Intel Investor Meeting 2015: November 19

[Arachnotronic]

Trust comes from a history and takes years to develop. Especially in a situation on dcg where server benefit of going from 22 to 14 nm is minor why make all that pr nonsense and eroding what is absolutely most valuable on the server side; trust.

What the heck are you talking about? Going from 22nm -> 14nm in servers buys the ability to pack in more cores/performance within the same power envelope, thereby increasing (often dramatically) performance/watt, a critical component of the TCO equation.

 

[raghu78]

Intel lost the Tablet and Smart-Phone wars because they didnt want to compete against their own Big cores. The badly castrated ATOM for years, especially latest 14nm ATOM could be the best CPU for the tablet market but noooo, they castrated to hell because they wanted to push Core-M. Good lack competing against ARM CPUs using Intels big core x86 designs for the Tablet and Phone market.

well said. Intel is all about margins. They will never develop a Atom which runs too close to their big cores. The fact is an Atom SoC has to sell in the tens of $ while Core M sells in the hundreds of $. The minute Atom gets too good there goes your Core M sales and with it their margins. Even if Atom had Sandy bridge IPC with a 128 bit FPU it would be extremely attractive since it will be priced very competitively against ARMv8 SoCs. But no that would mean cannibalizing Intel Core M sales which are of higher margin.

Apple, Qualcomm and Samsung are going to hammer Intel in the mobile market due to its reluctance to develop a Atom which is right up there with their big core. Already we are seeing A9X go toe to toe with Skylake Core M. Snapdragon 820 seems to also have excellent single thread performance as Qualcomm finally accept that single thread performance is what drives the user experience in phones and tablets. Heck single thread performance is what matters even in desktops and notebooks. Thats why Intel rules and AMD sucks. Apple is most likely to implement SMT in A10/A10X and with the move to TSMC InFO and further process maturing at TSMC 16FF+ and other micro-architectural improvements I would not be surprised to see a 2C/4T A10X running at 2.4 Ghz and hitting 3.5k/8k in single / multi thread perf in GB3. I also look forward to SPEC CPU 2006 benchmarks on A9X and next year A10X. Qualcomm will have to keep up with Apple for competitive reasons as Apple is taking huge market share from the high end Android phones. Intel's mobile ambitions are going to be toast very soon.

 

[Arachnotronic]

well said. Intel is all about margins. They will never develop a Atom which runs too close to their big cores. The fact is an Atom SoC has to sell in the tens of $ while Core M sells in the hundreds of $. The minute Atom gets too good there goes your Core M sales and with it their margins. Even if Atom had Sandy bridge IPC with a 128 bit FPU it would be extremely attractive since it will be priced very competitively against ARMv8 SoCs. But no that would mean cannibalizing Intel Core M sales which are of higher margin.

Apple, Qualcomm and Samsung are going to hammer Intel in the mobile market due to its reluctance to develop a Atom which is right up there with their big core. Already we are seeing A9X go toe to toe with Skylake Core M. Snapdragon 820 seems to also have excellent single thread performance as Qualcomm finally accept that single thread performance is what drives the user experience in phones and tablets. Heck single thread performance is what matters even in desktops and notebooks. Thats why Intel rules and AMD sucks. Apple is most likely to implement SMT in A10/A10X and with the move to TSMC InFO and further process maturing at TSMC 16FF+ and other micro-architectural improvements I would not be surprised to see a 2C/4T A10X running at 2.4 Ghz and hitting 3.5k/8k in single / multi thread perf in GB3. I also look forward to SPEC CPU 2006 benchmarks on A9X and next year A10X. Qualcomm will have to keep up with Apple for competitive reasons as Apple is taking huge market share from the high end Android phones. Intel's mobile ambitions are going to be toast very soon.

You really think Apple is gaining share against Android because Apple has better ST perf in GB3 than the Android vendors?

 

[raghu78]

You really think Apple is gaining share against Android because Apple has better ST perf in GB3 than the Android vendors?

Its a lot of factors - the iOS ecosystem, the Apple brand and definitely the user experience (which depends on both hardware and software). btw single thread performance is probably the number 1 factor which affects the user experience and responsiveness on a phone. So even though its not the primary reason I think ST perf is one of the reasons. Why do you think Qualcomm has gone with a custom Kryo core which has twice the single thread performance of A57 cores ?

http://semiaccurate.com/2015/11/14/qualcomm-shows-us-the-snapdragon-820/

"The Kyro CPU cores offer up to double the single threaded performance over the vanilla A57 cores in the 810. Qualcomm has also reduced the total number of cores from eight in the 810 to four in 820. Specifically there are two of the Kyro cores and two low power Kyro cores in big.LITTLE configuration.

