Is Intel's "Process Lead" Somewhat a Sham?

[Arachnotronic]

Interesting, I though FinFET+ was just a density improvement. Then I think FinFET Plus will likely be released 1 year after FinFET, around the start of 2017, which is more than 2 years behind Intel's 14nm (and still 1.3x less dense).

No. I e-mailed TSMC and received the following response:

Dear xxxxxxx,

Thank you for the message! Shipment schedule is customers' call and we
cannot reveal their schedule details. However, I can tell you that we will
be shipping both 16FF and 16FF+ in 2015. Products include both mobile and
non-mobile related applications.

Hope this is helpful.

I mean, at this point, TSMC claims it'll be shipping 16FF+ products next year, so unless we want to outright call them liars, this seems legit.

Intel is likely to blow it with Broxton. Mid-2015 launch? That gives them maybe a quarter or two of an edge on a potential 16 FF+ design built at TSMC. And we all know how good Intel is at hitting schedules for mobile products.

I believe Intel may be misleading its investors with respect to the size/magnitude of its process lead. I will be speaking with Intel to get their side of this tomorrow.

 

[witeken]

I didn't expect that. I though it would be similar to Snapdragon 600 vs. Snapdragon 800 with HKMG. Even then, we'll have to see how fast Qualcomm can deliver SoCs with those process nodes. At this point, it seems that there will only be 20nm products (maybe except for Apple) in 2015, with FinFET in 2016. So that gives Qualcomm only 1H to come up with FinFET+ with 10nm launching in the second half of 2016. In any case, TSMC's costumers will still be at a serious cost disadvantage because both 20nm and FinFET give them basically no cost/transistor improvement.

I believe Intel may be misleading its investors with respect to the size/magnitude of its process lead. I will be speaking with Intel to get their side of this tomorrow.

If Intel uses its process lead from a cost perspective, they can still gain a lot of market because they could offer much better products at the same margins.

And in the end, not only silicon but also the microarchitecture is important, so even with similar silicon the difference could be substantial.

 

[jdubs03]

I didn't expect that. I though it would be similar to Snapdragon 600 vs. Snapdragon 800 with HKMG. Even then, we'll have to see how fast Qualcomm can deliver SoCs with those process nodes. At this point, it seems that there will only be 20nm products (maybe except for Apple) in 2015, with FinFET in 2016. So that gives Qualcomm only 1H to come up with FinFET+ with 10nm launching in the second half of 2016. In any case, TSMC's costumers will still be at a serious cost disadvantage because both 20nm and FinFET give them basically no cost/transistor improvement.


If Intel uses its process lead from a cost perspective, they can still gain a lot of market because they could offer much better products at the same margins.

And in the end, not only silicon but also the micro-architecture is important, so even with similar silicon the difference could be substantial.

The A8 is highly likely to be 20nm. The A9 however, I'm not as certain, as Apple now is three different foundries to source their CPUs from. But in all likelihood, if both TSMC and Samsung's PR machine is being 100% honest, with their 16FF+ and 14nm nodes respectively, then Intel's lead did definitely did decrease. 10nm may be the same with Intel having only a one-year lead. Samsung and GF are definitely a big threat now for sure.

 

[Arachnotronic]

I didn't expect that. I though it would be similar to Snapdragon 600 vs. Snapdragon 800 with HKMG. Even then, we'll have to see how fast Qualcomm can deliver SoCs with those process nodes. At this point, it seems that there will only be 20nm products (maybe except for Apple) in 2015, with FinFET in 2016. So that gives Qualcomm only 1H to come up with FinFET+ with 10nm launching in the second half of 2016. In any case, TSMC's costumers will still be at a serious cost disadvantage because both 20nm and FinFET give them basically no cost/transistor improvement.

Intel may see a cost/transistor improvement going from 22nm -> 14nm, but without knowing what TSMC's cost/transistor was at 28nm we're really in no position to make relative claims.

And in the end, not only silicon but also the microarchitecture is important, so even with similar silicon the difference could be substantial.

You're right. That's why Qualcomm is bending Intel over a barrel this round even with Intel's magical 22nm process. QCOM won the micro-architectural battle at the SoC level, so I don't see why microarchitecture would suddenly be something Intel has over QCOM by late 2015/early 2016.

For all of the knocking I do of AMD's CPU group, I hate to say it, but Intel's mobile group is to Qualcomm what AMD is to Intel as far as execution goes.

 

[witeken]

You're right. That's why Qualcomm is bending Intel over a barrel this round even with Intel's magical 22nm process. QCOM won the micro-architectural battle at the SoC level, so I don't see why microarchitecture would suddenly be something Intel has over QCOM by late 2015/early 2016.

