Two nits: i686 was introduced in the Pentium Pro; and, you do find a lot of Intel Atom processors in embedded products like NAS devices, network firewalls, control planes for switches. While some of those are moving to Xeon D or newer Atom processors with Gracemont or newer Atom cores, requiring v3 will be a tougher sell for those markets than for the consumer markets.

4:11 - Main content starts here

Thanks for this well done and informative video.

9:00 - Exactly - one of Linux call to fame has always been suport for ancient hardware, giving second or third life to what would otherwise be hard to recycle waste.

There is another reaason v3 support as requirement will get harder: Most office Linux users won't actually performance wise require to update their CPUs beyond a pre v3 point, as the demand on the hardware hasn't increased much outside of gaming in recent years. I can certainly name a few in my surroundings for who this is true. For example, my mother's PC is still running perfectly fine with my old first gen quad core Core i7 860 & 8 GB of RAM. As long as you have a SSD as boot volume, even Windows isn't really an issue performance wise. Yes, gaming is a whole different story, but as long as you don't do that, anything a desktop will do for you, will still run great. Servers will update, I'm sure, and occasionally failed hardware will be replaced new, but nothing like the rate that you had in prior generations with real requirement increases demanding hardware updates. And if they are upgrading, at least for some, x86 won't be the target architecture. Already true for som e Chromebooks and all new Macs, as well as Qualcomm based Windows on ARM devices, which will get a lot better when the new Snapdragon X Elite hits shelves. Also any Neoverse based servers, and many SBC's & NAS systems that might run Linux on ARM instead as well.

AVX-5 1 12 never heard of that instruction set

Thank you for this great video.

I came about this a long while back when I was trying to download an update for mpv-player and there was an x86 64 V3 version, but there was a post explaining what it was so it was all good.

Are there documents that accurately, completely, and, extremely importantly, concisely/tersely map and define baselines, extensions, instructions, and their CPUs? If so, which is the best? Thanks in advance!

If I remember correctly the 80386 used a memory controller that was located in the Northbridge, which at the time was a separate chip from the CPU. Today the Northbridge is part of the CPU, so we really don't mention it much anymore.

_"Intel does not officially support AVX-512 family of instructions on the Alder Lake microprocessors. Intel has disabled in silicon (fused off) AVX-512 on recent steppings of Alder Lake microprocessors to prevent customers from enabling AVX-512. In older Alder Lake family CPUs with some legacy combinations of BIOS and microcode revisions, it was possible to execute AVX-512 family instructions when disabling all the efficiency cores which do not contain the silicon for AVX-512"_ does that mean v4 will never be a "standard" like v3?

Thank you for the video Gary, very helpful!

Recently I "culled" my old, unused, Athlon-64 desktop. Now I see that this was a better decision that I had thought at the time! The new one is v4 😁

So basically, it's a marketing thing. That's why big companies designed they own computers and OSs.

Interesting. If I have a new CPU that supports x86_64_v4, does that mean my system will be faster if I compile the kernel myself rather than using the one distributed by my distro? I have never compiled the kernel before but it can't be that hard, might be worthwhile if there is an uplift in performance!

CachyOS (Arch bases) os v3 and v4 --- I use v3 with my AMD Ryzen 9 5900X, and anything is running really great.

Funny to see in the list "Gentoo is now offering x86-64-v3 packages". I assume it's on the binary packages, because normally you compile the packages, and it's trivial (and by far the easiest when compared to the other distributions) to compile it to use everything your CPU is capable of.

I'm confused, people are moving to ARM because it's RISC (reduced instruction set)... and yet they keep adding instructions to x86.

Thanks for a fascinating upload. Considering all target CPU instruction sets (including V1-V4 extensions etc) are selected via compiler configuration, the only limiting factors to this are... 1) The capability of the CPU running the code. 2) The extension-awareness / capability of the compiler. With that, the real question is why do devs take so long to release versions targeting these extensions - if (pretty much) all that is required is a 'simple' compiler swich? Certainly, it's no easy task to develop a compiler that knows how to produce machine code that always selects the best optimisation when particular destination flags are selected. Over the last 30-40 years, we've put more and more trust in the compiler, while losing, more and more, the best speed/size optimisation tool possible - a human assembly language expert. Want to 100% guarantee your core/critical routines are speed optimized - become an expert in assembly. Sadly, total compiler faith is too embedded today.

