4 ms·

Counterpoint: X86 is dying. It is my understanding that Intel/AMD will have a hard time making x86 chips have a competitive performance-per-watt with ARM chips

by andjd 4y ago

Counterpoint: X86 is dying. It is my understanding that Intel/AMD will have a hard time making x86 chips have a competitive performance-per-watt with ARM chips. This is because x86 has a large instruction set, and those instructions must be physically implemented in the silicon. It's why you never had an x86-powered smartphone. Efficiency is a killer app for laptops. It lets manufactures design thinner lighter computers with all-day or longer battery life. This is more valuable to most consumers than support for every legacy app that doesn't get ported to run natively on ARM. Most recent software written in C#, Java, or Electron can be trivially recompiled to target ARM architectures. There's a chicken-and-egg situation here, but Apple's M1 launch is likely to push Microsoft off the fence and get them to seriously support Windows-for-arm. Microsoft's x86 emulation is currently bad, but if they can make it half-decent, the corpus of software that can't or won't be ARM-native will be usable -- if slow -- so this won't substantially impair adoption of ARM PCs. There will still be some niche users who need run software that's optimized for x86, in the same way there are niche users who still need to use parallel ports or PCI cards. The industry will support those users, but because they are a niche, volumes will be low and prices high. Also important here, is that besides competing with Macs, Windows PCs are also competing with Chromebooks. Chromebooks have already dominated the large education market, and companies are beginning to make premium Chromebooks because there's a whole generation of young adults that grew up using them and it's the computer platform they're most comfortable with. There are already ARM-based Chromebooks, so this is a platform where backwards compatibility with x86 is largely a non-issue. This will further push Microsoft to support ARM chips to avoid becoming an also ran. I'm not saying this is how it's going to be. But your take is pretty maximalist on x86. There's a similar story where x86 becomes basically irrelevant in 5 years or less.

smoldesu 4y ago
I understand where you're coming from, but I don't think you have the entire story; x86 is old and massive. You're correct, it has hundreds upon hundreds of instructions, many of which require hardware accommodations that can be fairly expensive to implement on-die. On top of that, it has a bevvy of ISA extensions like SSE and AVX that require even more hardware, and can drive complexity even further. The issue is that ARM is getting to the same point. ARM is no longer RISC; it's instruction set is anything but reduced. Much like x86, it has hardware-specific instructions that correlate to esoteric software features like Javascript execution or OS sandboxing. It also has lots of extensions like NEON, as well as Apple-specific ones that cannot be open sourced or replicated by third parties that want to also adopt ARM. The push to "modularize" ARM has failed, so much so that we update the spec with new versions requiring specific extensions to qualify. ARM from a hardware implementation perspective has become a mess, arguably moreso than x86 has. On top of that, it's extensions are fairly disappointing when compared to x86 ones. NEON gets smoked by AVX by 2 orders of magnitude. Floating point math is utterly lethargic without hardware extensions. Extended precision on ARM is simply not very well-supported. All of these things come back to bite it when considering as a desktop architecture, and ultimately represent a conflict of interests with it's application; ARM was split between running on $1 microcontrollers and $1,000 iPhones. There was no good solution that would please everyone, so they tacked on complexity until both sides were happy. And that leads us to where we are today, with two bloated architectures fighting it out for superiority on completely separate terms. x86 is by no means a perfect architecture, but ARM also suffers a lot of the same issues. They're both old, they're both extremely specific in implementation and require lots of licensing to produce. The advantages of switching to ARM are controversial and marginal, much like the case for moving away from x86. On top of that, it's not like 5 watt x86 chips don't exist; they simply aren't comparable to the M1 or other high-performance ARM chips because they're manufactured on silicon that's 2-3x less dense. On paper, x86 has a lot of efficiency it can recoup by moving to more recent lithography, and until Apple is fighting x86 manufacturers on equal footing (which should kick up around EOY '23), I think it's too early to say who's "won" the desktop market. We simply don't know, but counting x86 out of the game is a pretty hype-driven musing in my opinion.
- wmf 4y ago
  Note that a lot of the ARM legacy has been dropped from ARMv8.
- kcb 4y ago
  I think a lot of the x86 performance per watt issues are really Intel issues. Intel has spent the last half decade increasing max power consumption to its limits to eek out just a bit more performance to justify launching a new generation. It's clear that they're turning a corner but even the just released 10nm big.little Alder Lake chips fall behind AMDs several year old TSMC 7nm chips in performance per watt.
  - smoldesu 4y ago
    I mean, you're right. But that's not really a surprise considering that transistor density directly correlates with power efficency. You can have the best engineers in the world designing your chips, but it won't really matter if your competitor can undercut you on node size. Conversely, you can have the most efficient silicon but it won't always give you the performance crown. Considering how "bad" Intel's chips should be on paper, it's still fairly impressive to see the kind of single-thread performance they dish out. There's probably also something to be said for process efficiency too. Sure Apple's 5nm silicon isn't power hungry, but getting those chips into mass production is a hugely wasteful effort. 5nm's yield was awful for those first few years, and you could definitely tell Apple was feeling it with some of their binned chip options. Kinda makes me wonder how much water, silicon, neon and power was wasted manufacturing all those busted dies that never made the cut.
- klelatti 4y ago
  Sorry, this is a complete misrepresentation of the whole Arm architecture in 2022. The idea that ARM64 running on M1 or any other desktop or server is hobbled by the fact that you have a completely different ISA running on a $1 microcontroller is just wrong.