For all of the knocking I do of AMD's CPU group, I hate to say it, but Intel's mobile group is to Qualcomm what AMD is to Intel as far as execution goes.

How did Qualcomm win the microarchitectural battle if Silvermont is (1) faster (per clock) and (2) less power hungry? What I was especially referring to was ARM's A57 CPU, the successor of the infamously power hungry A15, that Qualcomm will be using in 2015.

 

[mrmt]

How did Qualcomm win the microarchitectural battle if Silvermont is (1) faster (per clock) and (2) less power hungry? What I was especially referring to was ARM's A57 CPU, the successor of the infamously power hungry A15, that Qualcomm will be using in 2015.

Qualcomm solution arrived first on the market and with a better cost and profit structure than Silvermont. So yes, Qualcomm wins this round.

 

[Arachnotronic]

How did Qualcomm win the microarchitectural battle if Silvermont is (1) faster (per clock) and (2) less power hungry? What I was especially referring to was ARM's A57 CPU, the successor of the infamously power hungry A15, that Qualcomm will be using in 2015.

How many Silvermont based phones are there on the market? How many Silvermont based Android tablets are on the market?

Let's not BS ourselves here; the proof is in the pudding. Qualcomm's QCT is making $$$ and its chips are in every Android and Windows phone that matters. Intel's attempts at smartphone processors have brought its shareholders nothing but misplaced hope and massive losses as a result of hundreds of millions spent on R&D that will never be recouped.

Intel's mobile group just doesn't appear to be able to execute and I will continue to believe so until proven otherwise. If Broxton ends up being a late, design-win-less disaster then as an Intel investor, I will give up and move on (unless of course the PC makes a magical recovery...but I doubt this). I've waited years for meaningful financial progress in mobile and Intel has not delivered. You'd think things would have gotten better post-Medfield, but they really didn't.

 

[VirtualLarry]

How many Silvermont based phones are there on the market? How many Silvermont based Android tablets are on the market?

Let's not BS ourselves here; the proof is in the pudding. Qualcomm's QCT is making $$$ and its chips are in every Android and Windows phone that matters. Intel's attempts at smartphone processors have brought its shareholders nothing but misplaced hope and massive losses as a result of hundreds of millions spent on R&D that will never be recouped.

Intel's mobile group just doesn't appear to be able to execute and I will continue to believe so until proven otherwise. If Broxton ends up being a late, design-win-less disaster then as an Intel investor, I will give up and move on (unless of course the PC makes a magical recovery...but I doubt this). I've waited years for meaningful financial progress in mobile and Intel has not delivered. You'd think things would have gotten better post-Medfield, but they really didn't.

Perhaps there are some areas where x86 / x64 is really just a bad fit, all things considered.

Edit: I guess I'm speaking mostly of the power penalty for x86 decode. Intel is attempting to use their brute strength (process node lead) to pound a square peg (x86) into a round hole (true mobile low-power platforms).

 

[Arachnotronic]

Perhaps there are some areas where x86 / x64 is really just a bad fit, all things considered.

Intel's weakness wasn't the CPU; Silvermont was quite good. The problem was at the SoC and timing level. So, no, "X86" has nothing to do with Intel's failures in mobile.

Merrifield was a dual core with no integrated modem. Marketing fail on the dual core point and just technical fail with the lack of modem. Moorefield does better, but still no modem. Intel's chips also don't have the imaging/hardware encode/decode that Qualcomm's chips have.

In short, the CPU was the ONLY part of these SoCs that Intel got right.

 

[Hulk]

I think the question posed is an unintentional red herring.

The process lead, real or imagined, waxing or waning, growing or shrinking, takes secondary importance to the more pressing question of "does Intel's process lead matter?"

To x86 CPU customers clearly the process lead matters, and to AMD's employees, shareholders, and enthusiastic customers it also matters.

But does it really matter (has it ever mattered?) to Samsung, TSMC, UMC, etc that Intel has a <insert preferred adjective to describe the breadth of the gap> lead over them?

Until Intel leverages its process lead in areas that fabless companies (AMD being the only exception here) live and die in, does it really matter to those fabless companies or their foundries that Intel may or may not have a process tech lead?

It did not matter to Nvidia on the eve of Larrabee's launch. And it doesn't seem to have Apple, Qualcomm, or Samsung running to hills either.

To whatever extent Intel has a process lead, it appears to be equally adept at making its lead mostly irrelevant to all other business segments outside x86 processors in traditional form factors.

Of course that is about to change, but only if you read Intel's PR...which goes back to the OP's point (regarding reading PR and extrapolating reality from it)

While I'm not anywhere near an expert in this field I think that in this particular case perhaps Intel's process lead matters if it gets x86 into tablet and possibly phones. If that can happen at 14nm (perhaps for tablets) and 10nm for phones then it seems like the lid will be blown off the kettle. Desktop power in tablets and phones because Intel has the wide, superscaler, OO, etc.. chips that it has been refining for 40 years, along with the process technology to go with it.

