ARM vs x86

[WhoDecidedThat]

Found a nice article on it. Wanted to share it.

ARM or x86? ISA Doesn’t Matter

For the past decade, ARM CPU makers have made repeated attempts to break into the high performance CPU market so it’s no surprise that we’ve seen plenty of articles, videos and discussions about ARM’s effort, and many of these pieces focus on differences between the two instruction set...

chipsandcheese.com

 

[R-T-B]

Anyone who really understands how a CPU works has been saying this forever.

ISA does not matter. Architecture does.

tl;dr version of the highlight:

In other words, this very popular Techquickie video is misleading, and the ARM ISA doesn’t have anything to do with low power. Similarly, the x86 ISA has nothing to do with high performance. The ARM based CPUs we’re familiar with today happen to be low power, because makers of ARM CPUs target their designs toward cell phones and tablets. Intel and AMD’s x86 CPUs target higher performance, which comes with higher power.

 

[WhoDecidedThat]

Anyone who really understands how a CPU works has been saying this forever.

ISA does not matter. Architecture does.

tl;dr version of the highlight:

You can't blame people for looking at the correlation between good battery life and ARM devices (smartphones, tablets and the new Macs) and thinking ARM is the cause behind good battery life. Intel has failed to deliver comparable battery life. AMD has somewhat succeeded.

 

[R-T-B]

You can't blame people for looking at the cor-relation between good battery life and ARM devices (smartphones, tablets and the new Macs) and thinking ARM is the cause behind good battery life. Intel has failed to deliver comparable battery life. AMD has somewhat succeeded.

No, I can't, but there were people who've been touting the "x86/CISC is bloated, throw it out" tagline since even before the PowerPC G5 / IBM 970. And that thing was NOT energy efficient vs anything. Even the Pentium 4.

It happened again with Sony and it's "supercomputer" cell processor in the PS3. Marketing pushed that hard.

None of those had anything to do with the ISA. And yet people thought it did.

 

[Kanan]

Anyone who really understands how a CPU works has been saying this forever.

ISA does not matter. Architecture does.

tl;dr version of the highlight:

ARM is a simpler ISA compared to X86, so while he isn't entirely wrong, he isn't entirely right either. It makes a lot of sense that ARM is associated with low power/high efficiency because these are the usual strengths of most ARM architectures. Yes, the ISA can do more, but then again, not as much as X86.

I want to add, that I would totally welcome ARM to challenge the X86 status quo and also Linux to challenge Windows so we have a more open and variable market, but probably wishful thinking.

Edited for clarification.

 

[R-T-B]

ARM is a simple architecture compared to X86, so while he isn't entirely wrong, he isn't entirely right either. It makes a lot of sense that ARM is associated with low power/high efficiency because these are the usual strengths of most ARM architectures. Yes, the ISA can do more, but then again, not as much as X86.

Read the article completely. Implementing these ISAs (not architectures) is taking an academically small level of die space. This might have been true in the 90s, but it just does not matter anymore.

The ISA can absolutely do as much as x86 were someone to make such a cpu... heck, some pretty much have.

ISA is a fun buzzword but it means nothing today. It's like arguing Spanish is "more effective" at getting the point across than say, any big language. It's silly.

 

[Kanan]

Read the article completely. Implementing these ISAs (not architectures) is taking an academically small level of die space. This might have been true in the 90s, but it just does not matter anymore.

The difference is, X86 minimum architecture level today involves AVX2 and more, with ARM the minimum level is much much simpler. Of course you can build a complex and powerful arch based on ARM ISA as proven by Fujitsu and later Apple but in general X86 is the more complex ISA. It has a reason ARM is more efficient, X86 is a bit too big, maybe a bit outdated, although I would say Ryzen is very modern, while Intel always reminds me why Apple has chosen to ditch X86.

 

[R-T-B]

The difference is, X86 minimum architecture level today involves AVX2 and more, with ARM the minimum level is much much simpler.

