Comment by jonstokes | Hacker News Reader

Comment by jonstokes | Hacker News Reader

"The entire x86 computer architecture is living on borrowed time. It’s a dead platform walking. The future belongs to ARM, and Apple’s A-series SoC’s are leading the way."

I can't believe that lines like this are still being written. I devoted the better part of a career to successfully debunking this idea that ARM has some sort of magical power efficiency advantages over x86. I can, however, console myself that such claims are exceedingly rare. Most people who lived through the "RISC vs. CISC" platform wars and made such claims conceded defeat a long time ago.

Anyway, I addressed this issue in one of my last CPU articles for Ars back in 2011:

http://arstechnica.com/business/2011/02/nvidia-30-and-the-ri...

Quote:

"It's also the case that as ARM moves up the performance ladder, it will necessarily start to drop in terms of power efficiency. Again, there is no magic pixie dust here, and the impact of the ISA alone on power consumption in processors that draw many tens of watts is negligible. A multicore ARM chip and a multicore Xeon chip that give similar performance on compute-intensive workloads will have similar power profiles; to believe otherwise is to believe in magical little ARM performance elves."

Also, I think that benchmarks run in this article for the purpose of comparing the iPad to the MacBook are pretty worthless. I don't have a good answer for how to comparatively benchmark them, or even if that makes sense. Meaningful cross-platform benchmarking is really hard, just ask the SPEC people.

Finally, in the article linked above, I predicted that the only way ARM is a threat to x86 is because it's cheaper, and that if ARM can get within something like 2X the performance of Intel's higher end CPUs, it might have a shot by virtue of being widely and inexpensively licensed.

Anyway, the TL;DR here is that those benchmarks say more about the software stack (the benchmark software included) than the do about the CPU and GPU hardware, and the reports of x86's death at the hands of PowerPC^H^H^H ARM are, as always, greatly exaggerated.

I don't know if outdated "RISC vs. CISC" thinking is what motivated Gruber's statement, but I think there are much more convincing ways to justify it.

If I could change Gruber's statement, I would say that "The future belongs to custom-designed SoCs"

The advantages that Apple derives from the A-series SoCs is not due to any inherent advantage of ARM vs. x86, but because Apple has full control over the design and manufacturing.

- Apple can design an SoC for a specific product given the manufacturing process available at the time: see last year's one-off 3-core A8X, because adding a 3rd core was a better tradeoff than increasing clocks. This year, the A9X is back to 2 cores but much higher clocked than the A9.

- Apple gains a competitive advantage by building processors only for themselves, and can catch the rest of the industry off guard (see: ARMv8 A7). They also get to follow their own principles (two fast, wide cores) rather than being forced into everyone else's marketing hype (8 heterogeneous, slower cores)

- If Apple wants a stronger GPU they can just license it from PowerVR, rather than having to lobby Intel and hope the resulting silicon is better (or worse, having to add an external GPU)

- Apple can even hedge its bets w.r.t fab processes: see the dual-sourced TSMC/Samsung A9

ARM is a threat to x86 because anyone can design/buy an ARM core, design an SoC around it, and manufacture it anywhere they want. Intel/AMD can't come close to that flexibility, and on platforms where Win32/Intel binary compatibility is irrelevant, x86 will decline/stay irrelevant.

I think this is a really important point that is overlooked. Other solutions get what Intel offers, sure they can suggest but they can't tell Intel "Hey don't waste transistors on that feature, make the cache bigger." Or other things that Intel does to widen the appeal of its machines.

The GPU issue is a good one. Every desktop system my friends have has two GPUs in it: one built into the processor that no one ever uses, and one they add separately so that they can play any games. AMD at least has half-decent GPUs built into their CPUs, but their CPUs are only half-decent anyway. Meanwhile, Intel refuses to license any Thunderbolt external GPU docks or anything of the sort, despite how popular they would be with laptop-toting would-be gamers, because they want to promote their own, awful GPUs instead.

Likewise the core count issue. Jeff Atwood's recent blog post[1] about how Android JS performance has stagnated because Android SoC single-core performance hasn't improved in years blew me away, but in retrospect it makes sense; it's presumably more engineering work to design a faster CPU core than it is to just put more of the onto a die and call it a day.

