Samsung Claims Mass Production on 14-Nanometers

[NostaSeronx]

14nm FinFETs from the Common Platform Alliance and sub-tier members. Will not have the mainstream value as the 28nm node till the 1H of 2017.

CPA and sub-tier members;
IBM(Now GF)
UMC
SEC
GF
SMIC

Mainstream value = Yields, performance, and power, or total cost of production. Overall, anyone producing FinFETs at the CPA/Sub-tier foundries will be in the red till 1H of 2017.

From my investigation for leading edge @ $10 to $500 selling rate for ASICs;
28nm bulk nodes at CPA/sub-tier -> Black (currently)
28nm FDSOI/SSRW planar nodes at CPA/sub-tier -> Blackest (currently)
20nm bulk nodes at CPA/sub-tier -> Red-ish (currently)
14nm FinFET nodes at CPA/sub-tier -> Hemorrhage Red (currently)
14nm FDSOI/SSRW (20nm) planar nodes at CPA/sub-tier -> Black-ish (currently)

~1 year from now;
28nm bulk nodes at CPA/sub-tier -> No data
28nm FDSOI/SSRW planar nodes at CPA/sub-tier -> Black
20nm bulk nodes at CPA/sub-tier -> Black-ish
14nm FinFET nodes at CPA/sub-tier -> Hemorrhage Red
14nm FDSOI/SSRW (20nm) planar nodes at CPA/sub-tier -> Blackest

FinFET includes FDSOI and SSRW as well as bulk undoped.

 

[III-V]

Well, except that they're "on track."

Exactly.

 

[krumme]

Well, except that they're "on track."

Doesnt say much if anything. From my perspective that doesnt mean anything more than that prior claims about 1 year from 20nm to finfet is still valid. They have the small production to back up that claim.
I think the question is at what cost and where the finfet will be used.
It surprices me to see apple use so strong cores now and so new 20nm tech. Its a change. Aparently the highend market and consumers is highly desirable and profitable.
Samsung r&d in process tech must be crazy high - evident from the total record r&d in tech.
A guess for reason : they must think the phones will be replacing the computers?

 

[krumme]

14nm FinFETs from the Common Platform Alliance and sub-tier members. Will not have the mainstream value as the 28nm node till the 1H of 2017.

CPA and sub-tier members;
IBM(Now GF)
UMC
SEC
GF
SMIC

Mainstream value = Yields, performance, and power, or total cost of production. Overall, anyone producing FinFETs at the CPA/Sub-tier foundries will be in the red till 1H of 2017.

From my investigation for leading edge @ $10 to $500 selling rate for ASICs;
28nm bulk nodes at CPA/sub-tier -> Black (currently)
28nm FDSOI/SSRW planar nodes at CPA/sub-tier -> Blackest (currently)
20nm bulk nodes at CPA/sub-tier -> Red-ish (currently)
14nm FinFET nodes at CPA/sub-tier -> Hemorrhage Red (currently)
14nm FDSOI/SSRW (20nm) planar nodes at CPA/sub-tier -> Black-ish (currently)

~1 year from now;
28nm bulk nodes at CPA/sub-tier -> No data
28nm FDSOI/SSRW planar nodes at CPA/sub-tier -> Black
20nm bulk nodes at CPA/sub-tier -> Black-ish
14nm FinFET nodes at CPA/sub-tier -> Hemorrhage Red
14nm FDSOI/SSRW (20nm) planar nodes at CPA/sub-tier -> Blackest

FinFET includes FDSOI and SSRW as well as bulk undoped.

You make serious profit selling phones with a name. Not producing soc.
Apple shows there is a highend market driving 14nm and fin fet forward.
But where is the market for soi?
With a a53 you go dirt cheap 28nm lp for a 2 cluster 8 core as shown by qcom. That is the low end and where does that place the midrange?

 

[TurtleCrusher]

None of us know
Who here knew Athlon would be better than P4 with HT?

Well, benchmarks got it wrong back then. The benchmarks clearly showed the A64 pummeling the P4 HTs. What they didn't test for was multitasking and input delay. When the CPU was heavily loaded, input delay from my logitech and xbox controller was huge. Mouse and keyboard was less, but that's likely an OS thread priority issue.

I had both an Athlon 3500+ and a P4 2.8 W/Ht. For daily usage I wish I never made the jump. It wasn't until I got my hands on an x2 4200+ was the system snappy again during multitasking, and then it really was.

