Arch Linux Developers Discuss Idea Of Providing An x86-64-v3 Port

[stiiixy]

Did you even read my damned post, or just chose to jump straight with a reply?

The issue is not about old hardware. It's about brand new hardware literally released less than two months ago that cannot boot and use a distribution optimized for x64-v3 because they lack AVX.

How will you like it if Arch suddenly decided that they will optimize the distribution for x64-v4 and cut out all consumer processors, limiting its use only to Xeons Golds and Epycs? You going to sink upwards of $7k to build a new machine capable of AVX-512 just to continue using Arch?

I'm sorry, did you not read what I wrote? I said keep the current, AND make a bleeding edge, so the rest of us can take advantage of it. Theoretically, there should be less maintenence on a build that focusses only on the modern.

[Sonadow]

Sorry, but people who bought CPUs without AVX support in 2020 made a bad buying decision. ISA support matters, the new feature levels show it and there were AMD alternatives on the market to these Pentiums which had AVX support.

Show me a fanless laptop or fanless mini PC running on AMD hardware.

Can't find it? Of course you can't; they simply don't exist. The fanless computing segment is entirely dominated by Atoms, Celerons and Pentiums. Who are you to tell people about what is a bad buying decision just because they make up a smaller segment of the market?

And your AVX-512 argument is absurd, as the market penetration of AVX-512 capable hardware is simply not there while AVX2 is widespread and common these days. Also as Smitty said, there is still the base variant for people who don't own capable hardware, so I don't really know what all the fuss is about. At least Arch is trying to accomodate to both users of legacy CPUs and capable hardware.

Playing the 'widespread and common' game, are we? Then I am perfectly justified to say that hardware vendors and OEMs have no need to accommodate Linux at all with any documentation or drivers because Windows is widespread and common but Linux isn't, therefore they should just simply not provide any drivers, source codes or documentation for Linux for all hardware.

[ms178]

Show me a fanless laptop or fanless mini PC running on AMD hardware.

Can't find it? Of course you can't; they simply don't exist. The fanless computing segment is entirely dominated by Atoms, Celerons and Pentiums. Who are you to tell people about what is a bad buying decision just because they make up a smaller segment of the market?

Look harder next time, here is such a fanless AMD system with AVX support which became available in October 2019. You are also talking about a niche of a niche now. These people still can use the base version of Arch, so what is your problem exactly? Missing out on all the free performance? Direct your anger at Intel for castrating these low-end CPUs instead of distributions which want to make use of that functionality or at yourself for not buying the AMD alternative.

Playing the 'widespread and common' game, are we? Then I am perfectly justified to say that hardware vendors and OEMs have no need to accommodate Linux at all with any documentation or drivers because Windows is widespread and common but Linux isn't, therefore they should just simply not provide any drivers, source codes or documentation for Linux for all hardware.

This is not a game, this is still the case as it comes down if it makes sense to put ressources into something which benefits only a few or in the example of a hypothetical v4-only distro build, who can use it at all. I hope you agree with me that it wouldn't make sense at all to make a v4-only build for a consumer distro as AVX-512 is only supported in a couple of processors today with next to no market share in the consumer segment. That might change once Alder Lake and Zen 4 and their successors penetrate the market. But today AVX2 is much more widespread and makes sense as a target for a performance-oriented consumer distro.

By the way, with a gaming market share of 1-2% on Steam Linux still doesn't get the attention of hardware and game companies the same way as Windows, at least when it comes down to gaming. Thankfully for Linux users Valve wants Linux to be a strategic alternative to Windows and keeps investing into it regardless of market share, we desktop users also benefit of the ongoing work from the hardware vendors for their HPC, enterprise and mobile users as Linux is much more important there but that work also serves us desktop users to some extent (e.g. GPU drivers, compilers, libraries).

[skeevy420]

As I have said (and a few others) the AVX instruction set extension may be 10 years old but there are brand new CPUs without AVX support. AMD seems to be phasing out all non-AVX capable CPUs in their inventory but Intel still treats AVX as a premium feature and will most likely continue to do so. Yes, the majority of these CPUs are Pentium/Celeron/A8/A10 based but a lot of companies are buying PCs with such CPUs inside (the place that I work at did buy quite a few non-AVX capable laptops recently). It just seems odd to me that big corporations want to shoot themselves in the leg like that, although we have yet to see anyone make AVX a minimum requirement (with the exception of Clear Linux).

Personally, I think it would be a better choice to set the minimum requirements to something that is common in all x86 CPUs from the last 8-10 years.

That's V2. Intel Nehalem from 2008. 12 years ago. All because of those non-AVX CPUs from Intel.

[TemplarGR]

[/B]

That's V2. Intel Nehalem from 2008. 12 years ago. All because of those non-AVX CPUs from Intel.

Yup! In my opinion, like i said before on this thread, the best suggestion is v2 as it was originally. It won't leave out that many people, it will still offer some gains for everyone, and it won't increase the workload/resources requirements that much for the Arch developers. It won't split the user base. Better bug reporting/fixing, etc.

If they want to offer a v3 alternative port of Arch, then it is their decision, but in my opinion if anyone really needs that they can make their own Arch-based distro focused on v3 or even v4 computers. There is not need for the official Arch distribution to split like that. Just increase it to v2 as a baseline, it covers far more hardware and is far more reasonable for 2021. v3 in my opinion should be considered in 4-5 years from now.

[skeevy420]

Yup! In my opinion, like i said before on this thread, the best suggestion is v2 as it was originally. It won't leave out that many people, it will still offer some gains for everyone, and it won't increase the workload/resources requirements that much for the Arch developers. It won't split the user base. Better bug reporting/fixing, etc.