ARM on other hand would have to compete by "opening" up current designs and hoping their current process technology could "fit" thermal constraints.

It's the top down (Intel) bottom up (ARM) collision we have been predicting for a few years now. Process technology will play a big part in this clash I think.

 

[Nothingness]

How many Silvermont based phones are there on the market? How many Silvermont based Android tablets are on the market?

Anway the supposed huge perf/W advantage of Silvermont vs A7 or Qualcomm does not translate into improved battery life in end products, so who cares? Of course we'll have to wait for Android BT tablets to be sure it's not Windows that is doing a poor job at power management, or if the BT perf/W wonder is useless.

 

[Qwertilot]

That sort of thing is going to be a big problem I think - things are already quite fast enough for large chunks of the market, with a very low power draw too and they'll only improve over time.

There is rather a large luxury market too of course but much of that is locked into using own brand things, probably inevitably due to the amounts they're selling.

Really, the thing that looks rather more likely to get them progress is if they can get their manufacturing costs really low and undercut everyone. Sounds rather like they're trying that actually.

 

[witeken]

Perhaps there are some areas where x86 / x64 is really just a bad fit, all things considered.

Edit: I guess I'm speaking mostly of the power penalty for x86 decode. Intel is attempting to use their brute strength (process node lead) to pound a square peg (x86) into a round hole (true mobile low-power platforms).

ISA doesn't matter at all: Power Struggles: Revisiting the RISC vs. CISC Debate on Contemporary ARM and x86 Architectures.

A related issue is the performance level for which our results hold. Considering very low performance processors, like the RISC ATmega324PA microcontroller with operating frequencies from 1 to 20 MHz and power consumption between
2 and 50mW [3], the overheads of a CISC ISA (specifically the complete x86 ISA) are clearly untenable. In similar domains, even ARM’s full ISA is too rich; the Cortex-M0, meant for low power embedded markets, includes only a 56 instruction subset of Thumb-2. Our study suggests that at performance levels in the range of A8 and higher, RISC/CISC is irrelevant for performance, power, and energy.

 

[witeken]

Anway the supposed huge perf/W advantage of Silvermont vs A7 or Qualcomm does not translate into improved battery life in end products, so who cares? Of course we'll have to wait for Android BT tablets to be sure it's not Windows that is doing a poor job at power management, or if the BT perf/W wonder is useless.

Snapdragon 800 will quickly throttle while Silvermont runs at its highest clock speed, effectively multiplying performance by the performance/watt advantage.

 

[Nothingness]

Snapdragon 800 will quickly throttle while Silvermont runs at its highest clock speed, effectively multiplying performance by the performance/watt advantage.

I agree that benchmarks that measure battery life lack the number of iterations they make that'd certainly help better understand what's going on.

I'm sure BT is better but as far as I know we don't have any analysis with a power virus as Anand did with Apple A7. And Intel didn't repeat the experiment they made to demonstrate how their previous tablet chip were better than the competition. Perhaps when Android tablets will be here?

 

[Khato]

I mean, at this point, TSMC claims it'll be shipping 16FF+ products next year, so unless we want to outright call them liars, this seems legit.

It's hard to call them liars with respect to 'shipping' wafers on a certain process because that label doesn't entail any level of functionality. Whether they're functional chips with the best power/density/performance in the industry or pretty pieces of silicon that go 'pop' when you apply power doesn't matter - they can still claim that they're 'shipping'. (Note that I've been extremely skeptical of finfet adoption timelines ever since they were announced in the wave of 'me too' press releases shortly after Intel's 22nm announcement - they all reeked of PR stunts where marketing was telling the engineers what to do.)

And the reason why I'm leery of any claim TSMC makes of late? It's pretty simple actually. They keep making claims about how they're upcoming finfet processes will compare against Intel's upcoming 14nm... despite the fact that they have no technical data on Intel's 14nm process upon which to base those claims. The only comparison point they have from Intel would be rough PR data, which I'm certain is what they're happily doing because they can easily interpret it to their liking within a rather large margin of error.

 

[Nothingness]

And the reason why I'm leery of any claim TSMC makes of late? It's pretty simple actually. They keep making claims about how they're upcoming finfet processes will compare against Intel's upcoming 14nm... despite the fact that they have no technical data on Intel's 14nm process upon which to base those claims. The only comparison point they have from Intel would be rough PR data, which I'm certain is what they're happily doing because they can easily interpret it to their liking within a rather large margin of error.