[1] https://meta.discourse.org/t/the-state-of-javascript-on-andr...

Thunderbolt 3 includes official support for external GPU docks. Intel's reasoning for not allowing them sooner was due to hot-plugging issues but they could be spinning it around.

Source: http://www.anandtech.com/show/9331/intel-announces-thunderbo...

When are you going to write us more amazing CPU explanations, Jon? :-) Thanks for all your great work over the years.

Gruber is clearly exaggerating a bit (x86 may not be on borrowed time, etc.), but there is something notable happening, which is that Apple’s (and everyone’s) ARM CPUs are getting fast enough to push into many applications where Intel previously had a big advantage.

Benchmarks aside, people can now edit video, render complex 3d scenes, analyze huge amounts of data, etc. on pocket computers with tiny power-efficient SOCs.

For consumer use, desktop workstations – and now even beefy laptops – are becoming a niche/specialist tool

> Gruber is clearly exaggerating a bit

Yes, even if the Apple ARM cpus are ready to displace Intel at the low end, no one besides Apple is going to be using them. So the issue is moot for the greater PC industry really.

Heh, I appreciate the encouragement :) However, good explanatory writing about tech is extremely difficult and time-intensive. If I were working full-time on writing about CPUs (or other tech) and trying to do the kind of stuff I used to do for Ars, I might be able to manage two articles a month... maybe realistically it might be like 1.5 articles a month. And that would be going at it full time.

I'm just not really sure that at this point in my life (I just turned 40), that I want to trade, say, three years of my 40's for about sixty or so CPU (or other tech) explainer articles. I dunno... I might... I'm not saying that I /don't/ want to do it. In fact, I think about it sometimes, because there's a whole lot to be said for doing something that you're really good it and that people appreciate, even if it's sort of niche and random like writing about CPUs. So maybe I could be talked into it. But I also think, maybe I'd rather try and write a zombie novel, or a graphic novel, or something else random. Who knows... First-world problems and all that...

I think the 'death of x86' issue is that currently, CPUs are 'fast enough' for most people; for that matter, they're faster than a lot of people need. Apple's iPad processors are fast enough for what most non-professionals need to do, and looking at these benchmarks I'm wondering how many e.g. students, even when they do need a laptop, even need one with as much power as the iPad Air line has right now.

Putting it another way: if Dell started offering ARM desktops with half the performance of the iPad Pro and the size and price of an AppleTV, AND they could run typical Office apps, would businesses (which were not bound by legacy software) start buying those instead?

I think the sense, in 2011, that ARM was the Intel-killing super robot from the future wasn't about ARM's ability to scale up, but about Intel's ability to scale down. Intel has held the performance crown for quite a while, but there's always seemed to be a minimum level of requirement involved; to buy a modern Intel system is to buy a system with a certain minimum power draw, weight, functionality, etc. ARM is appealing in the sense that it can start very small (see: RPi) and scale upwards from there (see: iPad Pro). Intel, is more than capable of reaching the high end, but the sentiment was always that they couldn't reach the low end, whether that was strictly true or not.

It's not a question of will x86 die; it won't, really. It's a question of does Intel bring something to the table with x86 that is necessary for the mass market? What's dying isn't Intel or x86, it's the assumption that we're stuck with Intel and x86 for doing any real work.

Edit: I sure hope I don't sound like an idiot in this post because I've been a big fan of your writing and insights so I'm hoping I'm not showing off how dumb I am.

My take here is not about magic ARM dust. I do remember the magic RISC dust days -- and I, as a lover of the elegant -- was on the losing side of that one. I think the basic threat to Intel and x86 is this:

- Intel operates as a high-margin business. Their CPUs aren't $10 components, and they don't really have any serious competition. The whole point of the word "Wintel" was that all the value in a PC was going to Microsoft (for the OS) and Intel (for the CPU). These mobile ARM chips aren't like that at all. They're just another component in the device, like the camera and the display.

- Scale. Ben Evans makes the case well here: http://ben-evans.com/benedictevans/2015/11/7/mobile-ecosyste... . He writes:

"And fourth, the broader scale advantage - the ARM/iOS/Android ecosystem is moving towards selling 10x more devices each year than the Wintel ecosystem. That's a similar disparity to that between PowerPC/Mac and Wintel 20 years ago."

