Comment by SlipperySlope | Hacker News Reader

Comment by SlipperySlope | Hacker News Reader

The article overstates the case against ARM in the mobile marketplace. Counterpoints not mentioned are: 1. The single company Intel can never match the wide variety of SoC (System on a Chip) components that the diverse ARM fabricators will offer. 2. The ARM fabricators will remain close behind Intel with regard to lithography generations. 3. Intel's Medfield has not been independently benchmarked on shipping mobile devices. 4. Industry partnerships established for the previous Intel mobile part did not create any notable results. 5. In order to dismantle ARM, Intel would have to significantly underprice commodity ARM fabricators - which would destroy Intel's historical margins. Note that Intel sold off a former ARM-compatible division - XScale - for this reason. 6. Furthermore some ARM partners plan to invade the x86 server market with low power, but capable alternatives to Intel servers, thus enlarging the potential ARM market and possibly counter-disrupting Intel's highest margin business.

Furthermore, the major proprietary player (Apple) will likely never give up on their ARM chip to favor Intel. Unlike in the 80s where IBM used Intel chips and Microsoft software and created two heirs to their monopoly, this time will be wildly different.

ARM (and the mobile market) will crush Intel. Intel is seriously late to the game here, and never took performace/watt seriously for decades, and ARM is going to disrupt them hard.

Interesting that you mention Apple/ARM -- don't forget that, eventually, Apple had no choice but to shift from PowerPC chips in its laptops and desktops to Intel. The performance difference was just too vast.

I am fairly sure that the same thing will now happen with x86 and ARM -- it just depends on the timeline.

Apple had to move to x86 because IBM wouldn't invest in development of a fast and low power processor for just one client (Apple) and Intel had a product (Core Duo) ready for them. One client, even Apple, is a very small market. Unit sales of ARM processors combined dwarf x86 sales and there is a huge market for higher performance ARM devices, something that didn't exist for PPC.

That's correct. If you remember, Tim Cook described this limits on PPC based products as a "thermal problem" which is what drove them towards Intel. Presumably, ARM's licenses don't have this problem, in part, as @rbanffy implies, because the ARM architecture doesn't rely on a single source as the PPC did.

That was rather different, as Apple was entirely dependent on IBM and Motorola to produce new chip designs (and Motorola had effectively quit the desktop market). The G5 was actually reasonably competitive with the P4 at the time the move was made, and beat it on some SIMD work. The real trouble was that IBM had repeatedly slipped on its deadlines to produce a low-power laptop-appropriate G5, and the aging G4 architecture just wasn't good enough.

Today, the story is very different. Apple can produce its own SoC designs, as with the A5, or even produce its own ARM _core_ designs, like Qualcomm does, or it can buy from one of a number of healthy vendors who show no signs of quitting ARM anytime soon. If this Intel core had a genuine dramatic performance advantage over the A15 (which it will be competing with, given the timescales), at the same power usage, that would be one thing, but that's far from clear right now.

It wasn't just that Intel chips were faster. They also consumed less power. Just before the switch, Apple shipped a G5 desktop Mac with liquid cooling built in, which shows just how desperate the situation was for them.

Things are much different now. Intel chips are faster than ARM, of course, but they are also vastly more power hungry. Power consumption matters more than speed in most devices these days (otherwise Intel would already be in everybody's phone) and will likely continue to do so for a long time. Intel shows no sign of being able to reduce their power consumption to the level needed while still retaining decent speed, whereas ARM shows every sign of being able to keep increasing their speeds while keeping power consumption where it needs to be.

Granted that the graphs are coming from Intel, but assuming they're accurate I'd say it's a sign Intel can reduce power consumption and retain decent speed.

http://www.anandtech.com/show/5365/intels-medfield-atom-z246...

Maybe they can, but I'll believe it when they see it. Going by past accomplishments, ARM has done very well in making their chips faster, and Intel has not done so well in making their chips more economical. Certainly there could be a change in the game, but precedent heavily favors ARM in this one.

Is that 'They also consumed less power' correct?

http://datasheet.digchip.com/205/205-00367-0-970.pdf gives typical power dissipation as 42W @1.8 GHz, 1.3v

For a later design, http://datasheets.chipdb.org/IBM/PowerPC/970/PowerPC-970FX.p..., section 3.1.5 gives 60W for the 2.2 GHz version at 105 degrees Celsius.

http://en.wikipedia.org/wiki/List_of_Intel_Pentium_4_micropr... gives comparable values for the TDP for 2002-era Pentiums.

I get the impression that IBM's data is real max (but 'Maximum power at 85 C is the ONLY value that is guaranteed by Manufacturing Test.'), while TDP may be lower than max (Reading http://en.wikipedia.org/wiki/Thermal_Design_Power)

Or am I overlooking a factor for #CPUs?

I always thought the water cooling was either an effort to decrease noise or over engineering or preparation for more cores and/or more speed (either of which were a necessity to keep up with Intel chips).

