[PCWorld] Intel kills the Atom line

[frozentundra123456]

More open than iOS or even Android. (specifically in practical terms for end-user, I am aware it is more open for other user scenarios)

Althrough I admit MS is certainly working to close their ecosystem with Windows 10.

Yep, I agree. One of the best features of windows has been the wealth of built-in programs and the availability of programs from other sources. Now, almost everything you want to do directs you to the Window's store.

For instance, I wanted to download Nook to my win 10 tablet to use it as a reader. Previously, I just went to the Barnes and Noble site and downloaded it. Now, have to do it through the Windows store.

 

[Sheep221]

Open x86 Wintel , the funniest oxymoron I ever read :biggrin:

Which platforms are more open than x86 linux and windows? Please list them here so all of us can finally move on from these evil x86 computers and their communist operating systems!

 

[Nothingness]

Which platforms are more open than x86 linux and windows? Please list them here so all of us can finally move on from these evil x86 computers and their communist operating systems!

Linux is open (and runs on other CPU than x86), Windows is certainly not open.

 

[videogames101]

Which platforms are more open than x86 linux and windows? Please list them here so all of us can finally move on from these evil x86 computers and their communist operating systems!

Linux runs on like every ISA ever.

 

[cytg111]

Which platforms are more open than x86 linux and windows? Please list them here so all of us can finally move on from these evil x86 computers and their communist operating systems!

Windows? Open? Open to malware and stealing your moms maiden name perhaps, other than that, open it is NOT. Well, it may be open to opensourcing all your privates to 1st 2nd and 3rd parties to ram advertising into every ori... Damn, sorry, snapped again. Like Intel, microsoft has reinvented it self as of late with going the whole telemetry/big-data as a business model route. If you ever pondered about linux the time to switch is now. Open source to protect your privates, noone else is doing it for you.

 

[PPB]

Atom was dead on their feet the moment it was arbitrarily relegated in favor of the Core line. I have a friend with a 2 in 1 with an Z3740 (1.33ghz base, 4 cores). After so much babbling about SMT's performance gain at a minimal TDP and transistor cost, I cant think of a reason of such bizarre product stack of 4 core Atom products with those awfully low clockspeeds. Performance on that convertible is barely acceptable and besides the better platform and lower power consumption, performance hasnt gone upwards compared to my almost 9 year old, t5300 14'' lappy.

Considering it's low IPC, a 2C/4T atom die with clockspeeds north of 2ghz base would give a totally better experience for the single threaded still heavy, consumer oriented workloads. Its like Intel with their Atom line behaves like a totally different company (moar coars anyone?), when in reality both Core and Atom should aim for similar goals withing their capabilites in the cost-die size- power envelope areas. The IPC difference alone will relegate Atom sku's to lower tiers of mobile products, with their respective lower costs to make up for the performance deficiency. Now if you castrate your mobile core to the extent of even underperform ARM variants that have their platform cost advantage compared to Atom, it is just a silly game.

TL;DR: Give Atom some love, maybe cut to an end that silly thing of having 2 design teams devoted to Core and compete-contribute to each other only to make Core line 5% more performant each 1.5years and make one of the fully committed to atom, the other one fully committed to Core line, and make useful patent-grade technologies developed by either team to be considered for the other's team future uarch designs.

 

[IntelUser2000]

PPB: Atom Z3740 has a Burst Mode Technology clock of 1.86GHz. And it commonly reaches that.

Core M 6Y75's Base clock is at 1.2GHz, yet it performs far higher.

Atom performs less because the IPC is far lower than Core chips. Also, rest of the platform is slower, like memory subsystem and storage.

Atom needs 2.5x performance it reaches now to be a leadership chip in mobile. However, that's exactly Core M's performance. Also, part of the reason they moved to 2 teams with Tick/Tock is because that turned out to be better for them.

For all the discussions we have, it unfortunately won't make a difference. I doubt it was easy for Intel either. The problem lies more deeply. They'd have to do something very substantial to succeed. I mean like accepting margins to decrease from 60% to 50%. Cut Core M prices to $50-70. Implement eDRAM across their lines. At least have it for top of the line Atom, and all Core ix chips, including Core i3. At same or lower price.

