x86 is history

[Cogman]

Beating intel isnt impossible. The guys at VIA have delivered a better lowend chip than atom, despite a huge process handicap and despite an engineering team and R&D budget that is ridiculously small compared to what intel has.

But Samsung and Microsoft are not exactly small players, they certainly can afford the resources. I also think nVidia might well pull this off, they have assembled a team of very experienced CPU guys and nVidia arent new at designing complex highend chips.

The type of chips that nVidia designs vs the type of chips intel designs are very different. nVidia chips are mostly focused on doing very few things and then copying it 800x. Intel does a whole lot of things and copies it, maybe, 4x.

Just because you have experience making CPUs doesn't mean you have experience making GPUs and visa versa. Just because your a big company, doesn't mean that you have the right engineers to be competitive. Resources are a big issue, however, resource come in more forms than just the amount of money you have.

Samsung is better setup than microsoft to compete with Intel. However, I wouldn't say they are absolutely poised to strike at intel.

Now beating intel in single threaded integer performance will be hard for anyone not called IBM, but you dont need that to be competitive. A few things are happening at once; the importance of raw cpu performance is dwindling, software is increasingly becoming parallel and much of the heavy stuff is being offloaded, either to the cloud or the gpu's.

Bull. That is the whole reason for going to a new architecture. People aren't going to switch architectures because they give them warm fuzzies. They are going to switch for performance/cost reasons. A data center isn't going to spend millions switching to a technology that is slower than what they currently have (That is where intel gets most of its money from, not from the average joe consumers).

Until someone develops a CPU that is competitive (Even if it is 10% slower) to intel's cpus in price, power consumption and performance there is no way the x86 is going to just disappear. The only one arm has is power.

I understand that they could go the GPU route and have 10000x ARM cpus on a chip, but they don't right now. I'm not talking about individual CPU performance, I'm talking about package performance in a threaded environment when I talk about performance.

Therefore nVidia (or MS or samsung or anyone) doesnt need a cpu that outperforms intels highest end cpu's in single threaded integer performance. For desktop, they need something that is faster than the current purely mobile ARM designs and that achieves something Core 2 class. That shouldnt be that hard, in fact a cortex A15 with a desktop class IO and memory interface would probably get you there already. Add to that a better GPU and GPGPU capabilities than intels, if needed throw more cores at it, while still retaining a smaller (therefore cheaper) chip that is probably a lot more energy efficient too, and you have a competitive chip. Not core i7 beating, but something that can compete with intels mainstream product range, selling in to notebooks, all in ones and desktops running ChromeOS or windows for ARM. And that is going to undermine intels fat monopoly margins.

Again, the consumer market isn't the market to beat (Even though throwing more CPUs into the design is the way to go to attack intel).

For servers, competing with intel should be easier even. Many workloads, especially "cloud servers" are all about performance/W and performance/$. Small wonder Atom servers are getting so popular, you even see 512 cpu servers based on atom. That is, despite Atom currently not even having ECC support (ouch), despite having to chose between virtualisation, 64 bit support and dualcore. You cant have it all today in an atom, but thats not stopping big OEMs and startups alike to design servers around it.

It is, however, stopping the companies that intel derives most of its money from investing in atom servers.

Building an ARM chip that performs like an Atom on server loads, at similar or lower price points and power consumption is very doable. Id even say 'easy'. And you can do so while offering the features that intel deliberately disables on atom to help save its xeon margins. Marvell is working on this, Im sure nvidia is, and it doesnt seem to be something particularly unlikely to succeed.

So again, why do you think that nVidia is so much better poised to succeed at this than intel, a well established company already in the market. Why do you think that intel couldn't make a pretty dang good ARM cpu if they decide to throw their full weight into it.

Could intel react? Sure, they could drop some of the restrictions on atom. And allow $30 atoms to go in to server sockets where they used to sell $300 xeons. That will help their margins.

Then there is HPC; nvidia is already making some inroads with tesla. Why should it rely on x86 cpu's to feed their tesla's? You could cram a dozen cortex a9 class integer cores in a single fermi and it would barely have an impact on power consumption or die size. Now Cortex is probably too slow for this, project denver should fix that and offer nvidia a complete HPC building block.

Again, very small market that is even dwindling because of box markets Which, ironically the boxes generally use ARMs anyways... Soo, um, ARM and nVidia would be stealing the market from... ARM. Not exactly a game changer.

Its not quite dead, their new ISA just failed to deliver a better cpu than the competition, including the x86 competition. So from a "merchant' 64 bit server cpu, it has become an HP only chip to replace HPs previous risc cpu's. In that sense, its an utter failure, it delivered nothing that couldnt and hasnt been done with existing ISAs. It just costed 5 or 10 billion dollars more to get there.

Again, to think that intel developed a whole architecture that taught their engineers nothing is ignoring a lot. Whose to say some itanium solutions didn't creep back into the x86 architecture?