If they want to offer a v3 alternative port of Arch, then it is their decision, but in my opinion if anyone really needs that they can make their own Arch-based distro focused on v3 or even v4 computers. There is not need for the official Arch distribution to split like that. Just increase it to v2 as a baseline, it covers far more hardware and is far more reasonable for 2021. v3 in my opinion should be considered in 4-5 years from now.

IMHO, Arch should offer 2 and 3 and leave it at that. The .03% of the world with a v4 system either custom compiles their own software making v4 moot or they're rich and are one of those people who buys a new system every 5 months and posts pictures of it on Reddit making v4 moot. I reckon most of us end-users at home are on v2 or v3.

I hate to say it, but if a person is on a v1 machine they really do need to upgrade. Even v2 is long in the tooth. I know, I did a v2 to v3 change recently. Intel Westmere, basically the v2 cutoff line, to AMD Zen 2, well into v3. I was one of those "My 10 year old system is good enough, I'll tell ya what" people; then I upgraded and 4K switched from seconds per frame to 30+ frames per second, same HDD, same OS install, same GPU. Was being the key word. I'm still gaming on 1080p. 1080p60 maxed out looks and plays better than 4K with compromised settings to reach between 30 and 60.

IMHO, v2 is only for people who are happy or don't need any better than 1080p 30-60 fps. No matter what GPU you pair it with, that's about as good as you'll get because you're extremely CPU limited. Can't even play games like Death Stranding due to the dreaded AVX requirement. It freaking sucked owning that game and not being able to play it because my Westmere wasn't up to par.

[TemplarGR]

IMHO, Arch should offer 2 and 3 and leave it at that. The .03% of the world with a v4 system either custom compiles their own software making v4 moot or they're rich and are one of those people who buys a new system every 5 months and posts pictures of it on Reddit making v4 moot. I reckon most of us end-users at home are on v2 or v3.

I hate to say it, but if a person is on a v1 machine they really do need to upgrade. Even v2 is long in the tooth. I know, I did a v2 to v3 change recently. Intel Westmere, basically the v2 cutoff line, to AMD Zen 2, well into v3. I was one of those "My 10 year old system is good enough, I'll tell ya what" people; then I upgraded and 4K switched from seconds per frame to 30+ frames per second, same HDD, same OS install, same GPU. Was being the key word. I'm still gaming on 1080p. 1080p60 maxed out looks and plays better than 4K with compromised settings to reach between 30 and 60.

IMHO, v2 is only for people who are happy or don't need any better than 1080p 30-60 fps. No matter what GPU you pair it with, that's about as good as you'll get because you're extremely CPU limited. Can't even play games like Death Stranding due to the dreaded AVX requirement. It freaking sucked owning that game and not being able to play it because my Westmere wasn't up to par.

All this is fine and dandy until you realize only a tiny fraction of pc owners play video games...

[Mat2]

Yup! In my opinion, like i said before on this thread, the best suggestion is v2 as it was originally. It won't leave out that many people, it will still offer some gains for everyone, and it won't increase the workload/resources requirements that much for the Arch developers. It won't split the user base. Better bug reporting/fixing, etc.

There are no or virtually no performance gains from compiling for v2 (see the benchmark). I think that maybe ~10% of Linux users do not have v2, it would be pity if they lost support.

[skeevy420]

All this is fine and dandy until you realize only a tiny fraction of pc owners play video games...

It's more than that. Everything is much, much more snappy and reactive. Compiles get done faster. Firefox doesn't seem to hiccup as much. When I upgraded from a 1080p TV to a 4K TV at then end of my old system's life, I noticed a bit of a sluggishness on the desktop afterwards, Windows and Linux. That's gone. My Westmere could just barely push a 4K desktop. Mind you that I had dual CPUs, x5687s, and 48GB of 1333mhz triple channel ram. It was arguably one of the best spec'd Westmere systems around.

It's hard to quantify it other than it feels so much nicer and smoother. Like when I went from a P2 to Athlon X64. To me, Westmere to Zen 2 is that much of a change. I didn't think it would be. I thought it would be better, but not this much better. It wasn't like when I went from Athlon X64 to Intel Q6600. Some games LOST FPS with that change due to it having slightly slower quad cores . Then I learned about the Q6600 FSB mod and then it was faster . With this change I lost cores and ram and it's still that much better, no modding necessary. 8c16t 48gb 1333 ecc to 6c12t 32gb 3600 no-ecc....though I did pick up an iGPU...compromises and all .

But 4K gaming going from literal seconds per frame to a playable frames per second, you just can't not notice that.

[Grinch]

that cannot boot and use a distribution optimized for x64-v3 because they lack AVX.

IIRC from when reading the mailing list, the reason they went with v3 instead of v2 was because the performance increase from upgrading to v2 was practically non-existent, thus making it rather pointless.

As an anectodal, this morning I compiled Firefox with -march=native on a AMD 3700x, and then benchmarked it against the official Arch Firefox package using 'Benchmark Web 3.0' which one of the Arch devs had used, the increase in performance between the version optimized for zen2 versus the official one (just optimized for plain x86_64) was ~3%.

Granted, this was just one web benchmark suite, results may differ in others, and maybe Clang/LLVM 11 which was used when compiling doesn't make very good use of the zen2 architecture when optimizing (but we're talking zen2 here, not zen3, it should be decent at least), I had expected something around 5-10% at least, but I guess it once again underlines how hard it is for compilers to make remotely efficient use of special registers, the official Arch binary only supports up to SSE2, while the zen2 architecture supports SSE4.2, AVX, AVX2, yet compiling directly for this instruction set only yielded ~3% better performance.

I know Clang12 is imminent, maybe it will have improved code generation for zen2 and provide a better result, also it would be interesting to try GCC, but it's unsupported by Firefox and probably requires patching in order to compile.