Qualcomm reiterated its commitment to fewer, faster cores during the Snapdragon 820 event. They believe this is the right strategy because of how lightly threaded the vast majority of apps are. To Qualcomm heterogeneous processing investments like more fixed function hardware and better modems are massively more important than increasing core counts. Qualcomm’s leadership was very vocal about having a maximum of four cores on their high-end products until phone apps start using more than two threads reliably."

 

[Arachnotronic]

Uh, yeah, nobody is denying that ST performance is important, but what I'm saying is that pretty much no consumer is going to make a phone purchasing decision because of how the phone's CPU performs.

It's something that the customer appreciates after they've bought the thing.

Good performance will help Qualcomm keep its current customers from using solutions from other vendors, so in that sense this should benefit Qualcomm, but to think that Kryo performing as well as claimed is going to influence purchasing decisions is a bit naive, IMO.

 

[Nothingness]

Its a lot of factors - the iOS ecosystem, the Apple brand and definitely the user experience (which depends on both hardware and software). btw single thread performance is probably the number 1 factor which affects the user experience and responsiveness on a phone. So even though its not the primary reason I think ST perf is one of the reasons. Why do you think Qualcomm has gone with a custom Kryo core which has twice the single thread performance of A57 cores ?

http://semiaccurate.com/2015/11/14/qualcomm-shows-us-the-snapdragon-820/

"The Kyro CPU cores offer up to double the single threaded performance over the vanilla A57 cores in the 810. Qualcomm has also reduced the total number of cores from eight in the 810 to four in 820. Specifically there are two of the Kyro cores and two low power Kyro cores in big.LITTLE configuration."

Current Kryo Geekbench scores are certainly not twice the scores of Cortex-A57: http://browser.primatelabs.com/geekbench3/compare/2551393?baseline=4159755

And that still is the case if you compare to the A57 in 810: http://browser.primatelabs.com/geekbench3/compare/4263042?baseline=4159755

810 really looks like sh*t and is not representative of Cortex-A57 performance.

 

[mrmt]

You really think Apple is gaining share against Android because Apple has better ST perf in GB3 than the Android vendors?

We have basically two IHVs making money on the mobile market, Mediatek and Qualcomm, and Qualcomm makes the bulk of it with its modem IP. I think the trend here is very clear, the market has become a race to the bottom and costs, not a specific performance parameter will be the deal breaker here.

Apple is a very different beast because it controls a significant part of its value chain, so they may be able to leverage on the advantages of a fast processors. Think about it, do you think Microsoft would make Windows phone more compelling if it shipped with a faster processor? Would Samsung bloatware on Android become less annoying if we had faster ST performance? No to both questions, but the already excellent experience of iOS becomes even better with the faster processor.

Mobile was a necessity for Intel because it needs the volumes that it would bring, but it seems that they are shifting towards memories, and that means a clash with Samsung, an even more direct clash than with mobile. It will be fun to watch.

 

[Arachnotronic]

We have basically two IHVs making money on the mobile market, Mediatek and Qualcomm, and Qualcomm makes the bulk of it with its modem IP. I think the trend here is very clear, the market has become a race to the bottom and costs, not a specific performance parameter will be the deal breaker here.

Apple is a very different beast because it controls a significant part of its value chain, so they may be able to leverage on the advantages of a fast processors. Think about it, do you think Microsoft would make Windows phone more compelling if it shipped with a faster processor? Would Samsung bloatware on Android become less annoying if we had faster ST performance? No to both questions, but the already excellent experience of iOS becomes even better with the faster processor.

Mobile was a necessity for Intel because it needs the volumes that it would bring, but it seems that they are shifting towards memories, and that means a clash with Samsung, an even more direct clash than with mobile. It will be fun to watch.

Exactly the point I was trying to make. Great post, mrmt.

(Also love the comment about Samsung bloatware, it's so true)

 

[krumme]

What the heck are you talking about? Going from 22nm -> 14nm in servers buys the ability to pack in more cores/performance within the same power envelope, thereby increasing (often dramatically) performance/watt, a critical component of the TCO equation.