You are talking extensions, technically, of which there exist several for both architectures. It is not a core ISA concern.

X86 has AVX/AVX2 (and not on all CPUs), ARM has Scalar Vector Extensions, not on all cpus either of course. Really the same tech with different names. Only difference is virtually very modern x86 CPU is a high performance design vs what ARM manufacturers target, but that's not neccesarily how it has to be. It's a choice.

 

[Kanan]

You are talking extensions, technically, of which there exist several for both architectures. It is not a core ISA concern.

X86 has AVX/AVX2 (and not on all CPUs), ARM has Scalar Vector Extensions, not on all cpus either of course. Really the same tech with different names. Only difference is virtually very modern x86 CPU is a high performance design vs what ARM manufacturers target, but that's not neccesarily how it has to be. It's a choice.

Like I said I'm not saying the article is wrong. I'm saying it is partially right. X86 minimum spec is way higher than ARM, not even comparable. This leads to the conclusion that many people have, that ARM is a low power ISA/arch, not for high performance. This is 98% right to be fair, and this is the only thing I wanted to clarify basically.

If you take ARM ISA and stretch it to the point of rivaling x86, something which has not happened at this point, yes, it's theoretically the same. Has a reason why x86 reigns supreme though.

Apple can try and replace their WS CPU (Mac Pro) with ARM as well, it will be a big challenge for them to make ARM competitive with EPYC, Xeon is a joke at this point, I expect them to make it that far, if they try it at all. At this point, Apple has grown so powerful, I honestly expect them to be able to do it, if they want to.

 

[WhoDecidedThat]

Like I said I'm not saying the article is wrong. I'm saying it is partially right. X86 minimum spec is way higher than ARM, not even comparable.

You're saying that uArch of a Cortex A710 (ARM v9A) isn't comparable to say... Zen 2 in terms of complexity?

I specifically mentioned Zen 2 because Renoir (mobile Zen 2) is more efficient than both Golden Cove and Gracemont according to chipsandcheese.com

 

[R-T-B]

X86 minimum spec is way higher than ARM, not even comparable.

Really? I'm confident you can do 100% of the things with the core ARM ISA you can do with just x86. Just x86 is pretty limited, tbh. You are talking a 8-bit mode processor if you scale it back to it's roots. At least ARM mandates 32-bits, by virtue of being a more modern design.

If you take ARM ISA and stretch it to the point of rivaling x86, something which has not happened at this point

In some ways, it has. IPC is pretty close on certain chips. x86 just has the snot and power targets clocked out of it.

My point is neither arch is "high/low power" that is every bit the core under them, and it's design.

Case in point: MIPS, one of the arguably least complex instruction sets (at least in history, it used to make ARM look complex), is what China uses in it's TOP100 Supercomputers. They are not weak CPUs at all.

I know this sort of flies in the face of what many believe, but it's true. It's not been told to you this way, because marketing loves to market the ISA like it means something to look you into a software ecosystem. But it really doesn't.

 

[R0H1T]

X86 minimum architecture level today involves AVX2 and more

No it isn't, pretty sure Intel released at least one (Atom) core in the last 2 years without AVX2 support. Their mainstream core uarch didn't have AVX2 enabled on Pentium/Celeron forever till 11th gen IIRC.

 

[Kanan]

Really? I'm confident you can do 100% of the things with the core ARM ISA you can do with just x86. Just x86 is pretty limited, tbh. You are talking a 8-bit mode processor if you scale it back to it's roots. At least ARM mandates 32-bits, by virtue of being a more modern design.

Maybe should've specified that I'm not talking about ancient architectures, anything you can still buy today.

In some ways, it has. IPC is pretty close on certain chips. x86 just has the snot and power targets clocked out of it.

