The ARM v.s. Intel Thing - Let's Discuss It

[Cerb]

Are you surprised after how it is with everything the last 5-10 years? Remember when Core 2 was "overclocked Opterons" for example?

ARM makers are already canibalizing in their own ranks. And its not funny when the gorilla gets into the chimps camp to wreck further havok.

I'm surprised Renesas and MediaTek are where they are. I would have figured MediaTek wouldn't even be on the list, and Renesas would be much lower.

IMO, Intel hurt themselves quite a bit relaunching various mobile Linux projects while Android and iOS were grabbing tons of developer mindshare - Moblin > Meego > Tizen.

IMO, what really hurt them was not standing behind Meego. Moblin to Meego was more or less a nice name change and some consolidation, but they never gave it more backing than as a dev toy, then Nokia went schizo, and now we're here. Even Tizen they pretty much just support in the sense that it runs on their hardware.

 

[IntelUser2000]

they never gave it more backing than as a dev toy, then Nokia went schizo, and now we're here. Even Tizen they pretty much just support in the sense that it runs on their hardware.

Yea but how is that a bad thing? I've watched Intel for years, and one thing they have hard time is pushing something outside of CPUs, and especially software.

Back then when Nokia had support, MeeGo looked promising. Weeks before Nokia ran off on them, MeeGo looked like it was on its deathbed.

Android has application support, so they don't need to start from the ground up. Perhaps if Nokia hadn't abandoned MeeGo, it might have gone somewhere, perhaps its just a dream that was never meant to be.

IMO, their mistake was not pursuing Android from the beginning. They only started executing and getting design wins when they got Google support.

 

[Cerb]

Yea but how is that a bad thing?

They had excellent products, and Meego was pretty neat. Everyone in their right mind knew that Windows was a bad direction to go, at least as a sole option for the smart phones and such. I think a little more OS competition would be good, and Windows doesn't really present much, yet.

IMO, their mistake was not pursuing Android from the beginning. They only started executing and getting design wins when they got Google support.

How would they know Android was going to be the one? Even now that it is, they're largely working for ARM parity, and have more of an uphill battle than if they had pushed another platform, earlier.

Intel needs to find to find an Apple-like way to act, because they just don't have enough going for them, in mobile, to compete on an even playing field, with their historical margins. They have needed to, and still need to, stack the deck (without running afoul of antitrust laws).

 

[shady28]

Supposedly one of the reasons otellini was fired was his x86 or nothing approach, which resulted in intel missing the boat in the mobile space.

Intel could (and has) license ARM. Remember XScale?

They seem to be trying to use Atom right now though, which I think is not the best strategy. Mobile is already established on arm, people wont switch just for speed.

 

[Arachnotronic]

Supposedly one of the reasons otellini was fired was his x86 or nothing approach, which resulted in intel missing the boat in the mobile space.

Intel could (and has) license ARM. Remember XScale?

They seem to be trying to use Atom right now though, which I think is not the best strategy. Mobile is already established on arm, people wont switch just for speed.

Otellini was not "fired". He left of his own accord and was even urged by the board of directors to stay.

Yes, I remember XScale. It was sold to Marvell which has been wholly unsuccessful in making any real money with it.

 

[Fjodor2001]

What about CPU frequency? It seems like Intel has hit a wall at around 4 GHz. Does that limit also apply to ARM CPUs?

I know ARM has reached 2.5 GHz with the Cortex A-15. But if they will hit a wall at 4 GHz too, then there's not much room for improvement coming from higher frequency.

And since the ARM CPUs are less "efficient" per clock cycle than x86 (i.e. they "get less work done" per clock cycle), can ARM ever reach the performance level of Intel CPUs?

 

[ShintaiDK]

I think ARM has some pretty good business allies though. Google, Apple, Qualcomm, Samsung it's not like Intel is pushing into virgin territory.

IMO, Intel hurt themselves quite a bit relaunching various mobile Linux projects while Android and iOS were grabbing tons of developer mindshare - Moblin > Meego > Tizen.

