Samsung and GLOBALFOUNDRIES Forge Strategic Collaboration to Deliver 14nm FinFET

[AtenRa]

And dont forget we are talking about transistor/cost now, not process cost.

It is really funny that according to Intel Transistor/$ goes down with each process but suddenly everyone else transistor/$ stagnates at 20nm or even worse its rising at 16/14nm.

Yea that Alien technology from Alpha Century really helps Intel, i just wonder why Apple or anyone else even bother with TSMC or Samsung/GloFo at 20nm and bellow and not just go manufacture their ICs to Intel and be done with it.

 

[ShintaiDK]

You said wafer cost is 2x 3x more than what HE said, i dont see the links to support that.

Remember initial since thats what we deal with today. His price estimate way be true far into production. However the required IC design cost still kills it for the wast majority of companies.

 

[ShintaiDK]

And dont forget we are talking about transistor/cost now, not process cost.

It is really funny that according to Intel Transistor/$ goes down with each process but suddenly everyone else transistor/$ stagnates at 20nm or even worse its rising at 16/14nm.

Yea that Alien technology from Alpha Century really helps Intel, i just wonder why Apple or anyone else even bother with TSMC or Samsung/GloFo at 20nm and bellow and not just go manufacture their ICs to Intel and be done with it.

Its all about volume. You need to accept that Intel got a high volume with a revenue than is 2x higher than the rest of the entire MPU industry. Plus they can easily afford multibillion IC design cost that is a requirement for lower transistor cost.

Thats the difference.

 

[witeken]

Intel's yields are still unbeaten, and they will scale aggressive at 14nm and 10nm, unlike TSMC and Samsung.

 

[AtenRa]

ok guys you sold me, im buying Alien technology tomorrow

 

[witeken]

ok guys you sold me, im buying Alien technology tomorrow

Great idea .

 

[AtenRa]

Just a quick one, at what yields does Intel measure this Transistor/$ ??
Because i dont see yields in to the equation :whiste:

 

[witeken]

It should be calculated in the $/mm², but that probably didn't happen. But I guess Intel's yields are about the same, or that's their goal. TSMC on the other hand is rumored to have a 50% yield for 20nm, since Apple was willing to way for that. Intel isn't, so they delayed 14nm.

 

[tential]

And dont forget we are talking about transistor/cost now, not process cost.

It is really funny that according to Intel Transistor/$ goes down with each process but suddenly everyone else transistor/$ stagnates at 20nm or even worse its rising at 16/14nm.

Yea that Alien technology from Alpha Century really helps Intel, i just wonder why Apple or anyone else even bother with TSMC or Samsung/GloFo at 20nm and bellow and not just go manufacture their ICs to Intel and be done with it.

Because there is a massive initial investment cost?

 

[ShintaiDK]

Companies that can afford a proper IC design gets lower cost. But as Handel says, thats in the billion $ range. More or less excluding everyone besides Intel, Samsung, Qualcomm and Apple.

Everyone else gets higher cost.

 

[AtenRa]

It should be calculated in the $/mm², but that probably didn't happen. But I guess Intel's yields are about the same, or that's their goal. TSMC on the other hand is rumored to have a 50% yield for 20nm, since Apple was willing to way for that. Intel isn't, so they delayed 14nm.

Sorry but they dont say anything about Yields, the first box is Normalized Capital Growth. That one has nothing to do with yields.

 

[AtenRa]

Because there is a massive initial investment cost?

For what ?? could you explain ??

 

[witeken]

I literally said they probably didn't include yields, now that you've posted a higher resolution image and I decided to look at it more carefully, you're right.

To answer your question: they measure the price/transistor at a 100% yield, since the graph only shows the result of changes in wafer price and transistor size.

 

[Khato]

Here's the math behind the graph:

Handel said the gate utilization is an issue because of limitations of the design tools and parasitic effects. “The other factor is parametric yields, which are strictly tied into leakage control for the 20nm and of course for the 16nm FinFETs,” he said. “You can break this. Intel has shown that it can be broken and of course that’s an excellent achievement. But, it’s based on very high design costs, potentially $1 billion per design, so you need $10 billion in revenue. It also takes a number of years,” he said. He noted that, in the smartphone market, designs move very fast. “You can’t make that kind of investments in designs.”

