AMD AOCC 3.2 vs. GCC vs. Clang Compiler Performance On Zen 3

[phoronix]

Phoronix: AMD AOCC 3.2 vs. GCC vs. Clang Compiler Performance On Zen 3

Earlier this month AMD released AOCC 3.2 as the newest version of their LLVM/Clang-based compiler focused on delivering optimized Zen performance. With our initial AMD AOCC 3.2 benchmarks on Zen 3, there is nice incremental improvement compared to the prior 3.x releases. But how does this AMD-optimized compiler stack up against the upstream LLVM Clang and GCC compilers? Here is a look at the AMD AOCC performance against the current Clang and GCC C/C++ compilers.

https://www.phoronix.com/vr.php?view=30814

[skeevy420]

I have bad luck with AOCC. The two things I've ever tried to compile with AOCC didn't compile when Clang/LLVM worked. Godot 4.0 with 3.2 sometime the day after Christmas and...fsck me running, I don't remember what the other program was and now that's bugging the heck out of me...

[Danny3]

Can't AMD contribute to GCC development too ?

[jason.oliveira]

Hopefully, AOCC is another one of those "Let's see if we can even compete with open source projects using our own talent" experiments that leads to greater contribution to whatever free softare project they choose to lose to this time, and kills any remaining NIH Syndrome. If AMD were smart, they'd accept that they've _never_ had a top-tier software development team anywhere in the company, and begin contributing more to open projects, which would take a load off whoever was paid to fork LLVM/Clang into this monstrosity.

We get that software is eating the world, and AMD needs to have a semi-proprietary software suite they can market to vendors and enterprise, but the pure hubris of the "We're better programmers than the FOSS pool" keeps leading them to these footgun decisions that really cater to nobody but their own wounded egos.

[Michael]

Can't AMD contribute to GCC development too ?

They generally have SUSE do it for them.

[agd5f]

Hopefully, AOCC is another one of those "Let's see if we can even compete with open source projects using our own talent" experiments that leads to greater contribution to whatever free softare project they choose to lose to this time, and kills any remaining NIH Syndrome. If AMD were smart, they'd accept that they've _never_ had a top-tier software development team anywhere in the company, and begin contributing more to open projects, which would take a load off whoever was paid to fork LLVM/Clang into this monstrosity.

We get that software is eating the world, and AMD needs to have a semi-proprietary software suite they can market to vendors and enterprise, but the pure hubris of the "We're better programmers than the FOSS pool" keeps leading them to these footgun decisions that really cater to nobody but their own wounded egos.

AMD does contribute to other open source compilers. A lot of these optimizations do eventually make it into LLVM, etc. In most cases it's the same developers.

[onlyLinuxLuvUBack]

if you can get a gigabyte board and see if the supermicro platform is crippled for amd. supermicro clowns only discuss buying intel and they really tried to push intel and say wait for intel. Gigabyte vs supermicro on zen3 vs aocc/gcc ?

[pracedru]

Hi Michael,

Any chance you could add the Intel compiler to your tests?

br Magnus

[brucethemoose]

Hi Michael,

Any chance you could add the Intel compiler to your tests?

br Magnus

+1. Intel claims their LLVM-based compiler have some pretty significant performance gains: https://www.intel.com/content/www/us...plete-icx.html

[PerformanceExpert]

Hopefully, AOCC is another one of those "Let's see if we can even compete with open source projects using our own talent" experiments that leads to greater contribution to whatever free softare project they choose to lose to this time, and kills any remaining NIH Syndrome. If AMD were smart, they'd accept that they've _never_ had a top-tier software development team anywhere in the company, and begin contributing more to open projects, which would take a load off whoever was paid to fork LLVM/Clang into this monstrosity.

We get that software is eating the world, and AMD needs to have a semi-proprietary software suite they can market to vendors and enterprise, but the pure hubris of the "We're better programmers than the FOSS pool" keeps leading them to these footgun decisions that really cater to nobody but their own wounded egos.

I hate to break it to you, but AOCC, like ICC, exists only to get the best SPEC results. Nowadays both are based on standard LLVM with extra optimization passes. Why? Well, on typical (badly written spaghetti) code it is hard to beat GCC or LLVM. The extra optimizations are so specific to SPEC they are practically useless for anything else. Libquantum remains the best example of how insane this benchmark trickery got: with speedups of several thousand times, SPEC2006 scores became meaningless. Some of the tech press gave ICC scores a 40% penalty for all this trickery, others like AnandTech exclusively use GCC and LLVM for fair benchmark comparisons.

The reason you will never see these passes contributed to open source is that they are not general optimizations. The open source community has a strong dislike for benchmark tricks in general, and the nasty tricks done for SPEC are just too much to be acceptable. Also both Intel and AMD believe they can out-trick each other and so get an advantage by keeping their optimizations closed source (there were lawsuits over the dirty tricks Intel played to ensure people could not use ICC on AMD CPUs).

Other companies decided not to play these deceptive games and just make big/wide CPUs which are undeniably fast running real applications...