I realize too that running a benchmark like GeekBench on iOS is not apples-to-apples running it on Mac OS X. But even if it's not the case that the iPad Pro is a faster computer than a MacBook, it's close. But I really do think it's faster -- it shows in browser-based benchmarks, which I think are a fair comparison even across iOS and OS X.

But in short: Intel doesn't want to (and possibly can't) compete on price, and they can't compete on scale. Neither of those things are magic.

> I can't believe that lines like this are still being written.

Gruber is an unpaid Apple PR flack.

Unpaid?

I don't really think that's fair. I like and respect Gruber. I just think he's way wrong about this.

I think Gruber demonstrates wisdom about Apple almost all of the time. The trouble is, he is SO biased, that he tends to really go overboard from time-to-time which diminishes my overall faith in what he has to say. It's like he literally will never acknowledge someone else might be able to do anything at all better than Apple.

What if there's an upper limit in performance, where gains get smaller and smaller with each generation? What if Intel is already hitting that ceiling and ARM isn't? What if in the next few years this will help ARM catch-up with Intel, while still having its chips cost much less? That would be a problem for Intel, no?

Intel's competitor to Apple A9 is Atom. Core i7 is not a competitor. If the A9 is within 2x of the latest Core i7, then that's great for ARM/Apple, but the real point is the A9 is probably at least 50% faster than the latest Atom.

Intel's Atom has never really caught up with ARM chips despite its process advantage (even now, although much smaller than before, and Intel process generations are also growing further apart). It had a very brief moment in which Atom was more or less equal to the high-end chips in benchmarks - but only because Intel had a generation and a half process advantage at the time being on 22nm FinFET vs 28nm planar for the ARM chips.

Atom is so hopeless in mobile, that Intel has started licensing its design to Chinese chip makers - which will build it on 28nm planar for next year anyway, making it a very pointless exercise, indeed.

The only way it got some foothold in some mobile devices this year is because it baited and switched Asus with a heavily discounted high-end Atom that cost Asus as much as a mid-range Qualcomm chip (which I still think should be highly illegal and worthy of an anti-trust suit). I'm talking about the Zenfone 2 - and even then, the phone got popular mainly because it was the first with 4GB of RAM, and Intel Atom happened to be there for the ride.

If Intel didn't lie when it said it will stop subsidizing Atom in mobile, then that kind of tactic will end and Intel won't be able to trick any other OEM anymore. But I suspect it did lie, because at the same time it said that, it merged the mobile and PC chip groups together to hide the financials of the mobile group. It also started selling $150 Atom chips as "Pentiums", where they probably have a profit margin of 500%, and Intel will use that profit to once again subsidize its mobile chips to drive competition out of the mobile market (which again - should be very illegal).

The problem for Intel is that this will still be a very slow process, and by the time it gets a serious foothold in the mobile market, ARM chips will start to become a threat to it in the PC market.

> What if there's an upper limit in performance, where gains get smaller and smaller with each generation? What if Intel is already hitting that ceiling and ARM isn't? What if in the next few years this will help ARM catch-up with Intel, while still having its chips cost much less? That would be a problem for Intel, no?

There's no polite way to put this, but you don't know much about the problem domain so speculating like this is entirely a waste of time.

The problem isn't the fairly minute differences between ARM and x86, the problem is that your transistor budget (and thus power budget) is largely spent trying to keep alive the myth that there exists a single thread of execution and that one instruction executes until it's done and that's when the next one starts.

I say that there are minor differences between x86 and ARM because compared to something like the Mill or a grid processor out of UT Austin (your data pipeline was physically routed between functional units on the die by a reconfigurable fabric) the differences seem minor. They both have the same underlying microcodes, the x86 front-end is just more complicated in some ways and simpler in others.

You raise an interesting question. For the most part, CMOS scaling has dominated for the past 25 years or so--which has meant that investments related to carrying a specific volume design forward (e.g. x86) has dominated. Lower-volume and lower-capitalized designs just weren't able to ride the same curve. But in a world where CMOS process scaling doesn't dominate, it raised the possibility that more specialized designs may play a larger role.

Reply on news.ycombinator.com