The Intel Atom Thread

[erunion]

Just a theory based on them pushing SoFIA for tablets- could definitely be wrong

Of course you're right. Baytrail wasn't designed to go into $100-200, its too expensive for those devices. That is why we have contrarevenue.

dont undestand this. how are they regressing to a 28nm node for entry and value while in 2013 their value and entry tablets all had lower specced baytrail derivatives in them. smh

furthermore what is the need for an integrated 3g modem for entry and value tablets?

SOFIA isn't actually a replacement for baytrail, cherrytrail is. SOFIA replaces the Infineon modems that power cheapo feature phones, and pseudo-smartphones like the Nokia Asha 500.

Ultra budget tablets will likely transition to SOFIA, and away from medfield and merrifield.

 

[Idontcare]

also why the hell has it taken so long to get intel baseband inhouse. i just dont get it. how can they be spending so much in mobile with so little to show for it.

Maybe Idontcore could give some insight in how long it takes to port such things to another process?

It's not trivial, that is for sure.

But Intel's particularly sluggish response to this area is more likely driven by the fact you can't put your A-team (be it your best project managers or your best engineers) on every single project you have going on across the entire company. Some projects are destined to have the C-team.

Projects get prioritized, and the higher priorities get the resources. And those resources aren't merely financial, they are also internal support services and the quality of resources you are allowed to secure for your program.

In a big company like Intel, pretty easy to have projects floating around that just sort of muddle along with an impressively long timeline for deliverables. Broadcom did the same thing.

Intel knows how to get things done, but they (just like everyone else) don't know how to get everything done all at once

 

[liahos1]

thank you for your insight Idontcare. Digitimes is reporting intel is about to invest 1.5billion for a 20% stake in Spreadtrum. What you guys make of that?

 

[witeken]

About the 64-bit discussion from a few pages ago, some people are speculating that the Desire 510 doesn't even have a 64-bit SoC, but 4 A7 cores instead.

If you look at Wikipedia, the only SKU mentioned runs at 1.4GHz instead of 1.2GHz. In the press release, it also doesn't mention 64-bit at all.

 

[krumme]

It's not trivial, that is for sure.

But Intel's particularly sluggish response to this area is more likely driven by the fact you can't put your A-team (be it your best project managers or your best engineers) on every single project you have going on across the entire company. Some projects are destined to have the C-team.

Projects get prioritized, and the higher priorities get the resources. And those resources aren't merely financial, they are also internal support services and the quality of resources you are allowed to secure for your program.

In a big company like Intel, pretty easy to have projects floating around that just sort of muddle along with an impressively long timeline for deliverables. Broadcom did the same thing.

Intel knows how to get things done, but they (just like everyone else) don't know how to get everything done all at once

I dont think thats quite the right explanation. (and yes its not trivial - lol)

We have seen Intel using lots of money here for at least 3 years. Mobile is losing big time. They give away their solutions for practically nothing. There is plenty of money. That usually means allocation of ressources is there big time. At least now, and must have been for 2-3 years.

We have to note the entire solution is bad, besides baseband its dsp and probably many other parts.

I think the reason is the most straight forward. Intel have not in time developed the nessesary comptences, and hired the new people for what was needed. Therefore they can not develop the technology. Mastering new technology on the highest level - baseband and dsp solutions is that for Intel - is always a pain. They simply needed enough men with the specific competences - but they were not there. You cant move A person or even A person project manager to a new field within a year. He will be a D performer then.

Now there might be a culture of "we are designing cpu" or "we are a manufacturing company with best process competences" and they probably used the "C team" so to speak as a start, but for me, a blunder like what we have seen is a top management responsibility.

Getting that started and have full effect usually takes 2-3 years from you hire the guys (and its not small numbers here ). Then you can start for real. That probably also means we will see some fast acceleration from Intel here during 2015, 2016.

But they are competing with QQ that (my guess) probably have the same market cap or so, and guys like SS, that is just a fat moving train that prefer vertical control.

 

[Roland00Address]

do any of you think we'll find out anything at all about Sofia, Broxton or Cherrytrail at IDF? I mean prior to IDF last year they had ANAND and a bunch of other dudes in their offices checking out baytrail. its depressing. If core M is an amazing product but the devices they get put in are $600+ plus its depressing. Although I guess the value one gets out of baytrail tablets now is ridiculous for the price. I havent really seen too many new arm based tablet skus popping up recently. seems like a massive blur now and all the releases seem to be intel based. maybe im just looking at it through blinders. /rant

If I recall IDF happened the same week that Anandtech posted their baytrail preview. Thus we are talking the 2nd week of september we aren't there yet.

Furthermore Core M is effectively the same architecture as before but just in new form factor. How it turns out with final devices matter a lot.

