Apple Entrusts TSMC with All of APs for iPhone7? TSMC 10nm in Q2 2016?

[Daniel Nenni]

The news is breaking that TSMC will get the iPhone 7 SoC business. The first story I saw was from Korea Times news:

Apple Entrusts TSMC with All of APs for iPhone7

but it has been repeated many times and is now taken for fact. Here is the best part:

TSMC is planning to enter a state of mass-production system of 10-nano chips starting from June. “TSMC has taken all of Apple’s supplies in 10-nano and such plan was already confirmed in second half of last year.” said a high-ranking executive in EDA industry. A high-ranking representative of semiconductor IP industry said that this story was already known in rear-end industry of semiconductor in Taiwan.

Do you guys really think TSMC 10nm will be in full production in June?

 

[Fred Chen]

The news is breaking that TSMC will get the iPhone 7 SoC business. The first story I saw was from Korea Times news:

Apple Entrusts TSMC with All of APs for iPhone7

but it has been repeated many times and is now taken for fact. Here is the best part:

TSMC is planning to enter a state of mass-production system of 10-nano chips starting from June. “TSMC has taken all of Apple’s supplies in 10-nano and such plan was already confirmed in second half of last year.” said a high-ranking executive in EDA industry. A high-ranking representative of semiconductor IP industry said that this story was already known in rear-end industry of semiconductor in Taiwan.

Do you guys really think TSMC 10nm will be in full production in June?

Interesting that Samsung is now put into panic mode, trying to keep Qualcomm 10nm while accelerating 7nm, and these may compete with Samsung's DRAM business. That's the problem with doing everything.

 

[astilo]

Do you guys really think TSMC 10nm will be in full production in June?

I do not think so Daniel, unless this 10nm is going to be some kind of half node shrink.
As far as I know, it should be instead a full shrink from the 16FF+, requiring some triple patterning (not confirmed) and also a color-aware design (confirmed).

TSMC made a working test chip in its 10nm FinFET process. The process should deliver a 0.52x area scaling compared to 20nm and support either 18% higher speeds or 40% less power than TSMC’s current leading-edge 16FF+.

The official TSMC statement was to be in volume production in 2017, so a 2 quarters pull in would be a sensational achievement.
I believe that the A10 has been already designed at 14/16nm and probably at both Samsung and TSMC sites. What would be instead possible is that the IPad chip will be manufactured at 10nm by TSMC only (maybe this is the source of this speculation).
Let's be honest, with a 60-70% of the total Apple revenue coming from the Iphone only, would you risk such a desperate move? And moreover, why? 16FF+ or 14LPP is going to be more than enough for the next smartphone waves. If TSMC should miss, Apple would be in jeopardy. They are money makers, not gamblers.
Competitors would not release any 10nm product before the end of 2017, so no reason to do that. 90% of the Apple users do not even care about what's inside the phone, at least until it is performing good.
Sure, there is always the possibility to have the A10 designed at both 10nm and 16nm, as backup plan, but still this is too hard for me to understand (at the very best, is a waste of resources).
When released, the Iphone7 will be for a while the fastest phone on the market (as usual), no matters if the chip is built at 10nm or 16nm. What would be the benefit of having an even faster Iphone, when you can already claim you have the best phone out there? They need revolutionary ideas (like in the past), chasing only performances, would bring them nowhere.

 

[Hans de Vries]

I do not think so Daniel, unless this 10nm is going to be some kind of half node shrink.
As far as I know, it should be instead a full shrink from the 16FF+, requiring some triple patterning (not confirmed) and also a color-aware design (confirmed).

It's a full node shrink. Multiple slides from last years ARM Tech Con with first 10nm ARM IP implementation efforts show so:

 

[Daniel Nenni]

The Apple TSMC relationship is a close one. TSMC developed 20nm for Apple and contrary to popular belief TSMC developed a custom version of 16nm for Apple called 16FFT which was used for the A9. So it is certainly possible TSMC is building a custom version of 10nm for Apple that will be released early but word on the street is that "TSMC 10nm" will not be ready in time for the iPhone 7.

My bet is that the iPhone 7 A10 will be an improved version of TSMC 16nm with a whole new architecture. Kind of an Intel Tick-Tock thing. Next year it will be 10nm then 7nm in 2018. Just my bet of course.

 

[Speedfriend]

So Does that mean Intel will be first with 10nm in early 2017? And why is Intel finding it so hard to move from 14 to 10 versus the speed TSMC and Samsung are?

 

[Daniel Nenni]

So Does that mean Intel will be first with 10nm in early 2017? And why is Intel finding it so hard to move from 14 to 10 versus the speed TSMC and Samsung are?

According to what I heard at the IISS Symposium last month Intel 10nm has been delayed until Q1 2018. With TSMC and Samsung launching 10nm in Q4 2016 that puts them a good one year head start. Not all 10nms are created equal however but to most customers 10nm is 10nm.

 

[astilo]

So it looks like the rumor was as usual partially true and partially wrong. A10 might be indeed solely manufactured by TSMC, but at 16nm. It makes much more sense. A9 was already built on 16FF+, so the process is already very reliable, even for a single source approach. On the other hand, capacity is not a problem either, since TSMC lost the snapdragon 820 business.
A10 would have been built on 14LPP at Samsung, while A9 was 14LPE, and 14LPP is moving into volume production right now (creating a potential residual risk for a timely mature yield). Considering Samsung has got already the Qualcomm SD820 business (expected to be in most of the high end smartphones of the year), will manufacture the new Exynos chip and probably is also engaged by AMD, I do not think they had enough 14nm capacity to satisfy the full Apple demand. Probably they had also no big pressure to be competitive in price, while TSMC was clearly desperate to win at least one of the big boys. This is making the Apple choice almost obvious, from a management point of view.
I have seen many articles saying that the Apple decision has been performance driven, that of course is not the case at all. The confusion was generated by the A9 "gate", where a difference in the power consumption between the TSMC chip and the Samsung one has been spotted. What they do not say (most likely because they do not even know that), is that the the A9 has been built on 16FF+ at TSMC (since they dropped the early version due to a lack of performance), while just on 14LPE at Samsung.
Now, the 16FF+ is the enhanced version that should be directly compared to the Samsung 14LPP one.

