The IDM Business Model is Collapsing!

[Daniel Nenni]

View attachment 16815If I had to describe TSMC in one word it would be clever. TSMC is a VERY clever company, absolutely.

Today’s TSMC Technology Symposium was full of cleverness and packed with executives from all over the fabless semiconductor ecosystem. Tom Dillinger (Chip Guy) will be blogging in more technical detail about the keynotes so let me give you the rambling Apple version of today.

TSMC does not talk about customer orders (especially Apple) but they do talk about capacity and ramp time of new processes and with Apple’s huge capacity requirement and delivery schedule it is easy to make guesses about who those wafers are really for.

So here is my Apple view of the Symposium:

TSMC did confirm that they will double 16nm capacity in 2016 as we talked about here: https://www.semiwiki.com/forum/f293/tsmc-double-16nm-chip-production-7407.html

Apple is a big reason for this but also 16FFC/InFO which will be a game changing node in regards to price/power/performance. If you look at the current foundry landscape you have to wonder if TSMC 16FFC is going to be another 28nm where they left the other foundries behind. TSMC says they have more than 70 new 16nm tape outs this year in addition to 50 last year. The Apple A10 is of course one of those 16nm tape outs thus the double down on capacity.

Bottom line: TSMC has 100% of the Apple SoC business this year (iPhone 7).

TSMC is going to continue with the quick node strategy they started at 20nm. Based on the ramp chart, 10nm will start production in Q4 2016 with a moderate amount of wafers up until Q3 and Q4 2017 where wafers will be flying out of those Giga Fabs at a rate only Apple can explain (Apple A11). TSMC said 10nm tape outs will start this quarter but they did not say how many. I’m told it is 20nm déjà vu all over again with only a handful. Part of it is due to the huge popularity of 16FFC but also due to the quick 7nm ramp that is expected in 2018.

Bottom line: TSMC has 100% of the Apple SoC business next year (iPhone 7s).

The rest of the technical talk was about 7nm which should rival the popularity of 28nm and 16nm. There will be two versions of 7nm, one for mobile and one for HPC (servers). 7nm currently has 30% yield on 128MB SRAM and qualification is planned for Q1 2017, just in time for the Apple A12/ iPhone 8 in 2018. And of course the 10nm fabs will be converted into 7nm fabs the same as 20nm moved to 16nm.

Bottom line: TSMC really wants 100% of the Apple SoC business in 2018 (iPhone 8).

 

[Speedfriend]

Daniel, any view why TSMC have been able to get 10nm right, when Intel has had to delay its 10nm? And where is Samsung relative to the TSMC roadmap?

 

[Daniel Nenni]

Daniel, any view why TSMC have been able to get 10nm right, when Intel has had to delay its 10nm? And where is Samsung relative to the TSMC roadmap?

Yes, here is my view:

TSMC is not bound by Moore's Law like Intel. Moore's Law is a bottom up approach to process development. You start at Moore's Law and go up from there no matter how long it takes.

TSMC on the other hand develops processes top down with customers and partners. At 28nm and above the performance/power/area goals were 25% PPA (Performance/power/area) improvement over the previous node and the timeline was whatever it needed to be to accomplish that.

At 20nm that all changed. TSMC decided to take the conservative approach of doing double patterning on planar devices then move to FinFET at 16nm one year later, which is really 20nm BEOL with FinFETs added. This quick node approach was also used to get the Apple business. Apple wants a new process node every year for the iPhone launch in the fall. Having a similar BEOL also allows TSMC to use the same fabs for 20/16nm which saves on cost/CAPEX of course.

Bottom line: Instead of delivering a 25% PPA for one node TSMC delivered a combined 30% PPA for 20/16nm.

At 10nm TSMC uses quad patterning. Rather than introducing an aggressive BEOL (25% PPA) AND quad patterning, TSMC settled for a 10-15% PPA. At 7nm (one year later) TSMC will add another 10-15% PPA. This also fits with Apple's iPhone timeline. TSMC 10nm and 7nm will both be ramped specifically to make the iPhone fall delivery schedule.

And again like 20nm, TSMC is only working with a select group of 10nm customers (Apple) so ramping is much quicker. And once you yield learn on 10nm it will be much easier to ramp 7nm. 16nm for example was TSMCs fastest ramping process because of the 20nm yield learning and the same will be said for 7nm which also uses quad patterning.

Sound reasonable?

Meanwhile, back at Intel, they missed the Moore's Law window with the 14nm delays and will miss it again at 10nm and 7nm. So instead of every two years of doubling transistor count it will be every three to four years. So yes, Moore's Law really is dead. Which brings me to one of my favorite semiconductor jokes:

The number of people who claim Moore's Law is dead doubles every year.......

D.A.N.

 

[Daniel Nenni]

Daniel, any view why TSMC have been able to get 10nm right, when Intel has had to delay its 10nm? And where is Samsung relative to the TSMC roadmap?

At 10nm Samsung is right with TSMC. At 7nm I'm not sure but I would guess Samsung again will be right with TSMC. I did hear at SPIE that Samsung will be using EUV at 7nm which could cause delay but I have not heard anything definite from the ecosystem.