Baytrail by contrast was a new architecture so the rough performance as important and reference design tablets were okay, with Core M the final products should be what it is judged by.

dont undestand this. how are they regressing to a 28nm node for entry and value while in 2013 their value and entry tablets all had lower specced baytrail derivatives in them. smh

furthermore what is the need for an integrated 3g modem for entry and value tablets?

Phones especially value phones and not the top end phones such as the Galaxy S line really really want an integrated package with everything to be built together to keep down costs. Intel bought their modem company (infineon) instead of developing it in house, this modem company were mostly developing on tmsc.

Furthermore the modem part is one of the hardest parts to make smaller and it is a part that does not scale well due to the analog parts of the modem. Thus it is harder to port to intel's fabs. Intel is keep on saying it is going to happen but the what it is actually happening is "not yet." We were supposed to get an integrated modem on intel 22nm. By contrast we see Intel porting its 22nm designed silvermont architecture to tmsc 28nm, and then going ahead with the 14nm integrated everything design.

Lets see if this changes again in a year or two.

 

[TreVader]

Bay Trail has what it takes in pure MT/ST performance (mostly) but it's GPU is atrocious for a mid 2014 SoC. If I wanted a tablet to do encoding (ha) , bay trail all the way. Anything else; Apple A7 or S800.

I really like the idea of intel dropping Atom like Itannic and focusing on Core M. If I can get core performance I a fabless tablet, why even bother with Atom unless it's for a $79 kids toy.


Really hoping for a Core M 12" rMBP. That would renew my faith in intel.

 

[Roland00Address]

Bay Trail has what it takes in pure MT/ST performance (mostly) but it's GPU is atrocious for a mid 2014 SoC. If I wanted a tablet to do encoding (ha) , bay trail all the way. Anything else; Apple A7 or S800.

I find this ironic for while I will always want better gpu performance, I think better storage performance and better temporary cpu performance (aka burst/turbo) would suit my usage better with the atom devices I have interacted with. That and better battery life aka idle.

Until they increase the resolutions of the atom tablets (which means android or fixing windows horrible resolution scaling, what chrome just now fixed the high resolution mode this week?) or that they release games for atom tablets that need more gpu power I am satisfied with the intel gpu in the current atoms.

That said I am salivating at the thought of Tegra K1 and Intells Cherry Trail. Something faster than a desktop kepler gt635 but about the same performance of a gt640 (640 has higher gigaflops but it was bandwidth limited so in real world scenarios between those two) in a 5 watt soc that is amazing.

I really like the idea of intel dropping Atom like Itannic and focusing on Core M. If I can get core performance I a fabless tablet, why even bother with Atom unless it's for a $79 kids toy.

Do not doubt the power of the $79 toys, those things cause fundamental changes you do not realize till later. Sure you may only have 1 or 2 $1000+ toys/work items but you buy so many $80 dollar things with you realizing it and now you have a computer sub $80 that has more cpu or gpu power than anything out there on the consumer market 10 years ago.

Just imagine the people at pixar and what they would do to get a couple atom computers from 1993 to 1995.

 

[TreVader]

I find this ironic for while I will always want better gpu performance, I think better storage performance and better temporary cpu performance (aka burst/turbo) would suit my usage better with the atom devices I have interacted with. That and better battery life aka idle.



Until they increase the resolutions of the atom tablets (which means android or fixing windows horrible resolution scaling, what chrome just now fixed the high resolution mode this week?) or that they release games for atom tablets that need more gpu power I am satisfied with the intel gpu in the current atoms.



That said I am salivating at the thought of Tegra K1 and Intells Cherry Trail. Something faster than a desktop kepler gt635 but about the same performance of a gt640 (640 has higher gigaflops but it was bandwidth limited so in real world scenarios between those two) in a 5 watt soc that is amazing.







Do not doubt the power of the $79 toys, those things cause fundamental changes you do not realize till later. Sure you may only have 1 or 2 $1000+ toys/work items but you buy so many $80 dollar things with you realizing it and now you have a computer sub $80 that has more cpu or gpu power than anything out there on the consumer market 10 years ago.



Just imagine the people at pixar and what they would do to get a couple atom computers from 1993 to 1995.

The problem is atom devices are targeted and more than just you. They are actually pitching it as a "gaming" tablet which is just plain lying. Atom just is an sku that need not exist. Lowest common denominator is not a way to make profit, ask Apple.

 

[jdubs03]

The problem is atom devices are targeted and more than just you. They are actually pitching it as a "gaming" tablet which is just plain lying. Atom just is an sku that need not exist. Lowest common denominator is not a way to make profit, ask Apple.