14LPP (Performance boosted edition) is the 2nd FinFET generation which the performance is enhanced up to 10%. 14LPP is the single platform for every application designs with the improved performance for computing/Network designs and the lowered power consumption for Mobile/Consumer designs.

 

[Brianhayes]

Do you guys really think TSMC 10nm will be in full production in June?

Looks like the stock market thinks so. Shares jumped 5% today. I can find no other rumours that would cause such movement.

 

[Daniel Nenni]

Looks like the stock market thinks so. Shares jumped 5% today. I can find no other rumours that would cause such movement.

Was that a bounce back from the earthquake news?

 

[benb]

It's true. TSMC has been VERY MOTIVATED since Samsung took back Apple at 14nm and Qualcomm's leading chip. This is war, and revenge is sweet. 10nm is arriving in Apple products this fall.

 

[Brianhayes]

It's true. TSMC has been VERY MOTIVATED since Samsung took back Apple at 14nm and Qualcomm's leading chip. This is war, and revenge is sweet. 10nm is arriving in Apple products this fall.

The war, surely, is between Apple and Samsung. Based on all the statements made by Morris Chang I have always believed that the only reason why Samsung got part of the A9 business was that TSMC did not have the capacity in time and this was because they committed themselves (to Apple?) to build capacity at 20nm. As TC has said, Cupertino is more secretive than the CIA, but it seems to me to be a no brainier that they want to build the iOS lead over Andriod (quality not quantity) and part of this is the A9/A10 lead over Exynos and Snapdragon. That, on my hypothesis, is a good reason for their capital contribution to TSMC that Mark Hibben and DAN suggested that they made. If 10nm arrives this Autumn the main victory is Apple's. No evidence for this, just logic.

 

[Daniel Nenni]

I also think this is a "balance of nature" kind of thing for the fabless semiconductor ecosystem. TSMC had 90% market share at 28nm which was completely out of balance. Before 28nm TSMC averaged 40-50% market share if my memory serves...

Also, TSMC had to sign some pretty strict business terms at 20nm to get the Apple business so Qualcomm went to Samsung. TSMC was late with the Apple version of 16nm (16FFT) so Apple had no choice but to split it between TSMC and Samsung. I'm not convinced either Samsung or TSMC had the yield/capacity to single source the A9 anyway. FinFETs and double patterning were quite the challenge, absolutely.

The A10 will be a new and improved version of 16nm unless TSMC is able to pull in 10nm which I highly doubt. It all depends on how much of a PPA (performance/power/area) increase Apple is willing to accept. If it is 10% the A10 could be 10nm. If it is 25% 10nm will be delayed until the 2017.

It will all depend on how fast 7nm will follow. If 7nm is ready for the A11 then 10nm will be rushed out the door at 10% PPA. Remember, 10nm and 7nm will use the same fabs just like 20nm and 16nm so it will be deja vu all over again.

Just my opinion of course.

 

[lefty]

What do you think the chances that Qualcomm will switch back to TSMC once they have enough volume on their 16 FF+ and FFC nodes?

 

[Daniel Nenni]

Very timely question lefty. A QCOM friend of mine said that Dr. Morris Chang himself visited QCOM recently to ask that same question. Based on the history of QCOM I would say no. The QCOMM culture is 2nd, 3rd, and even 4th sourcing of wafers up until 28nm when they were forced to use only TSMC until UMC finally yielded. But Morris is the best wafer salesman in the history of the foundry business so it is certainly possible. It would however upset the balance of nature I mentioned before by having both the #1 and #2 SoC companies at the same fab.

I caught an interesting tidbit of information at the SPIE conference today that may change the foundry landscape. It seems that TSMC will not use EUV for 7nm but Intel and Samsung will. What would happen if Samsung and TSMC both release 7nm at the same time with Samsung having a significant cost/yield/capacity advantage?

View attachment 16665

 

[Fred Chen]

That slide ignores that the cost of EUV per mask is effectively several times that of 193nm at least. It would imply Samsung could not start 18nm DRAM this year, which is not true.

 

[Daniel Nenni]

I asked Samsung directly about EUV strategies for logic versus memory and was told those are two very different things and logic EUV would come before memory. The first reason was because Samsung is the memory leader but a logic follower (for now) and the second reason is that memory will not "require" EUV before logic.

The other thing I can tell you is that the Intel EUV update was very sanitized/lawyerized where as some of the other EUV presentations were much more believable. They are still going on so I will keep asking questions...

 

[Fred Chen]

That's an interesting perspective thanks, Daniel. Looking forward to checking out the papers this year.

 

[Staf_Verhaegen]

That slide ignores that the cost of EUV per mask is effectively several times that of 193nm at least.

What I remember is that mask cost is mainly determined by mask writing time and that EUV mask time writing should be lower than 193nm masks because of less aggressive OPC needed. Where do you get your information from ?

 

[Fred Chen]

What I remember is that mask cost is mainly determined by mask writing time and that EUV mask time writing should be lower than 193nm masks because of less aggressive OPC needed. Where do you get your information from ?

I mentioned cost of EUV per mask rather than mask cost; this would include throughput at the doses required to counter shot noise, as well as projected defect yield (very low right now).