 

[Idontcare]

Yield is damn secret. So ofcource nobody knows.

Yield is a sensitive topic in the industry but not for the reasons I suspect most people think. A yield number is meaningless without a handful of qualifiers, such as die size, density, operating parameters, etc.

If I told you my sram yield is 98% then it should take you all of 1.3 seconds to realize I told you absolutely nothing.

How big is the sram? What is the operating voltage, clockspeed, power consumption? What is the cycle time in the fab (6 month cycle time for 98% is not better than 3 month cycle time for 90% yield), how much did it cost to produce the sram, how much time and money must go into designing the sram on my node's design rules, how reliable is the sram once it goes out to the field (98% yield on chips that die in 2 weeks is not as desirable as 50% yield on chips that last 10+ years).

And so we come to the real reason no one talks straight about yields, they don't want to talk about all the other parameters that you need to also know about in order for you to be able to make heads or tails of the yield number.

Generally what people are willing to discuss is a normalized parameter called D0 (d-zero) which is a defect density number. But again even if I told you my D0 for my fab and process tech all you can do with that number is say "thanks, but I still can't do anything with this info unless you also tell me the following details: what is X, what is Y, what is Z, etc etc".

(and in the case of D0, you only have normalized comparable yield info for functional yield but not parametric yield or lifetime reliability because those are device specific as well as node specific, and very rarely do to companies make the exact same device such that parametric and reliability yield numbers are meaningfully comparable)

 

[krumme]

Yield is a sensitive topic in the industry but not for the reasons I suspect most people think. A yield number is meaningless without a handful of qualifiers, such as die size, density, operating parameters, etc.

If I told you my sram yield is 98% then it should take you all of 1.3 seconds to realize I told you absolutely nothing.

How big is the sram? What is the operating voltage, clockspeed, power consumption? What is the cycle time in the fab (6 month cycle time for 98% is not better than 3 month cycle time for 90% yield), how much did it cost to produce the sram, how much time and money must go into designing the sram on my node's design rules, how reliable is the sram once it goes out to the field (98% yield on chips that die in 2 weeks is not as desirable as 50% yield on chips that last 10+ years).

And so we come to the real reason no one talks straight about yields, they don't want to talk about all the other parameters that you need to also know about in order for you to be able to make heads or tails of the yield number.

Generally what people are willing to discuss is a normalized parameter called D0 (d-zero) which is a defect density number. But again even if I told you my D0 for my fab and process tech all you can do with that number is say "thanks, but I still can't do anything with this info unless you also tell me the following details: what is X, what is Y, what is Z, etc etc".

(and in the case of D0, you only have normalized comparable yield info for functional yield but not parametric yield or lifetime reliability because those are device specific as well as node specific, and very rarely do to companies make the exact same device such that parametric and reliability yield numbers are meaningfully comparable)

Thanx. Makes good sense. Perhaps revealing some of the parameters for understanding yield then also gives away secret product information?

 

[witeken]

@Idontcare:

They could simply disclose the average yield for a number of chips that are going into HVM. The other things you said (like cycle time etc.) to me seem separate issues. If someone asks for the performance of a chip, he isn't interested in power consumption or price (unless he also wanna know those metrics).

 

[Idontcare]

@Idontcare:

They could simply disclose the average yield for a number of chips that are going into HVM. The other things you said (like cycle time etc.) to me seem separate issues. If someone asks for the performance of a chip, he isn't interested in power consumption or price (unless he also wanna know those metrics).

What would you do with the info besides come to the conclusion that you need more info before you can do anything with the yield info?

If I told you right now, today, the yield on 4790's is precisely 89.637%, what can you do with that number?

A real-world example, when we made dram at TI we had a product that was yield-comparable to a lot of other memory IDMs at the time (because the device-specific nature of yield was then normalized across IDMs), and we were beating ourselves up over the apparent yield disparity between us and another high volume leading edge dram supplier at the time (they had higher reported yield than us) until we finally came to understand that they were able to hit their higher yields at the expense of adding so many more clean steps in their process flow such that their cycle time was literally weeks longer than ours.

As a result, our margins were actually better than theirs, despite our lower yields, because our cost to produce the chips was lower than their cost.

At the end of the day there is very little you can do with a yield number if you haven't been given a host of other details regarding the cost structure, which includes performance because that is parametric yield, unless you fancy yourself making a lot of apples-to-oranges comparisons.