The gaming aspect is exactly why I believe/ed that Cherry Trail could be a difference maker. Who knows whether that can be the case. But the reason why Atom exists is due to Intel wanting to covering the low-end segment without sacrificing their higher-end margins. With Core M, they still will be focusing on that high end.

Having Core M is great, but what if Tegra Denver outperforms it at 28nm as well as the A8 at 20nm? Atom surely cannot compete with that if Core M cannot. Intel is betting big with Core M+14nm; and if those two competitors out-benchmark it, it will be very worrying, mostly on the Android side of things. (And normalizing the performance for process shrinks makes it more dire)

 

[witeken]

Having Core M is great, but what if Tegra Denver outperforms it at 28nm as well as the A8 at 20nm? Atom surely cannot compete with that if Core M cannot. Intel is betting big with Core M+14nm; and if those two competitors out-benchmark it, it will be very worrying, mostly on the Android side of things. (And normalizing the performance for process shrinks makes it more dire)

The turbo and IPC of Core M are way too high to be a threat for any SoC at a 28nm or 20nm process. You can't get around physics.

 

[VirtualLarry]

The turbo and IPC of Core M are way too high to be a threat for any SoC at a 28nm or 20nm process. You can't get around physics.

I know that you are a great believer in Intel's Fab prowess, but consider:

Intel's design teams, have been able to extract only minor percents of performance increases for the last three generations of CPU.

Nvidia's design teams, have been able to leapfrog their performance/watt numbers, for their GPUs, on the "same" silicon process.

Great design (or lousy design) can make or break a product, just as much as the fab tech can.

 

[witeken]

Using GPUs to proof something about CPUs?

I don't think Core is a bad design. Its ability to scale from 2 to 140W is a testament of this.

 

[ShintaiDK]

I know that you are a great believer in Intel's Fab prowess, but consider:

Intel's design teams, have been able to extract only minor percents of performance increases for the last three generations of CPU.

Nvidia's design teams, have been able to leapfrog their performance/watt numbers, for their GPUs, on the "same" silicon process.

Great design (or lousy design) can make or break a product, just as much as the fab tech can.

But Maxwell is the exception, not the norm for GPU makers. And Maxwell increases the transistor count. GM107=148mm2 and 1870M. GK107=118mm2 and 1300M. A 44% increase in transistor count better yield something.

Without shrinks, you could actually get slower performance on some GPU uarchs.

And CPUs dont have anymore low hanging fruits. There is only so much a better design can do without increasing transistor count.

 

[jdubs03]

The turbo and IPC of Core M are way too high to be a threat for any SoC at a 28nm or 20nm process. You can't get around physics.

Turbo is a very neat feature that neither of the two SoCs I mentioned, but nonetheless what we may see out of the A8 could be quite impressive +50% clocks will help out immensely and that doesn't include any uArch tweaks (which I expect to be minor if any). Nvidia has seemed to deliver with Denver even at 28nm, and I hope the Nexus 9 performs well, as that would rightfully validate the choice to go dual-core.

Its hard to argue against the solutions being put forth by Intel's competitors even at process disadvantages.

And CPUs dont have anymore low hanging fruits. There is only so much a better design can do without increasing transistor count.

Indeed, it will be harder to manage such large improvements, but that isn't to say that Intel can maintain its absolute performance advantage. The ARM vendors, at least the ones I mentioned above have pretty much caught up, and may slightly surpass, at least for <5W TDP solutions.

 

[ShintaiDK]

Indeed, it will be harder to manage such large improvements, but that isn't to say that Intel can maintain its absolute performance advantage. The ARM vendors, at least the ones I mentioned above have pretty much caught up, and may slightly surpass, at least for <5W TDP solutions.

Its clear on the 64bit side, that ARM makers are out of low hanging fruits. Even Apples dualcore is 5W.

 

[witeken]

Turbo is a very neat feature that neither of the two SoCs I mentioned, but nonetheless what we may see out of the A8 could be quite impressive +50% clocks will help out immensely and that doesn't include any uArch tweaks (which I expect to be minor if any). Nvidia has seemed to deliver with Denver even at 28nm, and I hope the Nexus 9 performs well, as that would rightfully validate the choice to go dual-core.

Its hard to argue against the solutions being put forth by Intel's competitors even at process disadvantages.

Do you realize that the 3 SoCs you mention are not yet available in products and thus are also not yet reviewed?

I don't think Apple will (be able to) scale the frequency much. I think they will mainly focus on reducing power consumption.

There's only so much you can do when you are >2 nodes behind in terms of power consumption.

Indeed, it will be harder to manage such large improvements, but that isn't to say that Intel can maintain its absolute performance advantage. The ARM vendors, at least the ones I mentioned above have pretty much caught up, and may slightly surpass, at least for <5W TDP solutions.

Nvidia hasn't really caught up, since its IPC is still significantly lower.