Putting these two together is quite useful for those not as familiar with the current state of process technology. Namely, from the image note how the potential gates/mm^2 (kU) goes from 3946 at 28nm to 6992 at 20nm and 12391 at '16/14nm', however at the same time the gate utilization using the standard design tools drops from 76.3% at 28nm to 64.7% at 20nm and 54.2% at '16/14nm' and the parametric yield drops from 86.7% at 28nm to 72.8% at 20nm and 60.9% at '16/14nm'. That's why the effective gates/mm^2 (kU) - accounts for utilization using the standard design tools and yield impact) only increases from 2610 at 28nm to 3293 at 20nm and 4090 at '16/14nm'. And that's a somewhat favorable analysis of the '16/14nm' process potential.

Meanwhile the commentary regarding Intel is with respect to how they don't follow the same trend of diminishing gate utilization and parametric yield hits. Sure they also take a hit, but due to their volumes and ability to place greater restrictions upon their own designs they can invest the money necessary to minimize those effects and easily recoup it. My impression is that it was this iterative process of perfecting the layout to mask transition that was so troublesome for them at 14nm - makes sense when you consider the fact that they're patterning 14nm features using 193nm immersion lithography. At that point the 'light' has to be manipulated in such quirky fashion that, well, the mask for simply drawing a straight line wouldn't look much like a straight line. And even with double patterning, when you have two features in close proximity both have to be taken into account with respect to the mask. Now I'm sure TSMC and Samsung are taking the simple steps, but neither of them want to invest the absurd amount of money that would be necessary to enable all of their customers - no, they're content with the trend of those customers who have adequately deep pockets using them while waiting for EUV to temporarily solve the problem.

 

[Fjodor2001]

I think the most interesting info to take away from this discussion is that Intel now only will have a ~1 year lead on 14 nm compared to the competition. It really surprises me how fast the competition has caught up.

Also, I think it will be hard for Intel to sustain the R&D costs for new nodes unless they find some large company to be foundry for. All the other semiconductor companies (Apple, Samsung, Qualcomm, AMD, ...) can indirectly share process tech R&D costs through TSMC/GF, but Intel has to cover all its R&D costs by itself.

 

[ShintaiDK]

I think the most interesting info to take away from this discussion is that Intel now only will have a ~1 year lead on 14 nm compared to the competition. It really surprises me how fast the competition has caught up.

Also, I think it will be hard for Intel to sustain the R&D costs for new nodes unless they find some large company to be foundry for. All the other semiconductor companies (Apple, Samsung, Qualcomm, AMD, ...) can indirectly share process tech R&D costs through TSMC/GF, but Intel has to cover all its R&D costs by itself.

Its only 1 year in PR talk. Real world it stays at around 3½ years.

The R&D cost isnt the issue. The big issue is the ROI on the node. And TSMC and Samsung is feeling the pressure there. Specially when there is no cost saving for the wast majority of customers to go below 28nm due to IC design cost. Apple and Qualcomm will only carry foundries like TSMC so far.

 

[jpiniero]

I think the most interesting info to take away from this discussion is that Intel now only will have a ~1 year lead on 14 nm compared to the competition. It really surprises me how fast the competition has caught up.

Can't really compare the densities now esp since Samsung and TSMC have messed with the naming for marketing reasons.

 

[Fjodor2001]

Its only 1 year in PR talk. Real world it stays at around 3½ years.

According to earlier predictions in this thread Apple will release 14 nm products in 2015Q4, and Intel in 2014Q4. That's a one year difference.

The R&D cost isnt the issue. The big issue is the ROI on the node. And TSMC and Samsung is feeling the pressure there. Specially when there is no cost saving for the wast majority of customers to go below 28nm due to IC design cost. Apple and Qualcomm will only carry foundries like TSMC so far.

Same problem for Intel then. They have to get ROI for their 14 nm R&D investments too.

The problem is actually worse for Intel, since they have no other major company to share the process tech R&D costs with to get satisfying ROI.

 

[ShintaiDK]

According to earlier predictions in this thread Apple will release 14 nm products in 2015Q4, and Intel in 2014Q4. That's a one year difference.

But its not real 14nm or 16nm products. Its basicly 20nm with FF. Funny thing is that Samsungs 14nm is even worse than TSMCs 16nm. We could simply rename Intels 14nm to 10nm if you wish to play that game.

Same problem for Intel then. They have to get ROI for their 14 nm R&D investments too.

The problem is actually worse for Intel, since they have no other major company to share the process tech R&D costs with to get satisfying ROI.

Intel got much higher revenue and more volume on a newer node. GloFo gave up at 14nm. The question is if Samsung or TSMC gives up at 10nm. There isnt really many you can share with either. And IBM as a partner have been a complete disaster.

Note this is the revenue on a node for all the foundries minus Intel combined.