I think you're comparing things a little too early. In 2002, things were still very much up in the air. In hindsight, the PowerPC first started showing signs of real trouble when Apple had to de-rate the entire Power Mac G4 line by 50MHz between the initial announcement and the first ship, but it still did fairly well for quite a while. It wasn't until close to the switch when the promised 3.0GHz G5 failed to materialize that the differences really got big. The Intel switch wasn't announced until 2005 and didn't happen until 2006.

I don't have hard data, but as far as I know, once Macs went Intel, they all got significantly faster and cooler.

This was most visible at the top of the line. The last Power Mac G5 had the liquid cooling system and the top end ran four cores at 2.5GHz. I don't know about the liquid cooled model in particular, but the G5s in general were notorious for being noisy due to fans.

The first Mac Pro which replaced it ran 8 cores at 3.0GHz and was known to be astonishingly silent. I personally used one as my main work computer for years and it was impossible to get the fans to run fast enough to be audible unless you disabled the smart controls by e.g. leaving a kernel panic on the screen for a while. Obviously, this could be due to more advanced thermal design as well, but even allowing for that, you have twice as many cores, running at a higher clock rate, vastly faster in most real-world use, and with a much simpler cooling system.

The laptops had a similar story. The PowerBook G4s would get astonishingly hot, while the MacBooks were much less bad (although still not great initially).

I think that #1 is the more important "pro" for ARM, while #4 is the most important "con" for Intel.

How many times have we seen Intel announce a hardware partnership with a company (especially in mobile) that ultimately went nowhere? I think this comes down to, I'll believe it when I see the devices shipping.

Two other things:

* ARM's big.LITTLE promises to drastically reduce power consumption under low CPU load. It seems unlikely Intel will be able to offer something similar anytime soon.

* If you compare Atom Z2460 with SoCs based on Cortex-A15 (both are supposed to start selling by the end of the year), it doesn't look that well.

Yeah, due to the fact they're out-of-order, the A15s will should be back on top. When the Pentium Pro was introduced with OOO pipeline, it got a 20% speed increase over similarly clocked Pentiums. The issue might be can the A15s keep their pipeline full with such a small cache and memory bandwidth. And how good (if it even has one) is the pre-fetcher? If the float units can't be given data fast enough, it's not going to matter how fast it is in theory.

Note that OOO isn't a boolean checklist feature. It's a design choice that opens up a ton of optimization possibilities (the obvious big one being the ability to issue instructions around a DRAM load). But you need to dedicate silicon to actually make those work. How big is the register renaming pool? What are the edge cases where OOO doesn't work (flags registers are a big one here). Do the SIMD registers get renamed too, or are they serialized? Lots of questions, with AFAICT no public answers.

I think point 2 is the most important -- and I think you're wrong. Intel is already way, way ahead of TSMC, GloFo, and Samsung -- and the gap, if anything, shows signs of extending.

Price will definitely be important, but judging by how hard Intel is coming into this segment, I can only assume that it's willing to match or get very close to ARM component costs.

It looks like a worse ISA will win because it happens to be owned by the owner of the better lithography process. In a really efficient market, those would not be tied. Microsoft using Windows to push Office was bad for society. Why is Intel good for society?

The old RISC vs CISC is completely dead now. Intel have now shown they can match ARM's equivalent powered CPUs in terms of energy efficiency, and that was the last argument the RISC proponents had in the argument.

Intel's chips are internally RISC, with the only overhead of decoding and fusing the ops and the transistors to do this. And despite this overhead, they're still miles ahead.

The only argument left then is the actual instruction sets at asm level, and here, CISC has an advantage, because they can do so much more in one instruction (take memory addressing modes for example), so they're more compact from an executable point-of-view and instruction cache point-of-view.

CISC isn't necessarily more compact. Since there are more instructions, each instruction takes up more space. For example in ARM's thumb encoding, each instruction takes up 16 bits. A typical x86 instruction takes much more space.

I wonder how much of an x86 chip is dedicated to handling legacy instructions or largely unnecessary instructions. How much saving could be achieved by a radically simpler instruction set (for example no neg/call/ret/push/pop/bcd instructions). If you look at the http://en.wikipedia.org/wiki/X86_instruction_listings then most of the instructions seem unnecessary. Does anybody have some information about this?

> I wonder how much of an x86 chip is dedicated to handling legacy instructions or largely unnecessary instructions. How much saving could be achieved by a radically simpler instruction set (for example no neg/call/ret/push/pop/bcd instructions).

It's been a long time since I was in a closely related biz, but back then, the answer was "not much if you want to go fast".

Instruction decode isn't a huge deal. Micro-op sequencing isn't a big deal if you're going out of order (and many risc instructions tend to get broken into micro-ops).

The number of architectural registers used to be a big thing, but X86_64 is good enough. (I don't know if anyone got around to renaming memory.)

Okay, I probably should have been more specific - they might take up more memory, but you generally need less of them to do the same thing in RISC, so that usually more than compensates.