 

[witeken]

I'm not going to do anything with the yield info because it's just information. But if we know Samsung's yield is 15%, then we know it will still be quite some time before we get the chips. If the same company (e.g. Intel) makes a comparison of yield between 2 nodes, then we can also have an idea of its health because it will be apples to apples.

 

[Arachnotronic]

I think the yield percentages probably refer to the lead products of major (mobile) customers at these foundries. Think Apple, Qualcomm, Samsung LSI, etc.

 

[krumme]

I'm not going to do anything with the yield info because it's just information. But if we know Samsung's yield is 15%, then we know it will still be quite some time before we get the chips. If the same company (e.g. Intel) makes a comparison of yield between 2 nodes, then we can also have an idea of its health because it will be apples to apples.

As i understand it getting new process nodes to market is a matter of $ and time. We know ss is investing more in r&d than Intel - that was oposite 5 years back.

The finfet and euv is probably a result of that extreme investment. And we must asume it cost big time. Its comming. And next year will be showing the shift for real.

15% is perhaps not a viable result but the matter is what you can do with the process.
What product advantages does it give and what is the benefit for that?
What new market posibilities does that give?
What strategic advantages does it give your customer - and yourself?

Ip6 20nm have just reached market 3 months ago or so. So we are 9 months away from a A9 presumably 14nm.
With the margins Apple have on their product its safe to say that if they think a performance advantage is good for the brand - then paying for a soc with low yield is no problem.
Its all about cost benefit and apple recent behavior svows they go for speed and new nodes to get it.
We get tons of 14nm h2. High performant one imo. Because the money is there and the investment have been done. And 10nm and euv will carry it forward as long as high end mobile pulls.

Instead of this downplaying and unqualified guessing of ss process i think the interesting part is why ss does it. Its a huge risk imo because what about 2016...2019? Its billions and billions of investment done by professionals.

We know mobile is the only place where ss makes some serious profit. The multiple billions. Their power is there. Without it ss would just be a giant local producer of eg trucks with a LG attached.
A thing that can explain it is the strategic goal to be a center of mobile. Be it phones, wearables, iot. And then access to state of the art process is viewed as a nessesity here.

 

[Khato]

As i understand it getting new process nodes to market is a matter of $ and time. We know ss is investing more in r&d than Intel - that was oposite 5 years back.

I'd imagine that I'm not alone in my desire to see the proof behind the above statement. Yes, Samsung Electronics as a whole invests more in R&D than Intel as a whole, and Samsung Semiconductor as a whole invests more in R&D than Intel does in process technology... But recall that Samsung Semiconductor's general logic fabs only account for roughly a third of their revenue - the bulk of their revenue, and by extension R&D, is in DRAM and NAND. I doubt that Samsung is even to the point of out-spending TSMC on an apples-apples R&D comparison.

 

[witeken]

and Samsung Semiconductor as a whole invests more in R&D than Intel does in process technology...

Source?

And

 

[Arachnotronic]

As i understand it getting new process nodes to market is a matter of $ and time. We know ss is investing more in r&d than Intel - that was oposite 5 years back.

Uhh...Samsung might be spending slightly more as a whole in R&D than Intel, but keep in mind that Samsung's R&D includes fun activities such as:

- LCD/OLED/LED display R&D
- Memory process R&D
- NAND process R&D
- Camera sensors, apps processors, modems and connectivity R&D
- Smartphone development
- Tablet development
- Logic processes
- Air conditioners

So yes, if you take the entire R&D budget of Samsung which does everything under the sun in consumer electronics, it's slightly higher than Intel's R&D budget.

But I have little doubt that Intel is outspending Samsung on logic technology, not the other way around as you suggest.

 

[Black Octagon]

And they do more than what you just listed (e.g., medical devices). And that's just Samsung Electronics. The Samsung group does everything from cars to clothes to shipyards

 

[krumme]

I'd imagine that I'm not alone in my desire to see the proof behind the above statement. Yes, Samsung Electronics as a whole invests more in R&D than Intel as a whole, and Samsung Semiconductor as a whole invests more in R&D than Intel does in process technology... But recall that Samsung Semiconductor's general logic fabs only account for roughly a third of their revenue - the bulk of their revenue, and by extension R&D, is in DRAM and NAND. I doubt that Samsung is even to the point of out-spending TSMC on an apples-apples R&D comparison.

I know. But it can never be apple to apple comparison.