 

[Brianhayes]

Surely the better title for these views would be, "Winner takes all".

The industry has now reached the point where only TSMC has the volumes to deliver logic chips to the fabless designers at the price they require and turn a profit. Intel's costs are off the scale and Samsung cannot justify TSMC's Capex levels out of what is left after TSMC has taken 70%+ market share. Not unless it is 2.3x more efficient! Globalfoundries doesn't figure in this, IMO.

Depreciation/cost of Capex, is over 50% of the average cost of fabrication of a SoC. This can be seen easily enough from TSMC's financials. This is the killer.

Only Apple has both the money and market share to dictate its own terms. No reason why it should help Samsung.

 

[Brianhayes]

View attachment 16815
There will be two versions of 7nm, one for mobile and one for HPC (servers).

It is often the shortest of sentences that convey the most information but this one begs the really big questions. Who will the partner(s) be if not Samsung and/or Apple? Will the IC in question be Ares. For how long have TSMC and ARM been planning the move? What will be the (large) projected volumes needed to justify a second version of the node?. Most important question, however, is, why have they decided to declare war on Intel two years before they can attack its markets in a serious way?

 

[tty2]

Did the title of this post have anything to do with its content? This is just TSMC being ahead, not anything to do with the IDM model.

 

[KevinK]

DAN,
Didn't you mean "10nm will start production in Q4 2016" (not 2015) ?

 

[Daniel Nenni]

DAN,
Didn't you mean "10nm will start production in Q4 2016" (not 2015) ?

Fixed. Thanks.

 

[Daniel Nenni]

Did the title of this post have anything to do with its content? This is just TSMC being ahead, not anything to do with the IDM model.

Sorry, that was sarcasm in response to Intel's claim in 2012 that the fabless business model was collapsing:

https://www.semiwiki.com/forum/content/1239-intel-says-fabless-model-collapsing-really.html

But given that the high cost of fabs has driven the semiconductor industry to become fabless or fab-lite I think one could argue that the IDM model is in fact collapsing (consolidating).

 

[lilo777]

Sorry, that was sarcasm in response to Intel's claim in 2012 that the fabless business model was collapsing:

https://www.semiwiki.com/forum/content/1239-intel-says-fabless-model-collapsing-really.html

But given that the high cost of fabs has driven the semiconductor industry to become fabless or fab-lite I think one could argue that the IDM model is in fact collapsing (consolidating).

I think both models are doing just fine with some twists: weaker players in both worlds are collapsing. TSMC still has quite a way to go to reach the profit levels of major IDMs (i.e. Intel and Samsung - although I do not know how much profits Samsung is actually making on their IDM/RAM/NAND business).

 

[Daniel Nenni]

It is often the shortest of sentences that convey the most information but this one begs the really big questions. Who will the partner(s) be if not Samsung and/or Apple? Will the IC in question be Ares. For how long have TSMC and ARM been planning the move? What will be the (large) projected volumes needed to justify a second version of the node?. Most important question, however, is, why have they decided to declare war on Intel two years before they can attack its markets in a serious way?

I understand completely why QCOM is using Samsung 14nm for their Snapdragon SoCs. Apple has TSMC's full attention and Samsung is a very large customer. QCOM is also developing server chips and that would be my guess as to who the big 7nm customer ARM and TSMC are working with. I do know that Dr. Morris Chang spent time in San Diego with QCOM and I highly doubt he was asking for SoC business. QCOM is also historically committed to fab 2nd and 3rd sourcing so this would also fit that model of not being dependent on Samsung foundry for all products.

I wish I could say that Apple was getting in the server business but I highly doubt that. I also seriously doubt that it is AMD. I have also heard rumors that QCOM is considering acquiring Xilinx to boost their server offering. That would be clever indeed!

Speaking of FPGAs, the CEO of Xilinx presented at the TSMC Symposium yesterday as well. He presented a graph showing that being first to a node as an FPGA vendor equals superior market share:

Xilinx
28nm 65% market share
20nm 80% market share
16nm 100% market share (thus far)

Xilinx beat Altera to 28nm by a matter of months, Xilinx beat Altera to 20nm by more than a year, and Intel/Altera has yet to release a 14nm or 16nm part. So go on and tell me how great the Intel 14nm process is. The Intel-Altera foundry deal was announced in 2013 with products expected to be delivered in 2014. Then it was 2015:

John P. Daane - ALTR CEO: So, schedule for our Intel 14 nanometer product remains Q1 2015 for our first production part tapeout which will sample roughly midyear to customers.

And here it is 2016 and we are still waiting......

 

[Daniel Nenni]

I think both models are doing just fine with some twists: weaker players in both worlds are collapsing. TSMC still has quite a way to go to reach the profit levels of major IDMs (i.e. Intel and Samsung - although I do not know how much profits Samsung is actually making on their IDM/RAM/NAND business).

