period of time, here are our initial benchmarks of GCC 9.0 SVN on and AMD EPYC server compared to the GCC 8.2 stable release candidate when tested at various optimization levels as well as PGO (Profile Guided Optimizations).

http://www.phoronix.com/vr.php?view=26617

[audir8]

Good to see PGO make meaningful differences in stockfish and himeno. Also good to see -march=native make a difference for Epyc/Ryzen for math intensive libraries. Still waiting to switch to 8.1 from Gentoo though.

[cb88]

Good to see PGO make meaningful differences in stockfish and himeno. Also good to see -march=native make a difference for Epyc/Ryzen for math intensive libraries. Still waiting to switch to 8.1 from Gentoo though.

You misread the benchmark graph... it made both of those WORSE than the others.

[audir8]

You misread the benchmark graph... it made both of those WORSE than the others.

With Stockfish, 8.2 RC + PGO does better than anything else. With Himeno, PGO does worse on 8.2 RC and 9. Point taken though, I was more focused on the native flag performance, and in scimark, lame and others, there's a healthy jump in performance from -O3 and -O2.

[coder]

I wonder how well-tuned these were, to begin with. PGO should be a huge win for things like loop unrolling, software pipelining, and vectorization. Only if the programs have already been hand-optimized (which I'm sure some of these are), should it fail to deliver a significant benefit (i.e. on vectorizable, pipelineable code).