 

[PPB]

PPB: Atom Z3740 has a Burst Mode Technology clock of 1.86GHz. And it commonly reaches that.

Even if it can sustain turbo most of the time 1.86ghz is Core Pentium U territory. That is just low clockspeeds for . You see over the fence and the ARM competition is already taking a dive in 2ghz+ clockspeeds. And that is in similar or lower power envelopes, and higher IPC than atom.

I sustain my view that in a normal world Core and Atom wouldnt overlap if Intel gave more resources to Atom to pump up things a little. It is baffling that with Intel's wide IP in CPU design, and it's expertise in designing the Core uarch Atom doesnt any collateral benefits. A team of 80 engineers in an outsorced office gave AMD their Cat uarch, which to the disdain of some here it's succesor ships in similar volumes WITHOUT any contra-revenue (it is sad to think AMD made more money out of Cat derived products than Intel did with it's Atom line).

Someone will tell me Intel cant achieve more performance in their Atom line with their vast R&D budget that some posters are so fond on shoving in as an argument for some other company's doom and gloom?

Honestly, the more time passes the more I realize Conroe and it's sucessors were a home run, an one in a ten change development, because the very same company cant leverage the same expertise that made Core so successful into their small core designs. And obviously Krzanich wont touch those 2 design teams at all, it is their big cow to keep milking and their comfort zone to go back when other parts of the company cant give him some piece of good news whatsoever.

That doesnt change the fact that is a waste of R&D to make internal competition in CPU design teams giving the back and forths done to the Core uarch, the minimal gains in performance and lower power consumption from uarch changes alone, and the deep need of making Atom more compelling at some key point. If they couldnt fight with BoM and platform costs or power consumption, performance was the route to go after all and the distintive selling point to fight ARM.

This product stack and positioning even hurt the Core line too. Having your design go from 6 to 95W because your sub 10W design cant compete without contra-revenue against the rest of the ARM pack is something the usual suspects praise, but it cant be achieved with some compromises down the road. Having your CPU design specialized at a narrower TDP range gives you more options design wise and probably better performance in all relevant parameters. This is exactly what Apple is doing with their Ax and what makes it so dreadly close in IPC against Core, a design that is the product of the maturity after a decade long worth of changes.

 

[IntelUser2000]

And that is in similar or lower power envelopes, and higher IPC than atom.

Artificial segmentation + delays and defeaturing

A team of 80 engineers in an outsorced office gave AMD their Cat uarch, which to the disdain of some here it's succesor ships in similar volumes WITHOUT any contra-revenue (it is sad to think AMD made more money out of Cat derived products than Intel did with it's Atom line).

Yea, but their Cat APUs are much more attractive than rest of AMDs chips. They also didn't make much money as a company because of that. Intel makes vast majority of money with Core chips, partly because they are good at segmenting to stop overlap. And I disagree about Bay Trail being less attractive. It brought similar performance at a similar/lower price. Their domination of the low-end markets that AMD traditionally excelled in proves it. Contra-revenue was used just to take marketshare against ARM devices, but not really in PC.

Goldmont based cores would have brought everything the Cat cores were and much more. The Cat core based platforms were uncompetitive in terms of platform power consumption.

That doesnt change the fact that is a waste of R&D to make internal competition in CPU design teams giving the back and forths done to the Core uarch, the minimal gains in performance and lower power consumption from uarch changes alone,

Are you sure because its not the setup of the team but other reasons? Perhaps the Core is really at the peak of advancement, and it's just Intel's refusal to budge from its outrageous prices compared to competition that makes it look bad.

It's possible to think of having two teams compete each other brought us the age of minimal gains, but are we sure? The reason they moved to Tick/Tock and two teams was to mitigate their weaknesses.

Conroe/Merom/Woodcrest - Haifa Team, CPU uArch focused
Nehalem - Oregon Team, platform focused
Sandy Bridge - Haifa Team, more CPU uArch focused
Haswell - Oregon Team, platform focused, this time for power management

They were saying that Merom generation cores lost SMT because Haifa team didn't have enough experience and did not want to risk delays(which always brings defeaturing). The Oregon Team that was experienced with HT with Netburst brought HT.