Better tell the potential customers of this dramatic effect on tco going from hw to bw because it seems to have slipped their assessment.

 

[Arachnotronic]

Better tell the potential customers of this dramatic effect on tco going from hw to bw because it seems to have slipped their assessment.

Intel is shipping Broadwell to its "early ship" customers now (i.e. Facebook, Microsoft, Google, etc.) who are so eager to get their hands on the improvement that BDW-EP brings that they're willing to accept basically pre-production chips with potentially significant errata.

 

[lopri]

We can look back a year (or more) ago and see how things have changed. It is Thanksgiving week and some people might be home early and want to kill some time reading. ^^

[DigiTimes] TSMC 10 nm trial production in 2015, mass production in 2016

[GSMArena] Samsung demonstrates 10nm FinFET for mobile - ahead of Intel

And then there is the classic he-said-she-said.

 

[lopri]

Current Kryo Geekbench scores are certainly not twice the scores of Cortex-A57: http://browser.primatelabs.com/geekbench3/compare/2551393?baseline=4159755

And that still is the case if you compare to the A57 in 810: http://browser.primatelabs.com/geekbench3/compare/4263042?baseline=4159755

810 really looks like sh*t and is not representative of Cortex-A57 performance.

Those A57's are overclocked to unknown frequencies. S820, I suppose, is at 2.2/1.6 GHz.

 

[Arachnotronic]

Those A57's are overclocked to unknown frequencies. S820, I suppose, is at 2.2/1.6 GHz.

'tis true. Stock Exynos 7420 doesn't score that well although still owns the S810.

 

[krumme]

Intel is shipping Broadwell to its "early ship" customers now (i.e. Facebook, Microsoft, Google, etc.) who are so eager to get their hands on the improvement that BDW-EP brings that they're willing to accept basically pre-production chips with potentially significant errata.

I am sure the web guys love the go away arm edition (albeit we can asume its without errors). Future dcg profit less so.

 

[Arachnotronic]

I am sure the web guys love the go away arm edition (albeit we can asume its without errors). Future dcg profit less so.

Not sure what you are talking about at this point, so I think I'm done here.

 

[krumme]

Not sure what you are talking about at this point, so I think I'm done here.

http://www.anandtech.com/show/9070/intel-xeon-d-launched-14nm-broadwell-soc-for-enterprise

 

[IntelUser2000]

The problem with Intel making density comparisons are because even *IF* Intel's process is more dense on a transistor level they don't use products that show that benefit.

Atom and Core are both not something you'd call dense at all, even when normalized for performance.

With initial LG NUCLUN 2 results showing decent clock speed advantage of Intel variant vs TSMC, we now know the answer to the question: Intel mobile chips being bad, a fault of their process or their design? Its the latter!

I am pretty sure due to being greedy with margins and being sidetracked trying to keep up with mobile they negatively affected both the Core and Atom line. Its not all about R&D and smart engineers, you need a clear goal, and a focus getting there. Both of which Apple probably has. Intel's server division seems much better though.

 

[StrangerGuy]

The problem with Intel making density comparisons are because even *IF* Intel's process is more dense on a transistor level they don't use products that show that benefit.

Atom and Core are both not something you'd call dense at all, even when normalized for performance.

With initial LG NUCLUN 2 results showing decent clock speed advantage of Intel variant vs TSMC, we now know the answer to the question: Intel mobile chips being bad, a fault of their process or their design? Its the latter!

I am pretty sure due to being greedy with margins and being sidetracked trying to keep up with mobile they negatively affected both the Core and Atom line. Its not all about R&D and smart engineers, you need a clear goal, and a focus getting there. Both of which Apple probably has. Intel's server division seems much better though.

I don't get the process worshippers at all, it's only a means to an end, not the end itself. The characteristics of the actual finished chip is the only thing that matters. The Apple A9 is impressive solely because by how well it performs from the end user POV, not because of the 14/16nm process it's built on.

 

[Nothingness]

With initial LG NUCLUN 2 results showing decent clock speed advantage of Intel variant vs TSMC, we now know the answer to the question: Intel mobile chips being bad, a fault of their process or their design? Its the latter!