Google inst an ally. They worked heavily on the x86 Android. Google is only interested in selling Android, not ARM.

 

[ShintaiDK]

What about CPU frequency? It seems like Intel has hit a wall at around 4 GHz. Does that limit also apply to ARM CPUs?

I know ARM has reached 2.5 GHz with the Cortex A-15. But if they will hit a wall at 4 GHz too, then there's not much room for improvement coming from higher frequency.

And since the ARM CPUs are less "efficient" per clock cycle than x86 (i.e. they "get less work done" per clock cycle), can ARM ever reach the performance level of Intel CPUs?

Frequency wall hits everyone. Its not really design related.

ARM makers cant afford the cost to make them reach the performance level. If the ISA even allows it by design.

 

[Roland00Address]

What about CPU frequency? It seems like Intel has hit a wall at around 4 GHz. Does that limit also apply to ARM CPUs?

Intel has not hit a wall at 4 ghz, they are just getting less efficency past 4 ghz and thus they do not want to cross it.

Intel has this philosophy with die space but it could be apply to cpu speed as well. Do not add a feature that will add less than 2 percent speed increase for every 1 percent higher power consumption. After 4 ghz you are not getting the proper scaling with speed to power consumption. To go from 4 ghz to 5 ghz you are pretty much doubling the power consumption while only gaining 25% increase in speed (assuming the program scales perfectly to mhz and has no other bottlenecks which is not always the case.)

Someone has a nice chart they occasionally post here showing sandy vs ivybridge scaling in ghz vs power consumption. This charts illustrates this point perfectly.

---------

Intel as a current philosophy is not trying to scale up features wise but instead scale down for less power consumption. They want an i7 920 (2008) speed which has a 130w tdp and no integrated graphics to fit into something like a tablet. Currently their best 17w tdp i7 dual core is 3667U which is about 80% the speed of the i7 920 quad core when comparing multithreaded benchmarks. They want to scale down this chip and get this performance in a 10w envelope and eventually a 5w envelope.

---------

If Intel needs more performance for desktops they can always just add cores. Currently they have not done this for no one is competing with them in desktops and thus they haven't had to make 6 or 8 core processors for desktops if you discount 2011 socket workstations or xeon servers.

 

[Fjodor2001]

Intel has not hit a wall at 4 ghz, they are just getting less efficency past 4 ghz and thus they do not want to cross it.

...which was what I meant by Intel hitting a frequency wall.

The question is if ARM will hit that wall to at around 4 GHz. Being a RISC CPU it may be able to reach higher frequencies than the Intel CISC CPUs.

 

[ShintaiDK]

...which was what I meant by Intel hitting a frequency wall.

The question is if ARM will hit that wall to at around 4 GHz. Being a RISC CPU it may be able to reach higher frequencies than the Intel CISC CPUs.

The last x86 CISC CPU was the Pentium MMX.

 

[Chiropteran]

Besides Samsung and Apple. I dont think any of the rest got any control, they could get exactly the same as they get today from Intel so to speak.

Still, even if they aren't making the CPU's in house, they have the option to have them made in several foundries, right? With Intel CPU, you get them from Intel, period. With ARM there is some competitive pressure, with Intel once you get locked in you are largely forced to go along with anything Intel decides to do, or spend lots of money switching to ARM.

 

[ShintaiDK]

Still, even if they aren't making the CPU's in house, they have the option to have them made in several foundries, right? With Intel CPU, you get them from Intel, period. With ARM there is some competitive pressure, with Intel once you get locked in you are largely forced to go along with anything Intel decides to do, or spend lots of money switching to ARM.

They simply buy from whoever makes them. Its off the shelves components.

People should also consider why my Samsung S3 Mini for example dont have a Samsung CPU in it If there is something better or more suited, they will use that.

 

[Charles Kozierok]

The last x86 CISC CPU was the Pentium MMX.

Eh, you're getting into a bit of a nomenclature argument there. In terms of the ISA, all of x86 is still CISC.

In terms of microarchitecture, one can argue that internally x86 is "RISC-like", though even there it gets dicey (I believe Atom only breaks down some operations into micro-ops, for example).

