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

[erunion]

Why do you think this? Samsung is claiming 14nm in high volume by EOY 2014...

Any insight from somebody who actually can parse this sea of marketing BS would be greatly appreciated.

I believe the term used was actually mass production. I looked up Samsung 28nm as a benchmark. There is a press report from may 2011 that Samsung 32/28nm were in mass production. In July 2012 there are small orders from Qualcomm for 28nm. But Samsung's key customer, Apple, didn't release 28nm until 28 months later, Sept 2013.

If 14nm parallels 28nm we wont see Apple 14nm until quarter1 2017!

 

[Nothingness]

Sure: http://www.realworldtech.com/forum/?threadid=133235&curpostid=133589

Thanks :thumbsup:

 

[raghu78]

Intel's node naming isn't any more questionable than other foundries'. In contrary, TSMC, GlobalFoundries and Samsung will go from 20nm to 16/14nm without any meaningful increase in transistor density, which is exactly the opposite of what Moore's Law implies. If Intel's slide is correct, Intel will have a 1.5x to 1.3x (FinFET+) density advantage with 14nm for the next 2 years, which will increase even more at 10nm (certainly versus 16nm). You can have the fastest transistors in the world (which TSMC apparently claims they will have), but if you have to pay the foundry tax, don't have great yields, have expensive wafers and you run out of die area because of the lack of density improvements, you won't be competitive against a company who doesn't have all of those problems.

Didn't you read about TSMC 16FF+ and Samsung 14 FINFET. 15% density increase compared to 20 nm planar.

https://markets.jpmorgan.com/research/email/-kjegkq4/GPS-1336259-0

http://globalfoundries.com/docs/def...dries-14nm-collaboration---final.pdf?sfvrsn=2

Sure: Intel Vindicated, Very Competitive With Apple's A7.

The too long, didn't read version is that Apple simply doesn't have 3D transistors.

what a failed comparison. you showed a power virus run on A7 and compared it to Baytrail running Cinebench R11.5. here is the page from which there is a normal workload like kraken running at same power consumption as A6x. But it completes the task much faster and races to sleep.

http://www.anandtech.com/show/7460/apple-ipad-air-review/3

let me quote anand

"For kicks I wanted to see just how much power I could get the iPad Air to draw. Here I’m looking at platform power during our mini-power-virus test from above:

How’s that for dynamic range? Almost 12W running all out, but around half that in what we’d normally consider to be a stressful CPU test. I couldn’t get any actual applications/games on the iPad Air to behave like this so the results above are purely academic (for now). A quick run through GFXBench 2.7’s T-Rex HD test confirms that even pushing the GPU won’t hit these numbers. The max I saw running T-Rex offscreen was ~6W, and turning to an actual game (Infinity Blade 3) the iPad Air pulls less than 5W."

Get the same workload running on Baytrail and A7 to make a fair comparison on power consumption.
Eg: Cinebench R11.5 single thread (for single core power)
Cinebench R11.5 multi thread (for multi core power)
3D Mark 2013 (for GPU + CPU power)
Furmark + Prime 95 (Max Power consumption)

http://www.anandtech.com/show/7460/apple-ipad-air-review
http://www.anandtech.com/show/7428/asus-transformer-book-t100-review

ipad air - 9.7 inch display, 32.4 wh battery , t100 - 10.1 inch display , 31wh battery

http://www.anandtech.com/show/7460/apple-ipad-air-review/8

web browsing battery life
Apple ipad air - 10.0 hrs
ASUS T100 - 8.6 hrs

video playback battery life
Apple ipad air - 13.63 hrs
ASUS T100 - 10.0 hrs

Even accounting for the marginal display size increase on T100, A7 is as power efficient or better than Baytrail. In video playback A7 has a 36% higher bettery life than T100.

Intel has had a manufacturing lead for a lot of years now, and the economics certainly aren't getting any better for the dedicated foundries at this point of Moore's Law, so I'm very skeptical for claims that TSMC or Samsung will suddenly catch up.

Intel will still have a 1 year time to market lead and a slight density advantage. But it won't have the generational transistor performance and power efficiency lead it had in the past. Also by the time Intel releases Broxton which is a baseband integrated 14nm SOC by late 2015/early 2016, Qualcomm should have their first FINFET SOCs out. Apple's A9 with FINFET will also be out for holiday 2015.