Want to know about the most hacky x86-64 binary format? It's x32 (look up x32 ABI) which is normal x86-64, but uses 32-bit pointers, this for cases where speed is king, but comparability is not.

Intel Sandy Bridge introduced the 1st AVX extension.

Intel really held back v3 adoption. Also if anyone is wondering about small cores and meeting v2 requirement, Silvermont and up on the Atom lineup and Jaguar and up on AMD’s cat core lineup meet v2 requirements. As for how other OSes do it, macOS supported SSSE3 (not to be confused with SSE3 which it also supports because it is older than SSSE3) due to when they started transition to Intel. Since Sierra, they require Penryn so you can support up to SSE 4.1 by default. In addition, since macOS can support multiple architectures, Apple created x86-64h slice which meets x86-64-v3 well before the microarchitecture levels were a thing so you can optimize for both the baseline and Haswell out of the box. And Apple didn’t choose Pentium or Celeron or Atom so it can go back to 2014ish Macs while still supporting older with the normal x86-64 in the same binary. For Windows, it required SSE2 since Windows 8 for 32-bits and retroactively for Windows 7 since 2018. For 64-bit, Windows 8.1 required CMPXCHG16B (basically allowing atomics on 128-bits/16-bytes memory), LAHF and SAHF, and PrefetchW support (which is just to help the CPU prefetch relevant data. And your CPU doesn’t even have to support the prefetch, as Intel treated it as NOP (No Operation) until Broadwell). And so this excluded the earliest x64 CPUs only. Coincidently, these CPUs also support SSE3. v2 level requires SSE4.2, POPCNT and other instructions. Windows 11 finally requires processors with instructions beyond v2.

Am I the only one who gets frustrated by the use of Intel CPU codenames? Everyone seems to do that when talking about instruction sets and such, and I always have to go look up to see which processor it actually is, to be able to have a point of comparison with CPUs I actually know and/or own. I would love it if people would at least at one point say that Haswell means "4th generation Core" or the 4000 series. That way I'd be able to instantly compare it with, say, my new-to-me 5300u laptop. (Got it for $25, so very much worth it.) I much prefer the generation names because you can instantly compare them without having to memorize all the names and their order. It's not like Android or Apple who would at least use letters in alphabetical order.

I was under the impression AVX-512 is being heavily reworked for big little. So what happened to x86-64-v4.

My Gentoo system is optimized for my exact microarchitecture 😉

Why can't compilers build combo binaries which use the new instructions if they're there, and the old way of they're not? Sure, the binaries would be bigger, but space isn't really a big issue these days.

Windows 11 is going to actually raise the CPU requirements to x86-64-v2 soon in the next major release according to the latest builds, the existing official Windows 11 minimum requirements are actually on paper and refer more to technical support, since Windows 11 in practice It still able to run on x86-64-v1 CPUs, but is finally being recompiled to only support instructions that meet the x86-64-v2 criteria. I suspect this is just the beginning and may in a few years Windows will only support x86-64-v3, with x86-64-v4 as recommended due to AI features.

Watching this on a v2 Maybe I should upgrade this one of these years... I'm on ivy bridge too so I'm only *just* short of haswell