Nehalem architect was interviewed as saying that they could have made it to be faster, but they decided not to because it would have brought risks, and delays(and defeaturing).

That tells me delays due to not setting their goals right were substantial problems with them, which is why they moved from 5+ year BIG gains, with 2 year smaller gains.

Without being part of the management team, there's just no way for us to conclude what the real reason for Intel's failures were.

My conclusion?: They are just victims of a bygone era. They are Blockbuster. They are VHS/DVD rental places. They were Steam Engine makers and Horse Buggy companies.

 

[Arachnotronic]

Without being part of the management team, there's just no way for us to conclude what the real reason for Intel's failures were.

Not true. I suggest finding somebody who worked on the Atom team and left and talking to them.

 

[PPB]

Artificial segmentation + delays and defeaturing



Yea, but their Cat APUs are much more attractive for AMD than the rest. They also didn't make much money as a company because of that. Intel makes vast majority of money with Core chips, partly because they are good at segmenting to stop overlap. And I disagree about Bay Trail being less attractive. It brought similar performance at a similar/lower price. Their domination of the low-end markets that AMD traditionally excelled in proves it. Contra-revenue was used just to take marketshare against ARM devices, but not really in PC.

Goldmont based cores would have brought everything the Cat cores were and much more. The Cat core based platforms were uncompetitive in terms of platform power consumption.



Are you sure because its not the setup of the team but other reasons? Perhaps the Core is really at the peak of advancement, and it's just Intel's refusal to budge from its outrageous prices compared to competition that makes it look bad.

It's possible to think of having two teams compete each other brought us the age of minimal gains, but are we sure? The reason they moved to Tick/Tock and two teams was to mitigate their weaknesses.

Conroe/Merom/Woodcrest - Haifa Team, CPU uArch focused
Nehalem - Oregon Team, platform focused
Sandy Bridge - Haifa Team, more CPU uArch focused
Haswell - Oregon Team, platform focused, this time for power management

They were saying that Merom generation cores lost SMT because Haifa team didn't have enough experience and did not want to risk delays(which always brings defeaturing). The Oregon Team that was experienced with HT with Netburst brought HT.

Nehalem architect was interviewed as saying that they could have made it to be faster, but they decided not to because it would have brought risks, and delays(and defeaturing).

That tells me delays due to not setting their goals right were substantial problems with them, which is why they moved from 5+ year BIG gains, with 2 year smaller gains.

Without being part of the management team, there's just no way for us to conclude what the real reason for Intel's failures were.

My conclusion?: They are just victims of a bygone era. They are Blockbuster. They are VHS/DVD rental places. They were Steam Engine makers and Horse Buggy companies.

Expanding to your comments:

What you mention on Merom not having HT (god that would have be compelling considering the T5300 I am using to post this), or SB coupling L3 speeds to Core speeds, Haswell then undoing that, having Skylake lose the FIVR, probably Cannonlake having it back, are all signs of 2 design teams not really fiddling with each other about what design decisions they make. Obviously we cant tell because we are not in inside that bubble, but we can think about it with the evidence shown in the silicon they designed. And to give more insight on what you posted, Oregon is the platform oriented design team, yet Haswell was a bigger gain to Ivy bridge that Skylake is to Broadwell, probably even rivaling Sandy Bridge to Westmere/Nehalem. Probably that time the Oregon team had more time/freedom to implement performance oriented changes compared to Nehalem when according to that engineer they didnt get the greenlight to do so. Anyhow, either of those 2 teams going for Atom and the other one for Core, with exchange of knowledge/IP inbetween, would probably be R&D better spent if we take at face value what you said of Core already having all the low hanging fruit and reaching some kind of IPC milestone. Moreso to my argument, Atom has a lot to gain yet it is coasting to some degree.

 

[IntelUser2000]

Expanding to your comments:

What you mention on Merom not having HT (god that would have be compelling considering the T5300 I am using to post this), or SB coupling L3 speeds to Core speeds, Haswell then undoing that, having Skylake lose the FIVR, probably Cannonlake having it back, are all signs of 2 design teams not really fiddling with each other about what design decisions they make.

You are smart enough to know that nevermind teams, individuals often don't agree on things. That's why you see so many conflicting beliefs everywhere and so many problems. Of course Intel guys are affected because they are run by people just like us.