 

[Khato]

what a failed comparison. you showed a power virus run on A7 and compared it to Baytrail running Cinebench R11.5. here is the page from which there is a normal workload like kraken running at same power consumption as A6x. But it completes the task much faster and races to sleep.

http://www.anandtech.com/show/7460/apple-ipad-air-review/3

You do realize that it's the delta power consumption between idle and load that's primarily attributable to the SoC, right? Because the numbers in that review for delta power consumption quite clearly show approximately 8W for the power virus test and 3W for Kraken. Kraken, by the way, is a perfectly realistic single-threaded workload which could be considered comparable to a single-threaded run of cinebench. So even being kind and assuming that a full watt of that single-threaded delta is due to components other than the CPU you're still left with 2W for A7 and around 1W for Baytrail when it comes to single-threaded workloads.

I'll ignore questions of performance because I see no reason to get into that endless debate again. Same goes for wild speculation about future process nodes from TSMC and Samsung - we have a few tidbits to go along with a boatload of marketing. In both cases there's simply not enough empirical data to draw any hard conclusions and it gets tiring to try and discuss hypothesis with those who are convinced that they're correct and everyone else is wrong.

 

[Idontcare]

What does "high volume" mean to TSMC, Samsung or Intel. Different things, especially with Intel. Intel's transitions to newer nodes are very fast compared to the foundry companies (in term of % of product on a new node a year after product introduction).

You'd have to suss out exactly what Samsung is claiming. According to the EET article from above Samsung has multiple devices in production "in hopes" of reaching volume production by the end of the year. How many wafer starts qualify as Volume production? They don't say - a pretty clear indication that they don't really know. So, it is marketing BS put out there to try and generate more interest in their latest node.

One can also glean information from semicon equipment and service providers (the 'back story' basically). Take a ASML: http://phys.org/news/2014-04-asml-lull-chip-makers.html



So there is an additional reason not to take the PR as the Gospel Truth - how can a fab be absolutely positive, when a key fab equipment supplier is seeing lower bookings due to uncertainties from the very fabs proclaiming such great news! If all were so well, they'd be looking to ramp up fast, but they are not because their yields are not stable enough to support a higher capital equipment purchases.

Finally, the kicker for even the well informed public: Khato pointed out

If you really want to go to this level of detail for the sake of developing a qualitative investment risk/reward profile - then it will take allot of research. I would think that there are professional sources of information on process development just as The Microprocessor Report does for CPUs. Secondly, you'll need to develop some professional contacts in the fab business (or maybe at college research labs) that have a better sense of what is going on in their industry (and do so while avoiding violation of SEC requirements for disclosing insider information).

This is a task that could take years to perfect. So, it comes down to where you really want to spend your energies. I would think that most, if not all, high tech industries weave a tangled web where finding actionable intelligence on future products is difficult to obtain legally. This is why most large tech companies have a "competitive analysis" group that often play in the grey (or worse) areas of the law.

TL;DR - Those who really know can't talk.

You nailed it. Those who know can't talk, but if you want a good indication in reading the tea leaves then you look to what the suppliers are reporting and saying about their book:bill ratios. HVM at the fab cannot happen without HVM at the tool suppliers.

 

[raghu78]

You do realize that it's the delta power consumption between idle and load that's primarily attributable to the SoC, right? Because the numbers in that review for delta power consumption quite clearly show approximately 8W for the power virus test and 3W for Kraken. Kraken, by the way, is a perfectly realistic single-threaded workload which could be considered comparable to a single-threaded run of cinebench. So even being kind and assuming that a full watt of that single-threaded delta is due to components other than the CPU you're still left with 2W for A7 and around 1W for Baytrail when it comes to single-threaded workloads.

Again those are all total platform power numbers. So without accurate measurements we are guessing as to how much that A7 SOC alone draws at load. the battery life tests for web browsing and video playback for ipad air show it beating t100 running baytrail. Apple uses Cyclone's beastly single thread performance to finish tasks faster and race to sleep. So from a power/energy consumption point of view for a given task, Cyclone might still come out on top.

