RHEL9 Likely To Drop Older x86_64 CPUs, Fedora Can Better Prepare With "Enterprise Linux Next"

[phoronix]

Phoronix: RHEL9 Likely To Drop Older x86_64 CPUs, Fedora Can Better Prepare With "Enterprise Linux Next"

Red Hat Enterprise Linux 9 will likely see support for older x86_64 CPUs eliminated to focus on more modern x86_64 Intel/AMD families. With that, Red Hat developers working on Fedora have been working on an "Enterprise Linux Next" proposal to not only vet such x86_64 build changes but also to provide a feedback workflow for other changes...

http://www.phoronix.com/scan.php?page=news_item&px=RHEL9-Drop-Old-CPUs-Fedora-ELN

[edwaleni]

VMWare has already sunsetted anything prior to 2010 (Westmere) due to missing instructions for virtualization. I wonder how much virtio stuff for KVM can be taken out for anything less than IvyBridge or Opteron Bulldozer.

While AVX2 is a pretty high ceiling to make the cutoff at, I am sure it will come down to something like VMX Unrestricted Guest.

[microcode]

I wonder how expensive it would be to host two packages for everything: AVX2 and later for mainstream, and the generic x86-64 target for everyone else.

[Michael]

I wonder how expensive it would be to host two packages for everything: AVX2 and later for mainstream, and the generic x86-64 target for everyone else.

The actual hosting/bandwidth is likely not the issue, more pressing is likely the QA and related costs involved.

[birdie]

Enabling the AVX2 compilation flag for the code which was not written to take advantage of this extension won't automatically increase performance. Just saying.

[TemplarGR]

I think AVX2 is too high of a threshold for now. AMD only made AVX2 cpus after 2017*, and Intel after 2013. Even for a distro released in 2023, i think that's a little too high. I think the first AVX version would be a better compromise, since it goes back a couple of years for both companies (2011 for both Intel and AMD), but is still far more advanced than SS2, (S)SS3 and SSE4.(1,2) cpus.

I would personally make it AVX1 and upwords for version 9 and push it to AVX2 minimum for the next version.

*Technically Excavator was first in 2015 but IIRC Excavator was not a desktop/server chip but was only found in apus, or wasn't it?

[microcode]

Enabling the AVX2 compilation flag for the code which was not written to take advantage of this extension won't automatically increase performance. Just saying.

This is actually sometimes untrue. Autovectorization has come a long way in compilers.

[NotMine999]

I guess changes like this are inevitable.

Is there any online repository out there that archives older Linux releases that get "tombstoned" for whatever reason?

[angrypie]

*Technically Excavator was first in 2015 but IIRC Excavator was not a desktop/server chip but was only found in apus, or wasn't it?

Carrizo (Athlon X4 845 and A8-7680) and Bristol Ridge were released for the desktop, the latter being the first AMD CPU to use the AM4 socket, but I don't think AMD had a huge inventory, and nobody would buy those anyway since 1st gen Ryzen was impending (it was mainly for shitty prebuilts). It's almost impossible to find these CPUs nowadays except for the A8 and, again, nobody has any business buying those when a 2200G APU is so cheap right now. People still on FM2+ might be interested in the A8, though.

Of course setting a baseline on AVX2 is a stupid idea when you consider the aggressive marketing segmentation Intel does with ISA extensions. Even recent low-end CPUs have AVX2 lasered off. That would effectively kill Fedora.

[hajj_3]

Neither AVX nor AVX2 are available on intel celeron or pentium processors. Requiring AVX2 is therefore a terrible idea. Requiring SSE4 or SSE4.1 would be a better idea.