There's a different way of looking at it. It seems interesting how one idea is presented both a terrible one that should never be mentioned yet also a paradigm shifting one. We'd have thought of dual teams as the latter when Intel was seemingly unstoppable and doing fantastic, but now we see it as the former, a reason for their destruction. Perhaps before its inception they saw it as a way to learn from each other. Part of the reason for Netburst's failure was their insistence on one design, which was completely different from what Haifa was doing. Combine the two and its perfect?...

(Also like I mention delays must have been such a big issue to them. The reason the XScale team was sold was because of delays too. And they had so many. Now I think of it, what causes so many delays?)

And to give more insight on what you posted, Oregon is the platform oriented design team, yet Haswell was a bigger gain to Ivy bridge that Skylake is to Broadwell, probably even rivaling Sandy Bridge to Westmere/Nehalem.

Over time, it looks like there's a possibility that they started blending. But its kind of obvious that Core 2's primary benefit was on PC. It was Nehalem with IMC and QPI and all scaling changes that brought massive server improvements. That put the final nail on AMD's server coffin. And Sandy Bridge brought enormous gains on laptops. Anand's Sandy Bridge Quad Core article had people floored! Acceptable graphics for casual gaming, un-quad like battery life, and 60-70% faster CPU on Single and Multi-thread.

But that was when mobile was gaining momentum. Then with Haswell they focused on power management changes. That needed enormous ecosystem cooperation to bring platform power down. And it did! 50-60% gain in battery life. The biggest ever.

We can't tell much about Skylake right now. Perhaps the gains would have been better without 14nm troubles. Also there's that "high-hanging" fruit issue.

 

[IntelUser2000]

By the way, I think their decision to cancel development of Broxton was retarded. Perhaps their team was incredibly inefficient, meaning they need to POUR money to compete barely with new Chinese teams.

However, this means they'll NEVER be able to recover. This also shakes everyone's confidence in their ability to stick to a project. Why did they pay $Billions in contra-revenue and pull it back?

The management made the same decision as the people that pull out of serious decisions like marriages due to the most trivial reasons, and kids that throw tantrums because parents spoiled them. It's likely that habit will return.

Surface Pro Gen 1 was not a money maker and sold in pitiful volumes. Apple's MacBook Air did not sell a lot in the first generation either. Now, both are regarded as top of its class. Stick to the project, and develop it.

 

[ZGR]

Here's hoping that Core can scale down far lower than what we have seen it do.

 

[jpiniero]

However, this means they'll NEVER be able to recover. This also shakes everyone's confidence in their ability to stick to a project. Why did they pay $Billions in contra-revenue and pull it back?

PC sales have fallen enough that they couldn't keep doing it.

 

[JDG1980]

Here's hoping that Core can scale down far lower than what we have seen it do.

IMO, what Intel should really do is bifurcate Core - create a stripped down version with a special focus on power efficiency for laptops, AIOs, and Surface-style tablets, and a beefed-up version with a focus on raw performance for servers and desktops. We're seeing the beginnings of that with AVX-512 being reserved to the big server/HEDT chips, but it needs to go further than that. There are unquestionably areas where server parts are compromised by the need for a design that works well with laptops, and vice versa.

Intel's current "jack-of-all-trades" approach with Haswell/Broadwell/Skylake has worked only because AMD has basically been MIA during the Bulldozer era. That stops in Q4 2016. Intel will have to step up their game, or risk losing market share.

 

[ehume]

IMO, what Intel should really do is bifurcate Core - create a stripped down version with a special focus on power efficiency for laptops, AIOs, and Surface-style tablets, and a beefed-up version with a focus on raw performance for servers and desktops. We're seeing the beginnings of that with AVX-512 being reserved to the big server/HEDT chips, but it needs to go further than that. There are unquestionably areas where server parts are compromised by the need for a design that works well with laptops, and vice versa.

Intel's current "jack-of-all-trades" approach with Haswell/Broadwell/Skylake has worked only because AMD has basically been MIA during the Bulldozer era. That stops in Q4 2016. Intel will have to step up their game, or risk losing market share.