I'll ignore questions of performance because I see no reason to get into that endless debate again. Same goes for wild speculation about future process nodes from TSMC and Samsung - we have a few tidbits to go along with a boatload of marketing. In both cases there's simply not enough empirical data to draw any hard conclusions and it gets tiring to try and discuss hypothesis with those who are convinced that they're correct and everyone else is wrong.

There is not going to be any data until actual SOCs are compared in real world tests. We might hear about the technical details at IEDM 2014 and / or ISSCC 2015. But until we compare actual chips with similar TDP and die size we can never say. That is why the A7 vs Baytrail comparison is valid. Both SOCs are 100 - 110 sq mm. Both chips are in the 5w TDP range. And even when being built on an inferior Samsung 28nm planar process, A7 is every bit as efficient or better than Baytrail.

 

[Arachnotronic]

I believe the term used was actually mass production. I looked up Samsung 28nm as a benchmark. There is a press report from may 2011 that Samsung 32/28nm were in mass production. In July 2012 there are small orders from Qualcomm for 28nm. But Samsung's key customer, Apple, didn't release 28nm until 28 months later, Sept 2013.

If 14nm parallels 28nm we wont see Apple 14nm until quarter1 2017!

Samsung was sampling 32nm Exynos in Q4 2011 and obviously we saw 32nm A6 chip in Sept. 2012.

28nm seemed to show up in Sept. 2013 in high volumes, yes.

 

[Khato]

Again those are all total platform power numbers. So without accurate measurements we are guessing as to how much that A7 SOC alone draws at load. the battery life tests for web browsing and video playback for ipad air show it beating t100 running baytrail. Apple uses Cyclone's beastly single thread performance to finish tasks faster and race to sleep. So from a power/energy consumption point of view for a given task, Cyclone might still come out on top.

Quite correct that they're still guesses, but the Kraken power numbers have few enough variables for a reasonable margin of error. Especially since allowing a full 1W for other components that are fired up versus idle is almost certainly more generous than reality, so 2W per core under load is likely a low figure.

As for battery life tests. Congratulations on recognizing that the iPad Air is the superior platform. That is, after all, the primary reason why it wins out in that metric as those battery life tests are all about active idle and a good SoC should have pretty much zero influence on the results. Which, by the way, is a key flaw in Baytrail as it's quite obvious that it does have an impact on battery life when it comes to video playback.

 

[witeken]

You do realize that it's the delta power consumption between idle and load that's primarily attributable to the SoC, right? Because the numbers in that review for delta power consumption quite clearly show approximately 8W for the power virus test and 3W for Kraken. Kraken, by the way, is a perfectly realistic single-threaded workload which could be considered comparable to a single-threaded run of cinebench. So even being kind and assuming that a full watt of that single-threaded delta is due to components other than the CPU you're still left with 2W for A7 and around 1W for Baytrail when it comes to single-threaded workloads.

I'll ignore questions of performance because I see no reason to get into that endless debate again. Same goes for wild speculation about future process nodes from TSMC and Samsung - we have a few tidbits to go along with a boatload of marketing. In both cases there's simply not enough empirical data to draw any hard conclusions and it gets tiring to try and discuss hypothesis with those who are convinced that they're correct and everyone else is wrong.

I was thinking if I would reply, but this comment is exactly what I wanted to say, and very good formulated. This discussion is mostly speculation about, as Intel17 almost poetically said, marketing claims. This means the 'facts' are open for interpretation, which means it will be warped by people to fit their own bias. In my case, this means I'm skeptical about TSMC's and Samsung's ability (as their marketing departments claim) to suddenly be able to catch up.

 

[Enigmoid]

Quite correct that they're still guesses, but the Kraken power numbers have few enough variables for a reasonable margin of error. Especially since allowing a full 1W for other components that are fired up versus idle is almost certainly more generous than reality, so 2W per core under load is likely a low figure.

As for battery life tests. Congratulations on recognizing that the iPad Air is the superior platform. That is, after all, the primary reason why it wins out in that metric as those battery life tests are all about active idle and a good SoC should have pretty much zero influence on the results. Which, by the way, is a key flaw in Baytrail as it's quite obvious that it does have an impact on battery life when it comes to video playback.

Its impossible to directly compare the SOC's due to the different ecosystems.

However, due to Apple's optimizations, a significant penalty must be applied to the A7 if you are looking at technical merit alone.

Their mac lineup gets significantly better battery life than competitors using the same same chips due to optimizations. Devices using the A7 won't be any different.

For example

Lower driver overhead (100%). Fact is that Apple, writing custom drivers can get Mantle like performance increases. I would expect the same for sunspider, kraken, etc (browser benches).

These things need to be taken into account when looking at performance and power characteristics of chips.

 

[jdubs03]

Its impossible to directly compare the SOC's due to the different ecosystems.

However, due to Apple's optimizations, a significant penalty must be applied to the A7 if you are looking at technical merit alone.

Their mac lineup gets significantly better battery life than competitors using the same same chips due to optimizations. Devices using the A7 won't be any different.

For example

Lower driver overhead (100%). Fact is that Apple, writing custom drivers can get Mantle like performance increases. I would expect the same for sunspider, kraken, etc (browser benches).

These things need to be taken into account when looking at performance and power characteristics of chips.

For sure, just the amount of change in performance from ios 6 to 7 for the iphone 5, proves your point pretty well. The processing prowess for the A7 is definitely there, but the software improvement significantly boosts the benchmark performance.

Comparing the A7 to BT-3770/5 isn't a fair playing ground, the only way to get that true comparison is if each chip ran the same OS.

 

[Idontcare]

Quick question is samsung/global foundry 14nm a true 14nm, or is it old node with finfets added like tmsc? I ask since samsung only claims 15% increased densisty over industry standard 20nm. this is what i want to express. Anyway , i still want to say that Intel now is also at high level. Currently, its position can not be replaced so easily.

Samsung's 14nm process is a 20nm BEOL + 14nm FEOL hybrid node, no different in integration approach than TSMC's 16FF or GF's 14XM.

 

[Ajay]

You nailed it. Those who know can't talk, but if you want a good indication in reading the tea leaves then you look to what the suppliers are reporting and saying about their book:bill ratios. HVM at the fab cannot happen without HVM at the tool suppliers.

Thanks IDC! I really should have bolded the supplier part too. I've read allot of articles since Intel17 first posted as well as following this thread. Finally I recalled some discussion we, and others, had in the 2012-2013 time frame about those who really know the details aren't able to talk (unless they care to risk losing their jobs).

The final piece struck me after reading the phys.org article and noticing that this in a metric I use when picking applicable stocks - what is the long term guidance given by key vendors on bookings.

Other than that, it sure is a big hairball trying to discern the electrostatic performance, xtor density or Bose Einstein defect density for unreleased silicon (we are lucky if we get some of that post product release). Some info can be discerned from various semiconductor symposiums, but sometime the presentations are sandbagging (Intel's 22nm Finfet) or presenting more optimistic results (the norm for foundries).

BTW, it's really awesome having you back here posting. You were very much missed. I hope all is well in Taipei!

 

[ancientarcher]

Samsung's 14nm process is a 20nm BEOL + 14nm FEOL hybrid node, no different in integration approach than TSMC's 16FF or GF's 14XM.

As IDC pointed out earlier, I don't know why people are comparing Intel's FinFets to the foundries' (TSMC and Samsung/GloFo) since they don't compete as manufacturers in any segment (except perhaps FPGAs).

When talking about ARM chips, rather than comparing the foundries' manufacturing prowess to Intel, it will be more worthwhile to focus on what improvement gain can Apple/Qualcomm/Sammy/Nvidia etc get on their SoCs because of the improvement of the process node. And that is 30% for 20nm (planar) over 28nm planar and 20% for 20nm finfets (or 14/16nm whatever they call it) over 20nm planar.

So, there is approximately around 50% improvement on manufacturing alone -whenever it comes, in a year or two. Add to that architecture improvements, and the recent history of ARM chips suggests that significant architectural improvements can be wrung out. I think another 30-40% improvement based on fine tuning chip architecture alone is possible within the next 2 years. So, basically we are looking at a doubling of SoC speeds with FinFet transistors.

Intel should be skating where the puck is heading towards...

 

[Revolution 11]

I am among the first to be skeptical of claims by Intel fanboys but your guesses are just as thin. You can't tell Intel to "skate where the puck is" on only your gut instinct. These companies aren't selling lemonade, you know.

 

[sxr7171]

Its impossible to directly compare the SOC's due to the different ecosystems.

However, due to Apple's optimizations, a significant penalty must be applied to the A7 if you are looking at technical merit alone.

Their mac lineup gets significantly better battery life than competitors using the same same chips due to optimizations. Devices using the A7 won't be any different.

For example

Lower driver overhead (100%). Fact is that Apple, writing custom drivers can get Mantle like performance increases. I would expect the same for sunspider, kraken, etc (browser benches).

These things need to be taken into account when looking at performance and power characteristics of chips.

From a practical perspective it matters whether Windows RT or Android can offer that level of optimization to Intel.

 

[Nothingness]

Comparing the A7 to BT-3770/5 isn't a fair playing ground, the only way to get that true comparison is if each chip ran the same OS.

Or running benchmarks that don't rely on the OS or other software such as the browser; this disqualifies any Javascript benchmark (Sunspider, Octane, Kraken), and basically leaves us with a single benchmark, Geekbench, which is a pity as it's rather small, but at least it provides some interesting data to compare various CPU's.

 

[seitur]

Serious question for debate. With TSMC claiming that its 16nm FinFET+ will have similar performance to Intel's 14nm transistors, and with Samsung claiming 14nm high volume production this year, just how much of a process lead does Intel actually have? Why does Intel keep claiming that it's *extending* its lead here?

Seems to me that there is one but that the gap is shrinking...at least if these PRs from TSMC/GloFo are to be believed (and that's the real debate here...whether they are to be believed).

Discuss.

Intel has actual 22nm CPUs for anyone to buy since 2012.

Intel is going to have 14nm CPUs for anyone to buy in 2014-Q1 2015.

TSMC and others products are still 28nm.

What are we talking again?


Sorry I have heard TSMC, GF and Samsung boosting about 20nm process being ready, being in volume production, etc for almost 2 years already.
So point me to actual CPUs or GPUs in those processes I can buy?

What are talking about again?


Intel have huge lead.

 

[Idontcare]

When talking about ARM chips, rather than comparing the foundries' manufacturing prowess to Intel, it will be more worthwhile to focus on what improvement gain can Apple/Qualcomm/Sammy/Nvidia etc get on their SoCs because of the improvement of the process node. And that is 30% for 20nm (planar) over 28nm planar and 20% for 20nm finfets (or 14/16nm whatever they call it) over 20nm planar.

So, there is approximately around 50% improvement on manufacturing alone -whenever it comes, in a year or two. Add to that architecture improvements, and the recent history of ARM chips suggests that significant architectural improvements can be wrung out. I think another 30-40% improvement based on fine tuning chip architecture alone is possible within the next 2 years. So, basically we are looking at a doubling of SoC speeds with FinFet transistors.

Intel should be skating where the puck is heading towards...

Well said.

Intel is definitely skating in the right direction, it wasn't mere happenstance that Brian Krzanich was put into the CEO role after occupying the COO role (which came from a succession of promotions up through manufacturing ranks starting out as a process engineer).

That's a guy who has built his career around making sure Intel's production nodes accomplish what Intel needs them to accomplish. And now Intel needs their production nodes to accomplish something that no other foundry can touch; such that their IC design go places the mobile competition can't go.

He intimately knows the value that comes from process nodes in a direct and opposite way to that of Dirk Meyer. Talk about your night and day difference. Dirk knew design, not process. So he had no problem getting rid of the fabs and putting AMD's future in the hands of IC design alone.

Brian isn't cut from the same cloth. Intel's BoD knew what they were doing. I expect Intel to use 14nm, and then 10nm, as a pivot point to swing into territory they have never trod before (smartphones). Much to the chagrin of the fabless behemoths (Apple, Qualcomm) that have no choice but to rely on TSMC or GF/Samsung for their manufacturing.

 

[AtenRa]

Intel has actual 22nm CPUs for anyone to buy since 2012.

Intel is going to have 14nm CPUs for anyone to buy in 2014-Q1 2015.

TSMC and others products are still 28nm.

What are we talking again?


Sorry I have heard TSMC, GF and Samsung boosting about 20nm process being ready, being in volume production, etc for almost 2 years already.
So point me to actual CPUs or GPUs in those processes I can buy?

What are talking about again?


Intel have huge lead.

Xilinx Ships Industry's First 20nm All Programmable Product

Company achieves major milestone with first customer delivery of UltraScale ASIC-class programmable architecture


Nov 11, 2013
SAN JOSE, Calif., Nov. 11, 2013 /PRNewswire/ -- Xilinx, Inc. (NASDAQ: XLNX) today announced first customer shipment of the semiconductor industry's first 20nm product manufactured by TSMC, and the PLD industry's first 20nm All Programmable device. Xilinx UltraScale™ devices deliver an ASIC-class advantage with the industry's only ASIC-class programmable architecture coupled with the Vivado® ASIC-strength design suite and recently introduced UltraFast™ design methodology. The UltraScale devices enable 1.5X – 2X more realizable system-level performance and integration for customers, equivalent to a generation ahead of the competition.

Granted its FPGA but TSMC's 20nm process is ready since 2013.

 

[NTMBK]

He intimately knows the value that comes from process nodes in a direct and opposite way to that of Dirk Meyer. Talk about your night and day difference. Dirk knew design, not process. So he had no problem getting rid of the fabs and putting AMD's future in the hands of IC design alone.

If Brian is the anti-Dirk, I wonder whether he will get rid of the IC design and put Intel's future in the hands of the fabs?

 

[Arachnotronic]

If Brian is the anti-Dirk, I wonder whether he will get rid of the IC design and put Intel's future in the hands of the fabs?

Anybody who thinks Intel should just "be a fab" hasn't done the math.

Only 34% or so of TSMC's fab volume is leading edge volume, so that's what...$7 billion/year for ALL 28nm product and maybe $8-9B/year run rate with 20nm + 16FF?

Intel's PC Client Group does $33B+/year and DCG does over $11B. Giving access to this process tech to everybody and pretty much ending its competitive advantage in DCG and significantly damaging the LT prospects of PCCG would be financial suicide.

BK knows design is important...that's why they added that SoFIA chip, built at TSMC, to the roadmap. He's trying to make IA as widespread as possible, even if it means giving TSMC some short-term business before going in-house.

As an Intel shareholder, I will say that I have a lot of confidence that BK has an extremely solid strategy and won't clown Intel like Hector clowned AMD. Intel's problem in mobile has always been one of execution...something that BK will have to fix.

 

[Idontcare]

If Brian is the anti-Dirk, I wonder whether he will get rid of the IC design and put Intel's future in the hands of the fabs?

You are "" about it but I think you are serious (as am I) in thinking about it...if Intel has the best fab technology the world over then why wouldn't they "unlock shareholder value" and take on the foundry market with all guns blazing?

Pretty sure any high ASP IC (discrete GPU for example, or Apple/Qualcomm ICs) would be willing to support the price premiums necessary for Intel to have its margins while giving Nvidia/Apple/Qualcomm their higher performing product that can soundly best the competition (fabbed at TSMC...or worse).

We'll know where the BoD and BK are taking Intel when we hear of what happens with IBM's assets/IP in 3-6 months.

But Intel isn't about to get out of design, that is another one of their core competencies.

Intel has the ability to get into a business model of being the IC design house for both the fabless and the IC design-less. Bring them the idea, pay them to capture it in a world-class IC design, and then pay them to fab it for you in a fab that guarantees you'll out-class your competitors.

Just a question of price, of course.

But BK brings one specific suite of qualifications to the CEO position, and 14nm will be pivotal specifically because of the full node-on-node areal shrink that Intel has engineered itself to have.

 

[Arachnotronic]

Xilinx Ships Industry's First 20nm All Programmable Product


Granted its FPGA but TSMC's 20nm process is ready since 2013.

Yeah, what kind of volumes are they running? If you want to play that game, Intel has been "shipping" Broadwell since September ;-)

 

[witeken]

Like they're now doing on a small scale, they can become a foundry, but it would be dumb to allow competitors to use their leading edge fabs. Intel's huge income comes from the added value of their designs.