I'm still riding on an AMD FX-8370 with Windows 11 patched to bypass all requirements. Even without UEFI and legacy only booting, it's surprisingly useful and fast. When I first got the FX series, it was crap, slow and worthless, but something happened over the past decade, it somehow got faster in software updates, and with newer games actually using multiple cores, and now it can play a lot of the modern game libraries (not all, and not all run "great", some run "well"), but alas, as a day to day driver, it still boots in 5 seconds, it still plays UKposts in 4k60, it still handles everything I simply throw at it, except software that might require a newer AVX than I have. But even then, I haven't felt the need to update it yet. Ironically enough, the only v3 CPU/computer I have would be my ThinkPad, but it's GPU's no longer get driver updates, meaning it's essentially more useless than my v2 desktop since my desktop can still have a modern GPU plugged in and boot with it, whereas my ThinkPad wouldn't be portable anymore if plugging in a modern GPU over Thunderbolt (and not even supported properly, it's a W541, so that's kinda a hack which cannot even use the internal LCD). If I am forced to update the desktop, I'll have no choice, but I will resist as long as I can. Not out of spite, but more so because I don't feel I should pay money to buy a new computer when again, it still can boot up in mere seconds, and be ready to be used, and actually run circles around my ThinkPad which IS a v3. I don't spend money when I don't need to, so if software reaches a point that my CPU is no longer a target, I'll just hold onto the OS until the SSL certificates expire and I simply have no other choice. Some of us just can't afford the insane prices, it's still a reason my GPU is a 1080 Ti, and not even an RTX. Can't afford it, don't want to, because I have 11GB of VRAM, and most games are quite comfortable with this. Alan Wake 2? Yea vkd3d actually helped enough to let the game be playable at least, but I didn't care for the game so I never continued to play. Star Field? What a boring game "to me" (stressing this is only my opinion and not being presented as a fact), so I didn't care that my CPU was too slow for it. The games I do play, run at or exceed 60fps, so what matters most to me is that I am happy with the machine, thus I refuse to pay up for an upgrade I feel I don't need. I also understand at a point, it's not feasible to target this machine anymore, but as far as I'm concerned, that's not my problem until it actually is my problem, and right now my v2 is not actively refusing to run any software I have thrown at it, except the rare few that weren't even important enough to actually care at all. Linux isn't my OS of choice (it was for 2 decades but I stopped), and Windows is not actively targeting V3 yet (except maybe the Rust based kernel, which I await to try and see, I have the ISO to it, and VirtualBox ready, which using AMDv should tell me if it works or not without actually installing it).

Surprised that the Intel Atom processors didn't support V3 until last year. IE, Intel N100 or N95.

I may be wrong but Ryzen 7000 has v4 support. They do support AVX-512. And I think they are the only CPUs that do for now.

I think that developers should always try to target the lowest required hardware. For example there is no real need for libc to require AVX512. We should try to keep support for things for as long as they are useful, so long as they don't stagnate forward progress. Luckily UNIX oses do a good job of this as one thing these oses do is breathe a new lease of life into older hardware.

I wonder how hard it would be to _reliably_ ship only the source code, check your CPU on install and then compile (or recompile) what is needed for your system. So if your CPU does not support v4, the kernel or parts of it would be compiled right on install. But I'm sure if it was that easy, it would have already been done...

..is there a chance of 128-bit era...while most BIOS are still 16/32-bit apart newer UEFI

Great Video. Can you please do a video comparing S24 Ultra's Gen 3 NPU vs the Google 8 Pros Tensors NPU ? Which is actually the better NPU ?

main subject at 7:20

"Baseline x86-64 included MMX, SSE, SSE2". I believe that the earliest AMD64 processors did not have SSE2, and that's why 64-bit Windows hasn't supported it since .. what was it, Windows 7? Or was it Windows 8, or 10? I have one old machine that must use 32-bit Windows to run at all.

One minor point, is that it is trademark, not copyright that was the relevant factor for why Intel went for Pentium branding rather than using a number. You can't copyright a single word either. Copyright and trademark might sound very similar, but the former will run out in time and requires substantial creative content. However, trademark can last forever, provided that they are renewed. Trademarks also only protect things within a trading/commercial environment. Intel could sue me if I infringed their trademark by, say, selling a computer related product using the name Pentium, but they cannot sue me, or anybody else, for using the word Pentium in other contexts. The difference can be important. Disney try and protect their control over Mickey Mouse by trademarking the designs related to the character, not just by copyright. Indeed the first ever film in which Mickey Mouse appeared has just dropped out of copyright.

Some of these extensions seem to be ... really specialized; not exactly things you'd be doing in the kernel itself. There are probably some features that a kernel might care about (VM extensions), but outside of detecting features and managing extra context and resource allocation (i.e. treat like a device) in case there's a limited set of new resources associated with the extensions, the rest seemingly should be kicked to userspace. It just feels a little weird to totally break a distro for, what, some extra vector SIMD instructions. We've seen this with other architectures with their alphabet of feature sets and extensions.