 

[Khato]

I think the most interesting info to take away from this discussion is that Intel now only will have a ~1 year lead on 14 nm compared to the competition. It really surprises me how fast the competition has caught up.

Also, I think it will be hard for Intel to sustain the R&D costs for new nodes unless they find some large company to be foundry for. All the other semiconductor companies (Apple, Samsung, Qualcomm, AMD, ...) can indirectly share process tech R&D costs through TSMC/GF, but Intel has to cover all its R&D costs by itself.

But it's just as logical to take the opposite stance is it not? Where TSMC and Samsung encounter similar issues to Intel with ramping up their "16/14nm" processes to mass production and end up even further behind even with their creative node naming. Note that neither of them are even to the point in the process deployment timeline where Intel announced that they weren't able to ramp yields as quickly as expected.

Meanwhile on the matter of R&D cost structures you're trying to claim that Intel is at a disadvantage because all of the customers share the process tech R&D costs... But what happens when you only have a small subset of customers upon which to place that burden? And even worse, when that small subset switches from one foundry to the other randomly? For example, if the hints at the TSMC conference call are correct and they lost Apple on the "16/14nm" node then who exactly is subsidizing their "16 FinFET" node? Not to mention, who is going to fill up all that freshly built capacity? Especially if Qualcomm starts spreading their manufacturing to Samsung as well. If anything being reliant upon a small number of customers in such a competitive, high-capital environment makes the coming nodes far more dangerous for the foundries. (Have to remember that, unlike previous nodes, only customers who actually benefit from the power/speed improvements have any incentive to move below 28nm and that doesn't apply to the majority of foundry customers, as evidenced by TSMC's continued volume on 40nm and larger nodes.)

Even if you want to ignore the volatility associated with the customer base you're still going to be left with the fact that the total foundry revenue across TSMC, Samsung, and GF doesn't match Intel. Now sure it's not exactly a 'fair' comparison as design costs must also be accounted for with Intel, but even doing so you're left with the capability to invest as necessary. Of course with Samsung the same argument can be made, just depends upon whether or not they feel it's a good investment for the company. Whereas with Intel, well, that's their business model so it's pretty much a given that they're going to invest as necessary.

 

[ShintaiDK]

A few more facts from the report that Handle covers:

IBIS further predicts that:
- 28nm will have long lifetime ( Probability 80% )
- 20nm yields will improve, and will be high volume technology node in 2015 and 2016 (Probability 50%)
- 16/14nm will provide low cost gates and support high bandwidth interfaces in SoC environment (Probability 20% in 2016 and 50% in 2017)
- 10nm will likely be postponed, cost per gate will be prohibitive and unclear as to demand other than high speed processors and FPGAs (Probability 90%) .

Note tho still, that lower cost per gate below 28nm requires massive investment in IC design. Else the cost increases.

They basicly say 90% chance that only Intel can deliver 10nm and below with profit.

 

[Nothingness]

Great idea .

How many times did that guy claim Conroe was coming again? He's lost all credibility long ago.

 

[Homeles]

How many times did that guy claim Conroe was coming again? He's lost all credibility long ago.

Not sure why witeken is so fond of quoting him as of late. I'd like to pretend he doesn't exist.

He must be where AMD's marketing team derives their inspiration...

 

[witeken]

According to earlier predictions in this thread Apple will release 14 nm products in 2015Q4, and Intel in 2014Q4. That's a one year difference.

How many times do we have to repeat this?

TSMC/Samsung 28nm = 40nm shrinked + 1st gen HKMG
TSMC/Samsung 20nm = 28nm shrinked + 2nd gen HKMG
TSMC/Samsung 14nm = 20nm + FinFET

Intel 32nm = 45nm shrinked + 2nd gen HKMG
Intel 22nm = 32nm shrinked + FinFET
Intel 14nm = 22nm shrinked + 2nd gen FinFET

=> TSMC/Samsung 14nm ~ Intel 22nm

They're 1 year later with the node they call 14/16nm, but it's about half a node less dense and only the first iteration of FinFET. If this node wanted to be competitive, it should have been released in 2013, 1 year after Intel released its 22nm, so it will be 2.5+ years too late (for mass product availability).

 

[witeken]

Not to mention, who is going to fill up all that freshly built capacity? Especially if Qualcomm starts spreading their manufacturing to Samsung as well.

A few more facts from the report that Handle covers:

They basicly say 90% chance that only Intel can deliver 10nm and below with profit.

Certainly if Intel starts gaining loads of mobile market share with their 10nm products. Who is going to fill up all the (10nm) fab capacity when Qualcomm's revenues evaporate and Apple switches to Intel Architecture?