by andjd 4y ago

smoldesu 4y ago

I understand where you're coming from, but I don't think you have the entire story; x86 is old and massive. You're correct, it has hundreds upon hundreds of instructions, many of which require hardware accommodations that can be fairly expensive to implement on-die. On top of that, it has a bevvy of ISA extensions like SSE and AVX that require even more hardware, and can drive complexity even further. The issue is that ARM is getting to the same point. ARM is no longer RISC; it's instruction set is anything but reduced. Much like x86, it has hardware-specific instructions that correlate to esoteric software features like Javascript execution or OS sandboxing. It also has lots of extensions like NEON, as well as Apple-specific ones that cannot be open sourced or replicated by third parties that want to also adopt ARM. The push to "modularize" ARM has failed, so much so that we update the spec with new versions requiring specific extensions to qualify. ARM from a hardware implementation perspective has become a mess, arguably moreso than x86 has. On top of that, it's extensions are fairly disappointing when compared to x86 ones. NEON gets smoked by AVX by 2 orders of magnitude. Floating point math is utterly lethargic without hardware extensions. Extended precision on ARM is simply not very well-supported. All of these things come back to bite it when considering as a desktop architecture, and ultimately represent a conflict of interests with it's application; ARM was split between running on $1 microcontrollers and $1,000 iPhones. There was no good solution that would please everyone, so they tacked on complexity until both sides were happy. And that leads us to where we are today, with two bloated architectures fighting it out for superiority on completely separate terms. x86 is by no means a perfect architecture, but ARM also suffers a lot of the same issues. They're both old, they're both extremely specific in implementation and require lots of licensing to produce. The advantages of switching to ARM are controversial and marginal, much like the case for moving away from x86. On top of that, it's not like 5 watt x86 chips don't exist; they simply aren't comparable to the M1 or other high-performance ARM chips because they're manufactured on silicon that's 2-3x less dense. On paper, x86 has a lot of efficiency it can recoup by moving to more recent lithography, and until Apple is fighting x86 manufacturers on equal footing (which should kick up around EOY '23), I think it's too early to say who's "won" the desktop market. We simply don't know, but counting x86 out of the game is a pretty hype-driven musing in my opinion.

wmf 4y ago
Note that a lot of the ARM legacy has been dropped from ARMv8.
kcb 4y ago
I think a lot of the x86 performance per watt issues are really Intel issues. Intel has spent the last half decade increasing max power consumption to its limits to eek out just a bit more performance to justify launching a new generation. It's clear that they're turning a corner but even the just released 10nm big.little Alder Lake chips fall behind AMDs several year old TSMC 7nm chips in performance per watt.
- smoldesu 4y ago
  I mean, you're right. But that's not really a surprise considering that transistor density directly correlates with power efficency. You can have the best engineers in the world designing your chips, but it won't really matter if your competitor can undercut you on node size. Conversely, you can have the most efficient silicon but it won't always give you the performance crown. Considering how "bad" Intel's chips should be on paper, it's still fairly impressive to see the kind of single-thread performance they dish out. There's probably also something to be said for process efficiency too. Sure Apple's 5nm silicon isn't power hungry, but getting those chips into mass production is a hugely wasteful effort. 5nm's yield was awful for those first few years, and you could definitely tell Apple was feeling it with some of their binned chip options. Kinda makes me wonder how much water, silicon, neon and power was wasted manufacturing all those busted dies that never made the cut.
klelatti 4y ago
Sorry, this is a complete misrepresentation of the whole Arm architecture in 2022. The idea that ARM64 running on M1 or any other desktop or server is hobbled by the fact that you have a completely different ISA running on a $1 microcontroller is just wrong.