 

[Idontcare]

The last x86 CISC CPU was the Pentium MMX.

Eh, you're getting into a bit of a nomenclature argument there. In terms of the ISA, all of x86 is still CISC.

In terms of microarchitecture, one can argue that internally x86 is "RISC-like", though even there it gets dicey (I believe Atom only breaks down some operations into micro-ops, for example).

CISC vs. RISC semantics aside, the reason it was brought up was the notion that the ISA plays a role in frequency envelopes:

...which was what I meant by Intel hitting a frequency wall.

The question is if ARM will hit that wall to at around 4 GHz. Being a RISC CPU it may be able to reach higher frequencies than the Intel CISC CPUs.

The point that needs to be made here is that it is the microarchitecture, not the ISA, that determines the frequency envelope and power envelope for a given IC.

ISA impacts performance/watt and IPC, but doesn't have to impact frequency or power unless the designers elect to make it so.

 

[ShintaiDK]

Well we could do 250Ghz chips, my watch might be faster in pure performance tho

 

[Charles Kozierok]

My guess is that the need to decode x86 into micro-ops does have some impact on maximum frequency, but Intel's gotten so good at it by this point that I bet the impact is relatively small.

 

[Fjodor2001]

Well we could do 250Ghz chips, my watch might be faster in pure performance tho

Not if running ARM code though, which is what we're discussing...

 

[TuxDave]

Not if running ARM code though, which is what we're discussing...

Please elaborate.

 

[Fjodor2001]

Please elaborate.

We cannot produce 250 GHz chips running ARM code. Also, if we actually could do that, it would be a hell of a lot faster than ShintaiDK's watch. Unless he's got some secret quantum computer chip in his watch that is... :sneaky:

 

[ShintaiDK]

250Ghz is only reached with a handful of transistors at a time. ARM requiers around 12000 transistors for the M0 for example.

 

[Paul98]

If intel just continues what it's doing and as their process lead continues to grow. Seems in a few years they won't have to do much at all to dominate the mobile market.

 

[TuxDave]

250Ghz is only reached with a handful of transistors at a time. ARM requiers around 12000 transistors for the M0 for example.

I guess 4ps fo4 delay is a bit of a stretch. Not sure if you can build any chip of any ISA at that point. I guess my question is assuming something doable but not desirable. I don't actually know what an arm license entails. Can you build whatever you want as long as it is arm comparable or is the design dictated by the license. Because if you could do whatever you want, 10ghz chip can be done but it would be very limited in non ISA scope.

But I think the comment was assuming at LEAST one gate per cycle (so something like 10-20GHz). At that point, it would take a good amount of creativity to get a chip to work and you PROBABLY could get it running ARM code. You couldn't get anything done in a single cycle. Decoding would be broken up and several cycles long. Execution units would be several cycles long. If I were building some garbage chip just to get it to clock high, I probably would allow only 1op done the pipe at a time and nothing else until it completes. At that point, the watch would probably be faster.

 

[Fx1]

Apparently Intel conroed the ssd market? Didn't you hear? Anand said so in his typical fanboy style. Oh wait. How long did that last? 2 months?

Then they went sand force after admitting defeat to OCZ of all companies.

Now Samsung kicks their ass.

Sound familiar?

Mobile cpu is not the laptop or desktop market.

ARM doesn't have any big fabs to keep fed in hard times. It also has very reliant partners who are heavily invested in ARM licences like Apple and qualcom and Samsung.

Ask yourself who gives a crap about Intel? Last time I checked the only people who give a crap about Intel were Microsoft and look what they just did RT?

 

[blackened23]

If intel just continues what it's doing and as their process lead continues to grow. Seems in a few years they won't have to do much at all to dominate the mobile market.

I agree with this. R+D wise they're so far ahead of everyone, it is inevitable that intel will be one of the top players in coming years.

The AT article on ARM vs clover trail power consumption was an eye opener as well. Intel is very close on that (power consumption) front, and way way ahead on the performance front.