I dont care for who uses most on r&d (more on the focus), and who is first for that matter, but where ss beliewes the future business is.

The point was it seems tons of investments is poured into the semiconductor area - accelerating. And as you say the prior revenue is meager as is the ss 20nm volume on the market. Yet finfet seems to be comming h2 big time.

I am just looking for explanations for that.

I am not going more into arguments like "its just for bragging rights" or "yield is 15%". Its boring, repeated to death and brings no new ideas or argument to the table.

It still bogs me what they are going to use all that computing power for in a phone. I use a note 3. I read a lot of heavy pdf. It takes its toll on the qcom s800 and i could use a apple a9 or a a57 on finfet. But man how many user do that? I simply cant see them paying for it. But perhaps we will? Lol.

 

[Khato]

Source?

The charts you posted work Because last I saw the whole of Samsung Semiconductor R&D was around $2.8 B for 2013. Whereas the 'process technology' component of Intel's R&D is only that 'development' slice of the pie, which ends up being a bit over the $2 B mark. (They actually had another breakdown slide in their 2013 investor meeting slides I believe.) But yeah, that goes to show the marked disparity in process technology R&D investment - just that branch of Intel almost matches Samsung Semiconductor's total.

 

[Khato]

I know. But it can never be apple to apple comparison.

I dont care for who uses most on r&d (more on the focus), and who is first for that matter, but where ss beliewes the future business is.

The point was it seems tons of investments is poured into the semiconductor area - accelerating. And as you say the prior revenue is meager as is the ss 20nm volume on the market. Yet finfet seems to be comming h2 big time.

I am just looking for explanations for that.

Heh, it's somewhat telling that Samsung Semiconductor's revenue for Q3 has actually split to about 3/4 memory and 1/4 everything else. Regardless, it's quite clear that your initial statement was drastically in error... which, by the way, is also why most of those who look for facts rather than propaganda don't have the same confidence that Samsung is portraying to investors.

 

[Arachnotronic]

I know. But it can never be apple to apple comparison.

I dont care for who uses most on r&d (more on the focus), and who is first for that matter, but where ss beliewes the future business is.

The point was it seems tons of investments is poured into the semiconductor area - accelerating. And as you say the prior revenue is meager as is the ss 20nm volume on the market. Yet finfet seems to be comming h2 big time.

I am just looking for explanations for that.

I am not going more into arguments like "its just for bragging rights" or "yield is 15%". Its boring, repeated to death and brings no new ideas or argument to the table.

It still bogs me what they are going to use all that computing power for in a phone. I use a note 3. I read a lot of heavy pdf. It takes its toll on the qcom s800 and i could use a apple a9 or a a57 on finfet. But man how many user do that? I simply cant see them paying for it. But perhaps we will? Lol.

My take on what's going on:

1. Samsung once had big profitable smartphone business and very profitable fab business to Apple

2. Smartphones became commodity and Samsung lost Apple as a client

3. Samsung's profitability begins to crater and investors are unhappy

4. Samsung tells investors that their 14nm is going into mass production in Q4, beating out TSMC which has just spent the last 3+ years having a virtual lock on leading edge foundry technology. This is likely to try to convince investors that a big inflection point is around the corner and that Samsung will soon be fabbing tons of chips for all of the major mobile players.

While there is certainly a lot of design activity on Samsung's 14nm process, it remains to be seen which foundries ultimately get the big volumes. It's in the best interests of the fabless players (like Qualcomm, Apple, etc.) to have two strong foundries fighting for business to bring costs down of course, but we'll see how it all pans out.

The next iPhone teardown is going to be fun

 

[witeken]

While there is certainly a lot of design activity on Samsung's 14nm process, it remains to be seen which foundries ultimately get the big volumes. It's in the best interests of the fabless players (like Qualcomm, Apple, etc.) to have two strong foundries fighting for business to bring costs down of course, but we'll see how it all pans out.

Costs only go up. There's a certain foundry that knows this and they're not going to participate in a race to the bottom; they'd rather see Samsung and TSMC do that, seeing their margins and income evaporate so they could increase their own technology lead. And with their bleeding edge yields, they could probably ask a high price, nice margin on their wafers but still offer a competitive offer because there's more useful dice per wafer so the net amount of useful transistors per m² wouldn't even be that much worse, while the rest of the cost deficiency is compensated by their unique technology proposition. I'm just saying Apple will probably move on from both companies in the near or more distant future -- or somewhere in between.