Here's another way to look at it:

Would IDMs open up their fabs to external customers if the IDM model was doing fine, especially Intel? Historically speaking IDMs created the fabless market by offering open fab capacity to fabless companies. Then came TSMC and the rest is history as most IDMs are now themselves fabless or fab lite and the fabless companies continue to take market share from IDMs.

 

[lilo777]

Here's another way to look at it:

Would IDMs open up their fabs to external customers if the IDM model was doing fine, especially Intel? Historically speaking IDMs created the fabless market by offering open fab capacity to fabless companies. Then came TSMC and the rest is history as most IDMs are now themselves fabless or fab lite and the fabless companies continue to take market share from IDMs.

Good point. Although Intel trying to enter foundry business may or may not be directly linked to the issues with IDM business as such. There is a limit to what amount of business can be done profitably with IDM model and it looks like Intel already has all of it (niche markets would not make any difference for Intel). Another traditional motivation in such situations (apart from simply making more profits) could be the idea that any dollar made in foundry business will deprive their competitors (fabless companies) of funds for capital investments thus putting Intel in a better competitive position (obviously just a speculation on my part).

 

[Daniel Nenni]

Good point. Although Intel trying to enter foundry business may or may not be directly linked to the issues with IDM business as such. There is a limit to what amount of business can be done profitably with IDM model and it looks like Intel already has all of it (niche markets would not make any difference for Intel). Another traditional motivation in such situations (apart from simply making more profits) could be the idea that any dollar made in foundry business will deprive their competitors (fabless companies) of funds for capital investments thus putting Intel in a better competitive position (obviously just a speculation on my part).

That is certainly possible. An interesting note from an ecosystem friend at the TSMC Symposium:

"Intel Foundry has a 3-5x cost disadvantage compared to the other foundries"

 

[lefty]

Does anyone know how Globalfoundries is doing with their 10nm and 7nm?

 

[lilo777]

That is certainly possible. An interesting note from an ecosystem friend at the TSMC Symposium:

"Intel Foundry has a 3-5x cost disadvantage compared to the other foundries"

Is this about the cost of manufacturing for Intel or the cost of making a chip for customers? Either way it sounds strange.

 

[Daniel Nenni]

Does anyone know how Globalfoundries is doing with their 10nm and 7nm?

GF will not use the Samsung recipe for 10nm, they will use IBM's 10nm and 7nm. From what I have learned IBM's 10nm is closer to Intel's in metals and such. I have been pushing GF for info on their 10 and 7 plans but they are keeping quiet for now. I have a Samsung 10nm briefing next week so I will write more after that.

 

[hist78]

I think both models are doing just fine with some twists: weaker players in both worlds are collapsing. TSMC still has quite a way to go to reach the profit levels of major IDMs (i.e. Intel and Samsung - although I do not know how much profits Samsung is actually making on their IDM/RAM/NAND business).

I'm not sure what the "profit level" you are pointing to. If we look at the after tax net profit margin for the past ten years, TSMC has performed much better than Intel.

% Net Income of Revenue
[table] style="width: 434px"
|-
| style="width: 44px; height: 20px" |
| style="width: 37px" | 2015
| style="width: 39px" | 2014
| style="width: 39px" | 2013
| style="width: 39px" | 2012
| style="width: 41px" | 2011
| style="width: 39px" | 2010
| style="width: 39px" | 2009
| style="width: 39px" | 2008
| style="width: 39px" | 2007
| style="width: 39px" | 2006
|-
| style="height: 20px" | TSMC
| align="right" | 36.3
| align="right" | 34.6
| align="right" | 31.5
| align="right" | 32.8
| align="right" | 31.5
| align="right" | 38.7
| align="right" | 30.2
| align="right" | 30
| align="right" | 34.1
| align="right" | 39.6
|-
| style="height: 20px" | INTC
| align="right" | 20.6
| align="right" | 21
| align="right" | 18.3
| align="right" | 20.6
| align="right" | 24
<td align="right" class="wysiwyg_dashes_td">26.8
<td align="right" class="wysiwyg_dashes_td">12.4
<td align="right" class="wysiwyg_dashes_td">14.1
<td align="right" class="wysiwyg_dashes_td">18.2
<td align="right" class="wysiwyg_dashes_td">14.3
<colgroup><col><col><col span="3"><col><col span="5"></colgroup>|-
[/table]

Source: S&P Capital IQ Stock Report

In terms of Samsung, there are some difficulties to analyze their Semi's true profitability due to Samsung's organization structure. But Samsung semi division's operating profit margin (before tax) is 26.88% for 2015 and 22.10% for 2014. So Samsung's Semi after tax net profit margin should be less than TSMC.

Source: http://www.samsung.com/us/aboutsams...se/downloads/2012/20160128_conference_eng.pdf

Considering the huge CAPX needed to keep up with the technology and the huge volume needed to justify the new capacity, it's a bad situation for both Intel and Samsung Semi. Daniel's provocative title "The IDM Business Model is Collapsing!" may not seem too far from the situation IDMs are facing.

 

[Portland]

IDM struggles had to do more with investments in arizona and oregon not paying off than the model. Kiryat Gat doesn't have the same problem.