First the frequency advantage of 2.4 GHz for Intel vs 2.1 GHz for TSMC is a rumor. Second it's possible Intel customized parts of the design to improve frequency and show an advantage over foundries (as far as I know, replacing some RTL blocks by custom implementation is allowed by some ARM license agreements).

 

[Hans de Vries]

My conclusion:
Your pictures seems to suggest that L3 cache density has nearly doubled from TSMC 20nm to TSMC 16nm. How is that possible ?
Maybe A8 actually has more than 4 MB L3 ? Maybe some portion of it is used as fixed function memory.

I haven't seen a retraction from ChipWorks on the 8MB L3 cache yet
but they may well have to. It's more likely to be 4MB L3.

 

[Hans de Vries]

I haven't seen a retraction from ChipWorks on the 8MB L3 cache yet
but they may well have to. It's more likely to be 4MB L3.

O well, here it is http://www.anandtech.com/show/9825/correcting-a9s-l3-cache

 

[dark zero]

And now there are rumors that Intel lost LG Nuclun 2 design....http://www.fudzilla.com/news/processors/39339-intel-loses-lg-s-nuclun-2-to-tsmc

 

[PaulIntellini]

O well, here it is http://www.anandtech.com/show/9825/correcting-a9s-l3-cache

Funny. Initally the said that A9 has 8 MB L3, not "is rumored to have". So I thought this was confirmed by Apple. Fare thee well, journalism...

 

[Hans de Vries]

Funny. Initally the said that A9 has 8 MB L3, not "is rumored to have". So I thought this was confirmed by Apple. Fare thee well, journalism...

Well, to be fair, Apple gives very little detail, if any, about its SOC's
so we should appreciate any investigative work done.

In this case the 8MB simply can not fit in 4.5mm2 because this leaves
only 0.067 um2 per bit which is below the known sram bit cell size of
0.07um2 (Normally there is quite some area overhead for real memory
arrays besides only the sum of the bit cell sizes)

 

[witeken]

Okay, so time for another round of Intel delays. This actually isn't new news, but it kinda changes my opinion...

Raising mobility through the use of non-silicon materials in the transistor channel remains a favored theme with Gargini but he appeared to push back the prospect of deployment until about 2020.

Gargini, discussed compound semiconductor on silicon, as he has done before. But he also discussed the arrival of multi-gate HKMG III-V devices on silicon substrates arriving by 2020. However he also discussed the need to optimize for power consumption rather than performance.

2020 looks, certainly by their 2 year roadmap they had up to 2012/2013, like 7nm for III-V. It is very hard, like this also shows: http://spectrum.ieee.org/semiconductors/design/changing-the-transistor-channel.

So unless it was just a general, vague statement to avoid spoilers, which actually isn't rare in this industry I guess, I and others like Kanter may be wrong and disappointed when 10nm arrives at last.

Note, BTW, that EUV was also aimed for 10nm.

http://www.eetimes.com/document.asp?doc_id=1266393

However, the articles seems credible if you read some more of it. E.g.

At the same event a year before Gargini had said that compound semiconductors in silicon could be an option in 2015.

For example, they're talking about leakage as a major problem. A more recent interview, with Mark Bohr from 1 year ago:

At very low leakage, you are limited by band-to-band tunneling. Sub-threshold leakage may be more of a problem with some III-V materials. So developing a universal, or widely useful, transistor from high performance to very low leakage is quite a challenging task.

Also, in the EETimes article they're talking about InSb, but it's quite possible that Intel has since then decided to use InGaAs. Note, BTW, that Bohr is talking in the follow-up question specifically about other companies, which implicitly implies that Intel does have it ready for 10nm.

Other companies may choose to push out the adoption of III-V, because all of the problems have not been solved for the 10nm generation. Tool readiness doesn’t seem to be the issue. It’s mostly device physics.

And has he then points out, it's a tall set of requirements and the decision for 7nm hasn't/hadn't been made yet.

I gotta say though, that what he says about High-k doesn't really line-up with "The High-k Solution" article. Okay, high-k might have been ready, but it wasn't really delayed since it was simply planned to go along the metal gates at the 45nm node.

So interesting information that does lower my confidence in III-V at 10nm, even though the vast majority of the data points point to it.

Edit: The bold part for TLDR: http://forums.anandtech.com/showpost.php?p=34198524&postcount=1. Maybe they have other interesting, even secret, stuff planned for 10nm, though.