Instruction decode isn't really difficult - 4-way decode could arguably be called difficult, but there's not much of that.

Intel is beating ARM because Intel are a process ahead. If an Intel fab made ARM chips on the same process they use for i7s, that ARM chip would beat the pants off the Intel chip made on the same process.

Intel does waste lots of chip real estate on caching big instructions (http://www.anandtech.com/show/3922/intels-sandy-bridge-archi...) and on wide decoders that fault and go to microcode when they encounter some instructions. Think of how inefficient the SSE and AVX encodings are, because the need to escape 8086 code. Or the extra byte needed to access the extra 64bit registers. New CPUs are seldom limited by decoder width on regular code but it does cost watts and it costs space that could have been used for cache or for making the chip cheaper.

I'm not even sure ARM is the best ISA. It doesn't matter which alternative is the best: I just predict that the best won't win because Intel will use their superiority in manufacturing to artificially boost the success of their lackluster CPU design guys.

What are you talking about?

Intel's 65nm (Pentium M) can still beat similarly clocked (1.5 Ghz) ARMs at 28nm - that's not due to process, that's due to a much more powerful architecture designed for speed over power consumption (Pentium M, then Core, Core 2).

The instructions that fault are the very rarely used ones like weird string processing ones. Those are not going to be used in the middle of number-crunching loops.

Escaping float ops hasn't been needed since the 486 days, so I'm not sure what on earth you're talking about.

Intel's branch predicting and memory / cache pre-fetching are the best in the business. To say their CPUs are "lackluster CPU design", is pretty ridiculous, despite Atom being pretty crap.

> Intel's 65nm (Pentium M) can still beat similarly clocked (1.5 Ghz) ARMs at 28nm

You can only compare chips that have the same power consumption. A 1w chip against a 45w (or whatever) chip is not a comparison you can make.

Specifically, Intel now demonstrates a Penwell SoC with a Saltwell (Atom) core made on Intel's 32nm LP process and it benchmarks well against Cortex-A9 chips made on TSMC's 40nm process. I said that I guess that if Intel manufactured an ARM chip on Intel's 32nm LP process with the same die area and power consumption as their x86 chip, the ARM chip would be preferable. Like Anand said, We'll see how this SoC stacks up against multi-core Cortex-A15 chips on TSMC's 28nm process.

If Intel changes the priority and makes their Low Power process transform to the next node together with (instead of after) the high-performance process (currently one step ahead at 22nm), then they will be seriously ahead of whoever manufactures ARM chips (TSMC and others).

> The instructions that fault are the very rarely used ones

They complicate the decoders and take up space and I think take up latency (clocks) during decoding. This costs money even when these instructions are not used.

> Escaping ... so I'm not sure what on earth you're talking about

REX prefix: http://www.x86-64.org/documentation/assembly.html

> To say their CPUs are "lackluster CPU design", is pretty ridiculous

Their power efficient design is in-order and pretty crap. You should read what Linus Torvalds had to say about it. Yes, Conroe was awesome and Sandy Bridge is awesome, but that is not what we're talking about here.

Intel has just announced a 32nm Atom SoC. Meanwhile, the first products based on TSMC's 28nm process have started to ship.

Yes, Intel is usually the first to bring a new process online (and their 22nm Ivy Bridge is due to ship in a few months), but for the entire history of the Atom product line, Intel has failed to make use of their fab advantage. They probably never will, so long as it is more profitable to put the desktop/laptop/server product lines on a new process first.

They're certainly ahead on process. It remains to be seen if that will be enough.

Currently there are no notable shipping devices with Intel CPUs. Thus, up to now, Intel's fabrication process advantage has not been enough. I speculate that the rest of the world will stay close enough to Intel to make the advantage non-disruptive to ARM. There is however a technological chasm that ARM must successfully cross - namely the transition to high performance systems enabled by the ever higher transistor budgets permitted by Moore's Law.

One can imagine that big L2 caches, high memory bandwidth and eventually 64-bit addressing will be important for mobile devices. On one hand Intel can attempt to drive existing high performance x86 solutions to low power, and on the other hand ARM can attempt to extend their architecture for high performance - given more transistors and better power characteristics available to both sides, the rest of the world lagging Intel.

1. Intel have a deal with TSMC to allow TSMC to fab Atom silicon for custom SoC solutions. 2. Intel is way ahead of other fabricators - die size alone isn't the only metric, there's things like powergating which Intel are way ahead on. They're about 1 year ahead. You also can't compared process sizes between fabricators, because they're not the same. 6. Only relevant for non-CPU / memory bound stuff - ARM's still very poor on float throughput and memory bandwidth (bus size is crap and still only 32-bit access to memory), so it'll only be things like VMs and webservers that use them for power consumption. It'll probably hurt AMD more. They still can't get their 64-bit stuff going, they're trying to do PAE (Physical Address Extension), and we know how well that worked on Intel's P4...

Reply on news.ycombinator.com