You mean they all failed to become intels bread and butter. They where all destined to replace x86, with the exception of their ARM chips.



Sorry, in what transition?

Markets don't just change overnight. If ARM starts to replace intel in server applications it isn't going to be x86 today and ARM tomorrow. There will be at least a couple of years of transition. Intel will surely notice that and react to maintain their market share. There will be a transition. The death of x86 is not going to be an overnight thing.

VIA's garage company did just that from a technical POV. As AMD is doing arguably with their fusion chips. Besides, if you look at the ARM ecosystem, thats not exactly "joe's garage". Its a sky scaper that totally dwars intel's shop in size and R&D expenditure. Dont forget the ARM business model, where 100s of companies essentially pool a big part of their R&D money in to ARM holdings thats doing a lot of the design for all of them. Combined, they outsell intel 10-1. Thats why a standard ARM core comes with a ~$0.2 core license. I suspect an Atom chip carries a per core R&D cost that is closer to $5 or $10. Its certainly more than the silicon cost of the chip. The same problem that AMD and VIA have competing with intel (volume) is going to be intels problem competing with ARM.

I'm not forgetting ARMs setup or architecture.

What makes it possible for several companies that have never created any notable CPU to somehow make something that takes intel out of the picture, that intel couldn't do itself? Why wouldn't ARM licence to intel? Do you really think that the people that currently make the best CPUs on the market couldn't switch to something that has already been designed, and make it better? Do you really think that the people who currently dominate the server market share are going to suddenly just roll over and die? Do you really think intel is so inept as to not be able to do exactly what you say the other intel killers are going to do?

I just don't see Intel a company that is just going to roll over and die. Even if they were kicked out of the processor markets, that wouldn't totally kill them (Though, they would probably have to sell their fabs ALA AMD). They still make a ton of other products that aren't processors.

Same goes for microsoft. Even if they lost their entire OS market, they still are extremely diverse. They would be able to go along with their xBox sales.

The days of big electronics companies failing are behind us IMO.

 

[P4man]

Bull. That is the whole reason for going to a new architecture. People aren't going to switch architectures because they give them warm fuzzies. They are going to switch for performance/cost reasons. A data center isn't going to spend millions switching to a technology that is slower than what they currently have (That is where intel gets most of its money from, not from the average joe consumers).

Datacenters will gladly switch to a slower architecture, like x86 instead of Power 6/7 if that architecture gives them better performance per watt, per dollar and per U of rackspace. 3 criteria where ARM has significant and inherent advantages over x86.

Of course some workloads are latency bound and need fast CPU's that you cant replace with many more slower ones, but many more workloads are parallel enough to be run on slower cpu's than xeon or opteron. But I wouldnt even assume project Denver turns out to be slower than typical Xeon or opterons.

I understand that they could go the GPU route and have 10000x ARM cpus on a chip, but they don't right now. I'm not talking about individual CPU performance, I'm talking about package performance in a threaded environment when I talk about performance.

Quad core ARM chips for server loads are already available:
http://de.nuhorizons.com/news/news.asp?newsid=1031

Thats only just the beginning. BTW, most datacenters dont look at performance per chip or per package. They could care less how many chips or even cores are in there. They care about rackspace, watt and dollars. Only exception to that is software that is licensed per chip or per core, but thats a software licensing model that is going to have to change anyhow.

So again, why do you think that nVidia is so much better poised to succeed at this than intel, a well established company already in the market. Why do you think that intel couldn't make a pretty dang good ARM cpu if they decide to throw their full weight into it.

You may want to reread the thread title: it reads 'x86 is history', not 'intel is history'. Although it may amount to the same thing, because I just dont see intel again taking an ARM license once again, especially not to produce chips that compete with their own x86 monopoly.

Again, very small market that is even dwindling because of box markets Which, ironically the boxes generally use ARMs anyways... Soo, um, ARM and nVidia would be stealing the market from... ARM. Not exactly a game changer.

What are you saying? ARM white boxes used in HPC? HPC as in High Performance Computing? Where on earth are those? I certainly havent ever seen one yet.

Markets don't just change overnight. If ARM starts to replace intel in server applications it isn't going to be x86 today and ARM tomorrow. There will be at least a couple of years of transition. Intel will surely notice that and react to maintain their market share. There will be a transition. The death of x86 is not going to be an overnight thing.

oh, I agree. And it will take far more than a just a few years as well, it will take decades. But one thing will happen almost overnight, that is the assault on intels gigantic margins. Intel will very rapidly be forced to chose between cutting margins significantly or allowing a new ISA to make inroads in to its most lucrative markets.

What makes it possible for several companies that have never created any notable CPU to somehow make something that takes intel out of the picture, that intel couldn't do itself? Why wouldn't ARM licence to intel? Do you really think that the people that currently make the best CPUs on the market couldn't switch to something that has already been designed, and make it better? Do you really think that the people who currently dominate the server market share are going to suddenly just roll over and die? Do you really think intel is so inept as to not be able to do exactly what you say the other intel killers are going to do?