Didn't Intel do the same when they showed a slide comparing their process density to TSMC one? Note Intel were cautious, to the point of even stating their own figure for 14nm was a forecast (which is hard to believe for a slide published in Q4 2013, don't you think?).

 

[witeken]

It's hard to call them liars with respect to 'shipping' wafers on a certain process because that label doesn't entail any level of functionality. Whether they're functional chips with the best power/density/performance in the industry or pretty pieces of silicon that go 'pop' when you apply power doesn't matter - they can still claim that they're 'shipping'. (Note that I've been extremely skeptical of finfet adoption timelines ever since they were announced in the wave of 'me too' press releases shortly after Intel's 22nm announcement - they all reeked of PR stunts where marketing was telling the engineers what to do.)

And the reason why I'm leery of any claim TSMC makes of late? It's pretty simple actually. They keep making claims about how they're upcoming finfet processes will compare against Intel's upcoming 14nm... despite the fact that they have no technical data on Intel's 14nm process upon which to base those claims. The only comparison point they have from Intel would be rough PR data, which I'm certain is what they're happily doing because they can easily interpret it to their liking within a rather large margin of error.

You make some good points.

I think it will be very remarkable of TSMC if they manage to release 3 full generations of FinFET within basically less than 3 years. Maybe they're just focusing on performance while waiting for EUV for higher density to leapfrog Intel on all transistor characteristics :sneaky:.

 

[witeken]

Didn't Intel do the same when they showed a slide comparing their process density to TSMC one? Note Intel were cautious, to the point of even stating their own figure for 14nm was a forecast (which is hard to believe for a slide published in Q4 2013, don't you think?).

They used publicly available data from TSMC, while I don't think Intel has such public data for 14nm and 10nm. I think they simply didn't want to guarantee anything, because 14nm was still 4 months away from volume production. And because TSMC's 20nm on the graph is a forecast based on public data, the 35% advantage they claimed was therefore also a forecast.

 

[Phynaz]

No. I e-mailed TSMC and received the following response:



I mean, at this point, TSMC claims it'll be shipping 16FF+ products next year, so unless we want to outright call them liars, this seems legit.

Intel is likely to blow it with Broxton. Mid-2015 launch? That gives them maybe a quarter or two of an edge on a potential 16 FF+ design built at TSMC. And we all know how good Intel is at hitting schedules for mobile products.

I believe Intel may be misleading its investors with respect to the size/magnitude of its process lead. I will be speaking with Intel to get their side of this tomorrow.

TSMC could ship test wafers in Dec 2015 and meet the claims they made in the email.

 

[Nothingness]

They used publicly available data from TSMC, while I don't think Intel has such public data for 14nm and 10nm.

Oh I didn't know that TSMC data for 10nm was public. Do you have a link to that data?

I think they simply didn't want to guarantee anything, because 14nm was still 4 months away from volume production.

I hope that 4 months before volume production they had accurate figures for density, not just forecasts

 

[Nothingness]

TSMC could ship test wafers in Dec 2015 and meet the claims they made in the email.

TSMC already have taped out test chips on 16FF

 

[Arachnotronic]

Oh I didn't know that TSMC data for 10nm was public. Do you have a link to that data?

I hope that 4 months before volume production they had accurate figures for density, not just forecasts

10-FinFET will offer greater than 25% speed improvement the same total power compared to 16-FinFET plus. More importantly 10-FinFET greater than 45% compared to 16-FinFET plus. 10-FinFET will offer 2.2x of density improvement over its previous generation 16-FinFET plus.

http://seekingalpha.com/article/214...sses-q1-2014-results-earnings-call-transcript

 

[Khato]

I hope that 4 months before volume production they had accurate figures for density, not just forecasts

Haha, yeah, Intel has reasonably accurate figures for 10nm density too... good luck getting them though I can understand why investors especially are nervous about Intel's sudden change in behavior - previously we would have had a paper or two on Intel's 14nm process by now, and there would have been a 14nm SRAM wafer on display something like 2 years ago with a statement of how small they were able to make a 6T SRAM cell. But instead we have nothing more than rough PR charts that are meant to reassure while not giving any useful data to the competition. That's why all of the non-released nodes in that chart were labeled as forecast.

Of course at this point all indications are that Intel's 14nm process won't be groundbreaking, it'll merely be a continuation of the status quo. Along the lines of Intel's 14nm being to TSMC's '16nm FF+' as Intel's 22nm currently is to TSMC's 28nm HPM. It's Intel's 10nm that has the potential for disruption - if you look at that Intel technology leadership timeline it's pretty easy to see that Intel has been introducing new transistor technologies every other process node since strained silicon with 90nm.

 

[Nothingness]

http://seekingalpha.com/article/214...sses-q1-2014-results-earnings-call-transcript

Thanks a lot :thumbsup: