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

[Exophase]

I view TSMC 16 FF+ like I view 28nm HPm v.s. 28nm polysilicon.

I think 16 FinFET will ramp in volume during 2016 from guys like Qualcomm, then 16 FinFET+ comes a year later (which is actually more aggressive than what the 28nm poly-si v.s. 28nm HKMG delta looked like).

Someone correct me if I'm wrong here, but I always thought 28HP and HPL were HKMG processes, and not just HPM. 28HP was first used by discrete GPUs starting from the end of 2011.

EDIT: Okay, yeah, 28HP and HPL use HKMG: http://www.nu-vista.com:8080/download/brochures/2011_28 Nanometer Process Technology.pdf FWIW, HPL was used by Tegra 4.

 

[Homeles]

I view TSMC 16 FF+ like I view 28nm HPm v.s. 28nm polysilicon.

I think 16 FinFET will ramp in volume during 2016 from guys like Qualcomm, then 16 FinFET+ comes a year later (which is actually more aggressive than what the 28nm poly-si v.s. 28nm HKMG delta looked like).

That's... actually a really clever observation, IMO. Interesting.

exactly. Intel's actual products need to show the signs of process leadership.

Yep. Their Core designs definitely do... but the mobile side has yet to show it so far. Even Avoton is rather bloated.

Everything I've heard about their 14nm products is that they've corrected those issues... so we'll see soon enough.

Never make calculations before you drink your coffee in the morning
Everything is wrong not only the Intel numbers
I will revise them all.

The same thing happened with me at work this morning. God bless stimulants, when they kick in.

Someone correct me if I'm wrong here, but I always thought 28HP and HPL were HKMG processes, and not just HPM. 28HP was first used by discrete GPUs starting from the end of 2011.

EDIT: Okay, yeah, 28HP and HPL use HKMG: http://www.nu-vista.com:8080/download/brochures/2011_28%20Nanometer%20Process%20Technology.pdf FWIW, HPL was used by Tegra 4.

Yeah, LP is the only one that uses a poly gate.

 

[jdubs03]

The difference is you are quoting an Intel uber bull Ashraf Eassa who has no idea of semiconductor manufacturing while I am quoting TSMC's management on a earnings conference call. Who do you want to believe is upto you ?

TSMC is on record stating that the vast majority of volume production of FINFET production in 2015 is at TSMC 16FF+. TSMC 16FF+ looks likely to go into production by Q2 2015 (analysts confirm the same) so Apple can make the A9 at TSMC 16FF+. Also most analyst reports are stating that Apple is looking at multi sourcing A9 from TSMC and Samsung/GF . The 14nm GF licensing from Samsung is to help Apple avoid wafer shortages. You can choose not to believe them. As usual end customer products will reveal if TSMC was right or wrong. :thumbsup:

TBH I would want a 16FF+ A9 as the iphone 6s is likely my next phone purchase. So hopefully you're right.

 

[Arachnotronic]

The difference is you are quoting an Intel uber bull Ashraf Eassa who has no idea of semiconductor manufacturing while I am quoting TSMC's management on a earnings conference call. Who do you want to believe is upto you ?

All companies spin. Intel, TSMC, everybody. You seem to turn a blind eye to TSMC's historical roadmap slips, but you will be very critical of Intel. Interesting.

 

[raghu78]

Yep. Their Core designs definitely do... but the mobile side has yet to show it so far. Even Avoton is rather bloated.

The competition from AMD is rather weak so the best time to compare would be Intel's Skylake vs AMD's next gen cores based x86 SOC in 2016.

Everything I've heard about their 14nm products is that they've corrected those issues... so we'll see soon enough.

we will need to see how Intel Cherrytrail fares against Apple A8/ Snapdragon 810 and Broxton against A9/FINFET Snapdragons before concluding that Intel has done so.

 

[Arachnotronic]

The competition from AMD is rather weak so the best time to compare would be Intel's Skylake vs AMD's next gen cores based x86 SOC in 2016.

Skylake comes in 2015 (likely Q2 for the desktop parts, and Q3/Q4 for the ULT/ULX parts w/ on-die PCH). How about comparing the 2016 Cannonlake to AMD's next gen X86 products in 2016? ;-)

 

[Homeles]

TBH I would want a 16FF+ A9 as the iphone 6s is likely my next phone purchase. So hopefully you're right.

Call me crazy, but I think smartphones have gotten to the point where they're "fast enough." There's obviously a lot more that can be done for them, but cameras, storage speed, RAM, and processing speed have gotten to the point where they're good enough for what I use it for... light gaming, web browsing, communications, and the occasional photo. I don't see my use cases changing over time, either.

 

[jdubs03]

Call me crazy, but I think smartphones have gotten to the point where they're "fast enough." There's obviously a lot more that can be done for them, but cameras, storage speed, RAM, and processing speed have gotten to the point where they're good enough for what I use it for... light gaming, web browsing, communications, and the occasional photo. I don't see my use cases changing over time, either.

I completely agree. Currently I have a Lumia 928 (iphone 4 before that), so I have time to wait, and then yea it's future-proof for a long time.

Skylake comes in 2015 (likely Q2 for the desktop parts, and Q3/Q4 for the ULT/ULX parts w/ on-die PCH). How about comparing the 2016 Cannonlake to AMD's next gen X86 products in 2016?

Skylake will still likely be the Intel chip at that time, but it'd be easier to compare that to a 16FF+ AMD new x86 core rather than cannonlake (to at least try and normalize process impacts, to isolate uArch comparisons).

we will need to see how Intel Cherrytrail fares against Apple A8/ Snapdragon 810 and Broxton against A9/FINFET Snapdragons before concluding that Intel has done so.

Cherry Trail should have leadership in GPU, with 16 Gen8 EU's, depending onn clocks surpassing HD4600. Idk about CPU though, my guess is it'll be at the top for multi-core, and highest ST of the quad-cores.

 

[raghu78]

Skylake comes in 2015 (likely Q2 for the desktop parts, and Q3/Q4 for the ULT/ULX parts w/ on-die PCH). How about comparing the 2016 Cannonlake to AMD's next gen X86 products in 2016? ;-)

I wouldn't count on Skylake being available 5 - 6 months after Broadwell. More like a year which is how Intel's tick tock works. So Skylake should be available towards end 2015 and the 10nm Cannonlake products from Intel won't be out before Q2 2017 given that Intel's ivybridge launched in Apr 2014 and Broadwell launches in Oct 2014. Node shrinks are not getting any easier .

 

[Arachnotronic]

I wouldn't count on Skylake being available 5 - 6 months after Broadwell. More like a year which is how Intel's tick tock works. So Skylake should be available towards end 2015 and the 10nm Cannonlake products from Intel won't be out before Q2 2017 given that Intel's ivybridge launched in Apr 2014 and Broadwell launches in Oct 2014. Node shrinks are not getting any easier .

Intel's own executives claimed 10nm in HVM by Q4 2015 at its analyst meeting. Does "node shrinks aren't getting any easier" apply only to Intel and not to TSMC/Samsung/etc.?

 

[carop]

Yes, I read it, and I think it does a poor job refuting the claims TSMC made that their 16nm to 10nm will have a higher density improvement than Intel said. Regardless of how either graphs look, TSMC made it clear that they intend to pick up some of the slack in density at 10nm due to not improving the BEOL at 16nm. Which seems plausible given that they'll have had more time than usual between BEOL improvements.

Cliff Hou of TSMC provided some details during DAC 2014. These details explain why TSMC 10nm will be denser. They intend to use spacer multiple patterning to deliver 40nm metal pitches:

In lithography, Hou said TSMC would push immersion approaches as far as possible, using spacer technology to deliver 40nm metal pitches for 10nm, and double spacer technology to achieve 30nm pitch for 7nm. TSMC is also considering directed self-assembly as an option at 16 nm half pitch.

http://www.techdesignforums.com/blog/2014/06/05/cliff-hou-tsmc-vp-rd-10nm-7nm-3dic/

Using the usual 0.7x scaling, the (foundry) 10nm node will have 44nm - 48nm metal pitches. For example, the common platform published the details of their 10nm node during VLSI 2014. Samsung-GlobalFoundries will have 48nm metal pitches at their 10nm node.

As for Intel, it goes without saying that you need the Intel specs to make a comparison. Or, you will have to wait until Intel parts are spotted in the wild so that they can be analysed.

 

[witeken]

Thats your belief while the fact is the reverse is happening. TSMC and Samsung are reducing the process gap and getting more aggressive at future nodes. The foundries will continue to grow at Intel's expense. The foundries will enable Intel competitors like Apple, Qualcomm, AMD, Nvidia to compete and gain market share in the respective businesses that they compete. The war is between the entire semiconductor tech industry and Intel and make no mistake Intel will not win.

Facts are backed up by sources, like I did, so this story is your own belief. The real fact is that Intel's been leading the semiconductor industry with a 2-4 year lead (node and transistor technology respectively) for the past decade or so. The fact is also that Intel is the most wealthy semiconductor company and costs are only increasing. While Intel likely goes to 450mm wafers at the 7nm node, it's no so sure if any other company will ever be able to do that too, probably except Samsung, but not within a short time anyway.

Intel's been leading the industry with its high performance transistors and now they're going to do the same with density. Due to the increasing costs, TSMC and other foundries can only go for longer upgrade cycles.

Intel will have a gap of roughly 30 months from the time ivybridge was first available (Apr 2012) and broadwell is first available (Oct 2014). Future nodes are not getting easier. So the first 10nm products should be available in Q2 - Q3 2017. 7nm products won't be available till late 2019 or early 2020.

Your mid-2017 and 2020 predictions for 10nm and 7nm are about as far away from reality as you can get.

 

[Arachnotronic]

Facts are backed up by sources, like I did, so this story is your own belief. The real fact is that Intel's been leading the semiconductor industry with a 2-4 year lead (node and transistor technology respectively) for the past decade or so. The fact is also that Intel is the most wealthy semiconductor company and costs are only increasing. While Intel likely goes to 450mm wafers at the 7nm node, it's no so sure if any other company will ever be able to do that too, probably except Samsung, but not within a short time anyway.

Intel's been leading the industry with its high performance transistors and now they're going to do the same with density. Due to the increasing costs, TSMC and other foundries can only go for longer upgrade cycles.



Your mid-2017 and 2020 predictions for 10nm and 7nm are about as far away from reality as you can get.

To add to your point, I offer yet more quotes from Intel's management on a conference call (since everything executives say on calls is gospel):

Christopher Roland - FBR
Okay. Great. Thank you for that. Also on the Broadwell quarter push out, do we make up for that at some point in the future, or would this also push Skylake?

Brian Krzanich - Chief Executive Officer, Director
We do not think it will push Skylake, so this is just - and in fact will even make up in total volume within a quarter or two after the one quarter push, so we tend to make these up very quickly. The two are independent. The early learning that's going on during the process development has no impact on Skylake's ability to come to market.

http://seekingalpha.com/article/174...-results-earnings-call-transcript?part=single

 

[witeken]

So we're clear, when you talk about Intel launching a product you mean that product is on the shelves, right?

Correct.

I remember when Ivy Bridge slipped a few months, putting its release around 28 months after Clarkdale. And if you go back to 45nm, the delta between Penryn and Clarkdale was about 26 months.

After IB came out people were saying the same thing you're saying now, that Broadwell will release 24 months after when IB was supposed to come out, putting the actual delta at around 20 months. Instead it's looking more like it'll be out September at the earliest, December at the latest, putting the delta around 28-31 months after when IB did come out. Which is a very different result. So your estimations of 24 months after when Broadwell "should" have come out (I assume about now) look very optimistic and goes against the trend of Intel's product launches at new nodes growing in distance instead of shrinking.

That's certainly possible, but as far as I know, 10nm is still planned for 2015, nothing other than Broadwell/14nm has been delayed and Intel still wants to relentlessly pursuit Moore's law. That guides me to the conclusion for a H1/mid-2016 release for 10nm and information from TSMC itself guides me to an EOY 2018 or later release of their 10FF node. Since the available information guides me to the fact that that node competes with Intel's 14nm, I come to the conclusion that Intel's lead is only growing. I'd like to be proven wrong, but I haven't seen anything suggesting that.

 

[Arachnotronic]

Correct.


That's certainly possible, but as far as I know, 10nm is still planned for 2015, nothing other than Broadwell/14nm has been delayed and Intel still wants to relentlessly pursuit Moore's law. That guides me to the conclusion for a H1/mid-2016 release for 10nm and information from TSMC itself guides me to an EOY 2018 or later release of their 10FF node. Since the available information guides me to the fact that that node competes with Intel's 14nm, I come to the conclusion that Intel's lead is only growing. I'd like to be proven wrong, but I haven't seen anything suggesting that.

witeken

I would imagine that 10nm parts go into production by the end of 2015, but I would expect first Cannonlake based parts on the shelves during 1H 2016. If Intel can really pick up the pace, then high end Atoms could end up on shelves in 1H 2016 (and that'd be a game-changer), but for now I'm going to keep expectations relatively modest and expect 2H launch for 10nm Atoms.

p.s. Skylake-ULX is already showing up on Zauba.
https://www.zauba.com/import-skylake-hs-code.html

 

[raghu78]

Facts are backed up by sources, like I did, so this story is your own belief. The real fact is that Intel's been leading the semiconductor industry with a 2-4 year lead (node and transistor technology respectively) for the past decade or so. The fact is also that Intel is the most wealthy semiconductor company and costs are only increasing. While Intel likely goes to 450mm wafers at the 7nm node, it's no so sure if any other company will ever be able to do that too, probably except Samsung, but not within a short time anyway.

Fact is Intel's node naming means nothing. Their 22nm FINFET has no significant density advantage over foundries 28nm. Transistor performance is superior due to FINFET but that too goes away at 14 nm. But importantly when we compare actual products to see if Baytrail beats A7 in perf/sq mm or perf/watt, it does not.

Intel's revenues shrunk for 2 straight years while TSMC is growing rapidly at 15 - 20% Y-O-Y. Intel's PC business shrunk by 2.5 billion dollars from 2011 to 2013. Why didn't Intel's so called silicon process leadership stop the marketshare loss to ARM ecosystem.

Intel's been leading the industry with its high performance transistors and now they're going to do the same with density. Due to the increasing costs, TSMC and other foundries can only go for longer upgrade cycles.

The foundries know what they are up against. Thats why they are investing aggressively at the leading edge. TSMC 16FF+ matches Intel 14nm in performance. You need not believe it. But we can evaluate final products in the market for the truth.

Apple A8 vs Intel Cherrytrail vs Snapdragon 810
Apple A9 vs Intel Broxton vs FINFET Snapdragon

take your bets. I know I am backing ARM/Apple/Qualcomm/TSMC/Samsung

Your mid-2017 and 2020 predictions for 10nm and 7nm are about as far away from reality as you can get.

come on. Its June 2014 and broadwell is well past the 2 year timeline from ivybridge launch. The traditional 2 year cycle is getting difficult as ivy to broadwell shows.

http://www.anandtech.com/show/5771/the-intel-ivy-bridge-core-i7-3770k-review

As always actual products and not Intel or foundry presentations will speak the truth. same holds good for TSMC. You can evaluate TSMC by Apple and Qualcomm mobile products availability as they are the companies who get the first wafer allocations ahead of AMD/Nvidia.

 

[ShintaiDK]

If TSMC starts risk production of 10nm in 2017. That means 2019 for products.

Intel is already running what TSMC defines as 10nm risk today.

 

[ViRGE]

I've already handed out a couple of infractions and I'd rather not hand out more. Please separate your feelings for other posters from the facts at hand. Other posters and their beliefs are NOT up for discussion
-ViRGE

 

[witeken]

You did the math correct for TSMC 16 FinFET (TSMC reported 0.07um^2), though your Intel number seems off.

A 2.2x reduction is 1/(2.2) meaning that you MULTIPLY by 0.455 not 0.55, leading you to a 14nm SRAM cell size for Intel of ~0.042um^2. If this is actually true, then this is denser than the 10nm SRAM cell size that Samsung/GloFo/etc. reported for their 10nm node at the recent VLSI conference (0.052um^2).

I eagerly await proper process disclosures from Intel.

So Intel's 14nm is 20% more dense than TSMC's 10nm? And almost 2x as dense as TSMC's 20nm. That's more than according to Intel's slide, which states that 14nm is 10% less dense and 50% more dense than 10nm and 20nm.

 

[witeken]

Fact is Intel's node naming means nothing. Their 22nm FINFET has no significant density advantage over foundries 28nm. Transistor performance is superior due to FINFET but that too goes away at 14 nm. But importantly when we compare actual products to see if Baytrail beats A7 in perf/sq mm or perf/watt, it does not.

I don't think it will go away at 14nm, since that's Intel's 2nd generation FF vs 1st gen FF. When you compare performance/watt, Intel is clearly superior by a factor of ~2 (BT vs A7).

 

[witeken]

If TSMC starts risk production of 10nm in 2017. That means 2019 for products.

Intel is already running what TSMC defines as 10nm risk today.

TSMC will start risk production of 10nm at the end of 2015.

 

[Fjodor2001]

TSMC will start risk production of 10nm at the end of 2015.

How do you know it's the end of 2015? They have only said 2015, so it can be any time that year.

 

[witeken]

How do you know it's the end of 2015? They have only said 2015, so it can be any time that year.

"The good news is TSMC believes it can offer the full advantages of a new process node with a 10nm FinFET process, including a doubling in the density of logic gates relative to the 16nm node. Designers could start early work in the 10nm process before the end of 2015, Sun said."

http://seekingalpha.com/article/196...semiconductor-have-very-different-definitions

 

[ShintaiDK]

TSMCs roadmaps says it all. And they havent for as long as I can remember been able to deliver on time. Plus they more or less randomly change the nodes. There is no CLN20G node anymore for example.

 

[witeken]

Q3'12 risk production -> Q1'15 products = 2.5 years
Q4'15 risk production + 2.5 years = Q2'18 products

Give or take 1-2 quarters.