Benchmarking The Experimental Ubuntu x86-64-v3 Build For Greater Performance On Modern CPUs

[phoronix]

Phoronix: Benchmarking The Experimental Ubuntu x86-64-v3 Build For Greater Performance On Modern CPUs

One of the exciting innovations currently being explored by Canonical ahead of the Ubuntu 24.04 LTS release is an x86-64-v3 build of the OS / packages. The x86-64-v3 micro-architecture feature level makes AVX/AVX2 support assumed by default as well as other modern x86_64 ISA features typically common of AMD and Intel processors the past number of years (with exceptions). Canonical's determination around what to do with a possible complementary Ubuntu x86-64-v3 build/archive is still being determined but they had released an experimental Ubuntu 23.04 based build that I decided to take for some benchmarking.

https://www.phoronix.com/review/ubuntu-x86-64-v3-benchmark

[Kjell]

CachyOS is solid

[skeevy420]

CachyOS is solid

I use it on bare metal and for WSL2.

[jeisom]

Definitely some performance being left behind by not upping the flags, at least in these benchmarks. I feel like a more thorough benchmarking analysis on a greater number of benchmarks would definitely be interesting.

Not sure V4 would show as much benefit as much as a lot of time avx512 has coded pathways to use the available instructions on the processor in software that would benefit most from it. Would be interesting to see if I am wrong about that though.

[coder]

Michael, for the sake of comparison, can you benchmark some of the OpenSuse packages built to support hwcaps?

[ms178]

Michael, for the sake of comparison, can you benchmark some of the OpenSuse packages built to support hwcaps?

To be a fair comparison the packages would need to be on the same version as Ubuntu 23.04. But I would be interested in a comparison as well.

I see a consistent trend in the numbers presented here by Michael by the way. Hence it might bring some benefits when doing a whole distro build with x86-64-v3 vs doing the hwcaps-only with some libraries approach.

[coder]

To be a fair comparison the packages would need to be on the same version as Ubuntu 23.04. But I would be interested in a comparison as well.

I think the same version of the package + same compiler version should be sufficient, in most cases. Also, CPU governor, as Volta mentioned (below). Same Kernel... if it's not much trouble.

Overall, I think Ubuntu outperforms OpenSuse (at least, the last time Michael compared them), so there could be some difference due to the kernel, but I'm trying to keep the request reasonable.

I see a consistent trend in the numbers presented here

It's a small sample that's not chosen at random. Consider that, before you draw any sweeping conclusions.

Hopefully, there will be enough overlap between the two test suites (i.e. v3-baseline vs. hwcaps) as to further inform the discussion.

[Volta]

To be a fair comparison the packages would need to be on the same version as Ubuntu 23.04. But I would be interested in a comparison as well.

They also have to use the same CPU governor and kernel.

[Schalefer]

Not sure V4 would show as much benefit as much as a lot of time avx512 has coded pathways to use the available instructions on the processor in software that would benefit most from it.

Same is true for v3 though. I would be surprised if e.g. SVT-AV1 is not using the AVX2 code path even on the non-v3 builds.

It would be interesting to see benchmarks where it is established that applications don't have and use a dedicated AVX2 path already.

[rene]

about as disappointing as to be expected from most just SIMD extensions for not vector math code general purpose control flow code, ...