Can that be done without leaving blank areas of silicon? Making a "stripped down" version sounds like a whole new project. I thought Atom was their stripped down project.

 

[bhtooefr]

The thing is, server and mobile actually don't have that big differences in design requirements, as I understand, mobile just needing less cores, but with per-core power demands that aren't that different. (Yes, there's things like AVX-512, but I thought the units were still present, just disabled, in the client versions.) Now, the uncore has to be different, to satisfy server bandwidth requirements, and mobile's graphics requirements, but...

It's really desktop that needs something different, because desktop needs to clock higher.

 

[Exophase]

By the way, I think their decision to cancel development of Broxton was retarded. Perhaps their team was incredibly inefficient, meaning they need to POUR money to compete barely with new Chinese teams.

However, this means they'll NEVER be able to recover. This also shakes everyone's confidence in their ability to stick to a project. Why did they pay $Billions in contra-revenue and pull it back?

The management made the same decision as the people that pull out of serious decisions like marriages due to the most trivial reasons, and kids that throw tantrums because parents spoiled them. It's likely that habit will return.

Surface Pro Gen 1 was not a money maker and sold in pitiful volumes. Apple's MacBook Air did not sell a lot in the first generation either. Now, both are regarded as top of its class. Stick to the project, and develop it.

The decision to cancel a far along product in favor of accelerating successor products is certainly unusual. Does anyone know of any cases where a similar strategy paid off? I keep thinking of 45nm Bulldozer but we all know how that ended up. And the people who said that 22nm's delay didn't have any bearing an 14nm's release.

But I guess the different cellular generations are pretty different from each other and it may well be possible that having a good 4G implementation doesn't help them that much with a 5G implementation.

 

[Exophase]

The thing is, server and mobile actually don't have that big differences in design requirements, as I understand, mobile just needing less cores, but with per-core power demands that aren't that different. (Yes, there's things like AVX-512, but I thought the units were still present, just disabled, in the client versions.) Now, the uncore has to be different, to satisfy server bandwidth requirements, and mobile's graphics requirements, but...

It's really desktop that needs something different, because desktop needs to clock higher.

Intel has server parts with peak single threaded performance on similar levels as their desktop parts (and the laptop parts aren't really that far behind). If mobile CPU performance - and I'm talking phone and tablet stuff here - were suitable for server then ARM server SoC vendors would have had a much easier time eroding Intel's market dominance.

 

[imported_ats]

Windows? Open? Open to malware and stealing your moms maiden name perhaps, other than that, open it is NOT. Well, it may be open to opensourcing all your privates to 1st 2nd and 3rd parties to ram advertising into every ori... Damn, sorry, snapped again. Like Intel, microsoft has reinvented it self as of late with going the whole telemetry/big-data as a business model route. If you ever pondered about linux the time to switch is now. Open source to protect your privates, noone else is doing it for you.

That's funny. My PC is more secure than any Android or iOS phone. My PC has ZERO advertisements in anything, Android/iOS not so much. My PC allows me to install or distribute any application I want, Android/iOS not so much.

So that's the thing with open, it means different things in different scenarios. Can you get Windows source code? generally not. Does it matter to say 99.999999999999999999% of computer users? no. In every other way, windows is very open and certainly in every way that actual users care about.

Also FYI, open source does jack all to protect your privates. Open source is purely an issue of copyright, it has no inherent goodness.

 

[TeknoBug]

There are still new devices being released with Atom in it.

 

[cytg111]

My PC is more secure...

Ok.

 

[Essence_of_War]

So to make sure I understand correctly, Intel has killed only the smartphone/tablet Atom SoCs?

For the time being we're still expecting to see Atom for microservers (that is to say like, Avoton, Denverton, and whatever-comes-after-Denverton)?

And of course whatever Frankenstein-ian Atom thing the Xeon Phi is made out of is still alive?

 

[Arachnotronic]

So to make sure I understand correctly, Intel has killed only the smartphone/tablet Atom SoCs?

For the time being we're still expecting to see Atom for microservers (that is to say like, Avoton, Denverton, and whatever-comes-after-Denverton)?

And of course whatever Frankenstein-ian Atom thing the Xeon Phi is made out of is still alive?

Your understanding is correct.