Thats a lot of questions
Anyway, why intel is in trouble, is not because they have inept engineers. Its because they are a (virtual) monopoly that is about to face competition from a competitor that is larger than them, and that has no monopoly margins to protect. That is the crux of their problem. Its not an engineering one, nor is it one engineers can solve.

Mobile phone chip vendors will over the next years see an opportunity to upsell SoCs that cost around $5 to produce and that they now sell for $10-15, in to a market where ASPs are typically 10x or 50x higher. What do they have to lose ? Nothing. If they undercut intel by 50% per performance unit, they are still printing money.

Now even assuming intel can produce atom or xeon chips for the same cost per performance unit as the ARM vendors, its still in deep doodoo. Intel has huge margins on their desktop chips and even huger on xeons. But they need those margins, because they are heavily overweight and because they do their all their own R&D. Their xeons are subsidizing their atom and larabee attempts. For ARM its the other way around. Billions of mobile phones are paying the R&D and core development that they can tweak for desktop or server chips.

Really, its the same advantage that got intel in to servers in to the first place. 20, 30 years ago they made a killing selling desktop chips generating far higher unit volumes than the RISC guys, and that allowed them to sell slightly tweaked desktop cpu's in to workstations and later servers for fat margins. Remember the Pentium Pro?

Thats what killed 90% of the risc server market and gave it all to intel: high volume consumer chips who's R&D was already paid for, that got tweaked a bit to compete in the workstation and server markets. It wasnt that intel chips where faster, they werent. They were slower, sometimes a lot slower, but far cheaper and umbiquiteous. Over time intel chips did get faster because the RISC vendors couldnt spend nearly as much as intel on R&D, but that only happened after intel had conquered most of the server market already anyway.

Its not that the guys working on Vax, Alpha, PA-Risc, MIPS or the others where incompetent. They where every bit as competent as their collegues at Intel, if not more so. Its not that their ISA was worse than x86. It was better in fact. But their business model wasnt.

And now, its pretty much the exact same again, only its ARMs turn to tweak ultra high volume consumer chips to make them useable in desktops and servers. Whether its nVidia that will prove this or someone else remains to be seen. nVidia is the first one trying, but they wont be the last.

 

[P4man]

Oh btw, some intriguing evidence of Apple working on a full custom ARM chip for their Mac line here:
http://www.semiaccurate.com/2010/12/29/evidence-points-apple-designing-arm-laptops/

Have to agree with their analysis too, not something that happens very often with SA articles .

 

[HeXen]

whoooaahh!! i came in here and got smacked in the face with a Steven King Novel. think i'll just go and play some NFS: Shift or something.

 

[cbn]

But I wouldnt even assume project Denver turns out to be slower than typical Xeon or opterons.

http://www.dailytech.com/Microsofts+Ballmer+ARM++Theres+a+Windows+PC+OS+for+That/article20580.htm

ARM features a reduced-instruction set, versus Intel's cluttered instruction catalog. And it has more integers registers, which eliminates the expensive process of renaming registers. The net result of both of these architectural differences is that ARM can perform the same computation using less power.

Just wondering how that efficiency gain would translate into keeping heat under control as the xtor gates get closer together at smaller nodes? Could this translate into faster single thread performance and better single thread performance per watt?

 

[namtran512]

It's like trying to compete versus Microsoft in the OS business. Your best efforts may make a dent, but Microsoft would still rule supreme.


Intel / x86 aren't disappearing for a very long time. It would take at least a minimum of 10 years for ARM to become even close to "mainstream".

 

[Voo]

And it [ARM] has more integers registers, which eliminates the expensive process of renaming registers

Wait what? While having more registers is nice, register renaming as part of an OoO architecture offers other advantages that "more registers" don't have
And the whole premise of ARM having more registers only works if you don't count the additional registers you get from the SSE extensions (why should those not count as "integer registers"?)

 

[Absolution75]

Wait what? While having more registers is nice, register renaming as part of an OoO architecture offers other advantages that "more registers" don't have
And the whole premise of ARM having more registers only works if you don't count the additional registers you get from the SSE extensions (why should those not count as "integer registers"?)

They aren't general purpose registers and have specific uses - not really the same thing at all. You can't execute x86 instructions on them, they have their own set of instructions (not to mention you can really only rely on the fact that a user will have SSE2 or lower).

 

[jihe]

x86 5 years on: fudzilla is history, bad history

http://www.fudzilla.com/mobiles/item/21438-nvidia-believes-that-x86-is-history



:biggrin:

OK it's mostly humorous. But I wonder if we could finally be seeing the beginning of the end for x86? It's been around a long time and many say that its starting to show its age. GPUs are starting to take over the super computer realm and simpler chips are eating away the low end. x86 is stuck in the middle and that middle could disappear. Intel is the only company actually making a real profit on x86, so it's not nearly as widespread as ARM.

Granted we are talking a decade at least for a total collapse, but it could happen.