Shortly after the release of the 8-bit Nintendo Entertainment System, they already had 32 bit processors? Why didn't Nintendo contract with Intel?

For Windows 11 and AMD CPU's, the requirement from Microsoft is not the Zen version, but rather the model number series. The requirement is AMD 3000 series or higher. The Ryzen 3600 is Zen 2, but the Ryzen 3400G is Zen+. Both will meet the requirement. AMD's naming scheme is stupider than Intel's naming scheme.

Hmmm so buying used ThinkPad with coreboot with SSD and Ram upgrade seen unwise now especially if you use fedora or other distros mandating new x86 revisions

Might as well just make a replacement for x86 at this point, it's becoming a compatibility nightmare. Like a x86-64 without all the instructions that aren't really needed anymore, and just revise it every few years whilst keeping compatibility in mind.

*Suse:* "Suse-er" *"Can't copyright numbers":* There was a court case between Intel and AMD over AMD's licence to produce x86 processors. AMD claimed they could use _any_ instructions from any Intel processor with an x86 numerical name, so to save on court costs Intel responded by changing from numbers to alpha names ("Pentium" etc).

The first 32 bit CPU by Intel was the Intel iAPX 432 . It was released in 1981. One of the three big fails of Intel. Extremely complex and extremely slow.

At this point v2 should cause no problems for any enterprise linux distro (Windows has required v2 level since 8.1). Desktop distros might have to contend with people intentionally using modern linux on obsolete hardware

It's kind of uncommon & confusing in calling these microarchitectures, because in the recent processor architecture world the word microarchitecture specifically already means something different. Instead it's more appropriate & common to either call these architecture extensions or subarchitectures.

Debian still offers i386 builds as of today, which is kind of stupid as those are in fact i686 builds, as they won't even run on i586 CPUs. Other distributions at least name their builds correctly.

Can you make video on New harmony os announced by huawei vs Android vs ios

Just publically registered by disdain for this trend. This is TOTALLY opposite of from among the reasons Linux came to be to begin with. Supporting older hardware was one of the thing Linux was famous for.

No, 32 bits was not given up ... every single intel CPU yet can run a 32 bit OS, the backward compatibility is still there, all the way to mov al,2 in 8 bit. That is the beauty and very often advantage of x86, it may change in a near future.

For me the fact that my old PC will be *immune to Windows 11* is great news, I was afraid that any day in the morning it would appear updated to an OS that can't do more than Windows 10 but that has a dumb (or is it dumb oriented?) user interface that is reaching the same level of functionality as the MS-DOS Shell "GUI"

x86 as a CISC unit is kinda bloated too much already, with huge die area covered by almost never used instruction handlers. And now we get new ISA with even more bloat? Does it worth it? I wonder if extended instruction sets actually give any noticeable performance over old CISC or RISC architectures. Depends on application and compiler i guess, but as of now, LLVM Clang has very weak, almost negligible optimisation for latest x86 ISAs, same with gcc iirc. To me personally, it seems as just a pure marketing and potentional threat of hardware bitlocking, like it’s already happening with new TPM and Secure Boot modules, which obviously are created to try and let microsoft monopolize OS market and keep you from your right to own your devices(like it’s already happened with your phones).

Can't copywrite names. Pentium isnt Copywrite it is Trademarked :)

Debbie Anne

AVX was introduced in Sandy Bridge, not Haswell.

It is a chip architecture which is on its way out which is inefficient and Bloated with so many hidden spyware instructions from the likes of Intel and even AMD.

step to 128

It's actually good if Windows 11 requires newer hardware. If anything they are much too slow with requiring more modern CPUs. It took them ages to dish 32 bit for example. And with all the security flaws existing in older gen processes, it's better to have new OS require new processors.

Pentium Pro*

/ʊˈbʊntuː/

I think debbie-anne linux is a great idea to get more women into Linux😂

It is AMD that produced the x86-64 architecture (and saved us all from the train wreck Itanium), not Intel. @2:54, you wrote "for Intel EM64T compatibility", it is the reverse: it is Intel who is compatible with AMD64.

First

X86 is the dying CISC Instruction Set Architecture from Intel. Nothing to cry about. Just obsolete. It goes the same way, as any technology. From birth to death.

/ˈdʒɛntuː/