 

[jdubs03]

Do you realize that the 3 SoCs you mention are not yet available in products and thus are also not yet reviewed?

I don't think Apple will (be able to) scale the frequency much. I think they will mainly focus on reducing power consumption.

There's only so much you can do when you are >2 nodes behind in terms of power consumption.


Nvidia hasn't really caught up, since its IPC is still significantly lower.

I'm going off the rumor of the A8 being 2.0 Ghz. ~50% faster.

We'll see what happens. But, I think you're going to be disappointed.

 

[NTMBK]

Using GPUs to proof something about CPUs?

I don't think Core is a bad design. Its ability to scale from 2 to 140W is a testament of this.

There is no Core design with a 2W TDP.

 

[witeken]

I'm going off the rumor of the A8 being 2.0 Ghz. ~50% faster.

We'll see what happens. But, I think you're going to be disappointed.

I doubt it will be able to turbo long to 2GHz. 20nm is going to help somewhat, but they will need to do some clever tricks to reduce power consumption.

Why'd I be disappointed?

There is no Core design with a 2W TDP.

Because the CPU isn't the only part of the Broadwell-Y SoC. If you're gaming, the GPU will easily take half the power budget. Core M isn't meant for low-end and low-power devices either, so it wouldn't make sense not to let it use a full tablet TDP.

 

[Sweepr]

About that BioShock A6/A7 iOS vs Bay Trail-T Windows x86 discussion a few pages back, here's a first look from AnandTech: www.anandtech.com/show/8440/a-look-at-bioshock-for-ios

 

[liahos1]

The gaming aspect is exactly why I believe/ed that Cherry Trail could be a difference maker. Who knows whether that can be the case. But the reason why Atom exists is due to Intel wanting to covering the low-end segment without sacrificing their higher-end margins. With Core M, they still will be focusing on that high end.

Having Core M is great, but what if Tegra Denver outperforms it at 28nm as well as the A8 at 20nm? Atom surely cannot compete with that if Core M cannot. Intel is betting big with Core M+14nm; and if those two competitors out-benchmark it, it will be very worrying, mostly on the Android side of things. (And normalizing the performance for process shrinks makes it more dire)

if Core M is getting something approaching haswell y i5 levels of performance it's already defeated denver in the benchmark wars and unless apple magically increases performance by multiples of the a7 it seems that core m wins. now just need to see battery life of devices that core m is in

edit -

Frequency Memory TDP
Core M 5Y10 2 / 2 800 MHz / 2 GHz 4 MB 100 / 800 MHz DDR3L/LPDDR3-1600 4.5W
Core M 5Y10a 2 / 2 800 MHz / 2 GHz 4 MB 100 / 800 MHz DDR3L/LPDDR3-1600 4.5W
Core M 5Y70 2 / 4 1.1 / 2.6 GHz 4 MB 100 / 850 MHz DDR3L/LPDDR3-1600 4.5W

i think the haswell i5 in the surface pro 3 turbos to 2.9ghz. the 5Y70 sku purportedly turbos to 2.6ghz and also should have a 5% ipc increase relative to haswell.

for comparison to the a7 see below. i find it hard to believe a 28nm k1 or even a a8 at 20nm would get in spitting distance of performance or more importantly perf/watt

http://www.anandtech.com/show/8287/first-look-the-799-microsoft-surface-pro-3-with-core-i3

now we just need intel to get this in devices at reasonable prices

 

[liahos1]

About that BioShock A6/A7 iOS vs Bay Trail-T Windows x86 discussion a few pages back, here's a first look from AnandTech: www.anandtech.com/show/8440/a-look-at-bioshock-for-ios

ios version looks like its running on a dreamcast. its horrible

 

[witeken]

It's not trivial, that is for sure.

But Intel's particularly sluggish response to this area is more likely driven by the fact you can't put your A-team (be it your best project managers or your best engineers) on every single project you have going on across the entire company. Some projects are destined to have the C-team.

Projects get prioritized, and the higher priorities get the resources. And those resources aren't merely financial, they are also internal support services and the quality of resources you are allowed to secure for your program.

In a big company like Intel, pretty easy to have projects floating around that just sort of muddle along with an impressively long timeline for deliverables. Broadcom did the same thing.

Intel knows how to get things done, but they (just like everyone else) don't know how to get everything done all at once

Intel acquired Infineon in 2010. Something went very wrong if it takes more than 5 years to get the Infineon IP in-house.

 

[jpiniero]

ios version looks like its running on a dreamcast. its horrible

Huh, I compared it to the (very poor) port of Bioshock 1 on the PS3 that was a throw-in with Infinite. Even though it's quite ugly, it still seems like it is better than those screenshots.

OTOH Infinite on the PS3 is OK graphically for what it is.