I don't think any ARM arch can rival EPYC Milan upcoming in new supercomputers. I don't completely disagree though. There was a short time, where everything was stalled down due to Intel being a shitty company and AMD not there yet with Zen, where Fujitsu basically had something better, and they're still on Rank 1 with their ARM based Supercomputer until later this year when the new SC with Zen 3 are finished. I think, Intel on the long run, is the (would be) killer of x86 and AMD is the savior and already did it. This is what happens if you let a manager (not EE) that is only interested in bottom line manage a tech company vs the opposite happening at AMD. Intel is getting better but I don't like their big little approach.

My point is neither arch is "high/low power" that is every bit the core under them, and it's design.

Theories are fine, I'm a more practical person. Intel certainly didn't have much going in eg. high efficiency recently, it's just happening now with their newest arch. But I think the Zen approach of the golden middle way with cores that are efficient and strong is the way to go.

No it isn't, pretty sure Intel released at least one (Atom) core in the last 2 years without AVX2 support.

You're quoting me talking about "today" and then countering with the words "last 2 years". I'm not interested in old stuff. Actual new Atom cores found in 12th gen are very well able to do AVX2.

 

[R0H1T]

I'm not interested in old stuff.

What old stuff? Intel's still selling (big)cores without AVX, just because you think it's the "minimum" uarch level doesn't make it so! Intel or AMD can just as well sell a new chip without AVX in it ~ that simply depends on the requirements of the market they're addressing. AVX is not a necessity anywhere.

 

[Kanan]

What old stuff? Intel's still selling (big)cores without AVX, just because you think it's the "minimum" uarch level doesn't make it so! Intel or AMD can just as well sell a new chip without AVX in it ~ that simply depends on the requirements of the market they're addressing. AVX is not a necessity anywhere.

Sure you can buy outdated stuff, buy 3950X for 800€ despite 5950X costing less, etc. Doesn't make it a relevant product nor do I care. I have said pretty clearly that I talk about the newest architectures. Newest Atom does AVX2.

AVX is not a necessity anywhere.

Pretty much completely wrong. There are even old games that don't run without AVX, many programs don't and Windows 8 doesn't like CPUs that don't have AVX, let alone Win 10 or 11. There are also a lot of apps and a few games that use AVX2.

 

[londiste]

You can't blame people for looking at the correlation between good battery life and ARM devices (smartphones, tablets and the new Macs) and thinking ARM is the cause behind good battery life. Intel has failed to deliver comparable battery life. AMD has somewhat succeeded.

Intel did make at least one true attempt to get into smartphone-tablet space with lower end Atoms. The CPUs were not bad and power efficiency was very comparable to contemporary ARM competitors. Their failure was on one hand the entire ecosystem - everything in the SoC that is not the CPU, modems, buses, IO stuff and also the compatibility and breaking into the market part with a different ISA than normal - on the other hand Intel IIRC quite directly stated that profit margins in that market did not meet their expectations and immediately dialed down their efforts.

Current Atoms are not too bad either in both performance and power efficiency. Not at the level of Apple's M1 but for various reasons not much else in the whole market segment is for now.

 

[Wirko]

In some ways, it has. IPC is pretty close on certain chips. x86 just has the snot and power targets clocked out of it.

It may be my obsolete knowledge but on ARM, as the instructions are less complex then on x86, more of them have to be executed for the same amount of work done. Does it make sense to compare the IPC between ISAs?
It's not easy to find a meaningful comparison of instruction sets (while leaving architectures aside), here's one: RISC-V vs. ARM vs. x86 – What’s the difference? but it's quite brief.

This is what happens if you let a manager (not EE) that is only interested in bottom line manage a tech company vs the opposite happening at AMD.

That's a description of a bean counter, not a manager. He or she doesn't have to be an engineer but needs to understand very well what engineers are good for, what they are usually bad at, how to communicate with them and motivate them - I think that's what a good manager should be able to do, engineer or not.

 

[Kanan]

That's a description of a bean counter, not a manager. He or she doesn't have to be an engineer but needs to understand very well what engineers are good for, what they are usually bad at, how to communicate with them and motivate them - I think that's what a good manager should be able to do, engineer or not.

That's the thing. Pat Gelsinger seems to be much better than his predecessor and I don't even consider him ideal, not nearly.

 

[R0H1T]

Newest Atom does AVX2.

Yes & a big part of the reason is big little, why do you suppose they disabled AVX512 support on ADL?

Intel to Disable Rudimentary AVX-512 Support on Alder Lake Processors

Intel is reportedly disabling the rudimentary AVX-512 instruction-set support on its 12th Gen Core "Alder Lake" processors using a firmware/ME update, reports Igor's Lab. Intel does not advertise AVX-512 for Alder Lake, even though the instruction-set was much publicized for a couple of its...

www.techpowerup.com

Arguably the biggest reason they've done this is to make feature set (or extensions) the same across P & E cores.

There are even old games that don't run without AVX, many programs don't and Windows 8 doesn't like CPUs that don't have AVX, let alone Win 10 or 11.

That's not remotely true! I have an 8th gen Pentium on a laptop running win10 just fine, bought it for less than $200 after all the rebates & it's good enough as a backup machine/web browsing for basically any time I need a "spare" PC. Find a better reason than that.

The CPUs were not bad and power efficiency was very comparable to contemporary ARM competitors.

Not quite, they were pretty good in performance but at the time Intel also had a process lead with 22nm. Most mobile chips were on Samsung/TSMC 28nm, when they shifted to a smaller node Intel was left in the dust!

profit margins in that market did not meet their expectations and immediately dialed down their efforts.

You mean after spending billions of dollars & 2(3?) years on that failed "contra revenue" adventure.

 

[Kanan]

Yes & a big part of the reason is big little, why do you suppose they disabled AVX512 support on ADL?

Arguably the biggest reason they've done this is to make feature set (or extensions) the same across P & E cores.

Those have nothing to do with this argument. Don't deviate the argument because you lost your original point. All modern x86 CPUs support AVX2. And then again, I don't agree. AVX is a very important instruction set, that is a main point why it is supported in modern CPUs, it's not optional to drop it, whether big.little or not.

That's not remotely true! I have an 8th gen Pentium on a laptop running win10 just fine,

I had meant to say SSE 4.1/4.2, possibly SSSE 3, the 8th gen Pentium has those.

 

[R-T-B]

It may be my obsolete knowledge but on ARM, as the instructions are less complex then on x86, more of them have to be executed for the same amount of work done. Does it make sense to compare the IPC between ISAs?

That does not really matter. They all get decoded to very extremely RISC like microops at the end of the day anyways. The longer length instructions in x86 take longer to decode, thus aren't exactly time savers.

 

[Kanan]

Part of me wanted Nvidia to win the ARM deal, so that the PC market gets diversified. Now it is unlikely to happen and the x86 monopoly will hold.

 

[Vya Domus]

It does matter in practice for different reasons, for example a compiler might generate dramatically slower/faster code for stupid reasons depending on the target ISA.

 

[londiste]

Not quite, they were pretty good in performance but at the time Intel also had a process lead with 22nm. Most mobile chips were on Samsung/TSMC 28nm, when they shifted to a smaller node Intel was left in the dust!

Not really. The peak of it was Medfield and mid-2012 if my memory serves right. Intel was on 22nm but Atoms were still 32nm for whatever reason. Pretty much everyone else in phones was also on 32nm with 28nm coming in later in the year.

 

[Wirko]

Found a nice article on it. Wanted to share it.

Here, another nice article.

RISC vs. CISC Is the Wrong Lens for Comparing Modern x86, ARM CPUs

Try to investigate the differences between the x86 and ARM processor families (or x86 and the Apple M1), and you'll see the acronyms CISC and RISC. It's a common way to frame the discussion, but not a very useful one. Today, "RISC versus CISC" obscures more than it explains.

www.extremetech.com