TSMC owning the silicon manufacturing game

[witeken]

Moore's Law Slowing? Don't Tell TSMChttp://community.cadence.com/cadenc.../archive/2016/03/25/tsmc-manufacturing-status

Well worth a read. I wonder how Intel compares for comparison...

TSMC is a manufacturing powerhouse. It has twice the capacity of any other non-memory semiconductor company. It has the best yield in the industry, driven by collecting over 1M datapoints per second from the equipment in its fabs. They are the only company that can ramp from a standing start to full volume by adding 20K wafers per month for three consecutive months. So from 0 to 60K wafers per month in a single quarter (basically manufacturing 90K wafers in the first quarter and then 180K wafers each quarter after that).

20nm was an unprecedentedly steep ramp. However, 16nm was even faster, just three months to full capacity. In 2016, the plans are to continue the N16 ramp to 300,000 wafers per quarter. The defect measure D0 is now a record low <0.06. Was 0.1 when production started. The ramp for N10 is planned to be slightly steeper still.

Here some forward looking statements. As always, take with big grain of salt because no one can predict yield learning.

* 16FFC started ramping this quarter and so should be at 60,000 wafer per month by the end of Q2 or earlier (since the ramps only seem to get faster)
* 10FF starts ramping in 2H 2016 with a planned capacity of 200K wafers/quarter. Risk production is just about to start.
* 7FF qual is planned for Q1 2017 (two processes, high performance and mobile). Risk production March 2017. Ramp in Q2 or Q3.

Two years later they will be at 7nm with a huge density advantage and either a big performance increase for the HP flavor of N7, or a huge power reduction for the mobile version of N7. If all goes to plan, that is something that has never been seen before.

Well, we'll see.

 

[Arachnotronic]

Moore's Law Slowing? Don't Tell TSMChttp://community.cadence.com/cadenc.../archive/2016/03/25/tsmc-manufacturing-status

Well worth a read. I wonder how Intel compares for comparison...




Here some forward looking statements. As always, take with big grain of salt because no one can predict yield learning.




Well, we'll see.

Kind of a puff piece IMO. Don't expect huge 10nm volumes until 2H 2017.

 

[Lonyo]

Their nodes are not what they claim relative to Intel. This is not a new thing.

http://www.extremetech.com/computin...-10nm-production-this-year-claims-5nm-by-2020

Further, EUV tech is primarily being driven by ASML, in which Intel, Samsung and TSMC have invested, with IBM also having use of the same tech. TSMC can't magically leap ahead of everyone when they are all basically waiting on ASML's EUV tech to get things working properly at smaller nodes.

TSMC are basically catching up with Intel while Intel are having issues, but the article seems to assume TSMC won't have the same issues. Further, once TSMC hits "7nm" (10nm) they will probably hit a problem if they don't have ASML equipment up and running.

 

[ShintaiDK]

Seems TSMC got just as big issues if not bigger than Intel at 10nm.

 

[imported_bman]

I really hope they hit their 7nm targets, I would hate to see the 16nm/14nm nodes last as long as 28nm.

 

[raghu78]

Kind of a puff piece IMO. Don't expect huge 10nm volumes until 2H 2017.

Sorry but you are underestimating the threat of TSMC to Intel's server business. TSMC enables Intel's competitors so the threat of TSMC 7nm High performance products competing against Intel 10nm process products in 2019 is real. TSMC 7nm will have the same transistor density as Intel 10nm and the fact that TSMC is having a 7nm high performance version optimized for enabling 4 Ghz clock trees is a good sign that the ARM ecosystem wants TSMC to enable them in their fight against Intel. TSMC's model of BEOL changes followed by FEOL changes at 20nm/16nm has proved very successful as TSMC has had the steepest ramp for any process node at 16FF+. 10nm/7nm will follow a similar model. The big advantage is Apple will guarantee them the 10nm volumes and the crucial yield learning while the rest like Xilinx, Nvidia, Qualcomm wait for the 7nm process for their FPGA, GPU, mobile and server product needs. 7nm will use the same BEOL as 10nm and will benefit tremendously from the 10nm yield learning as TSMC will only make changes to FEOL at 7nm.

 

[dark zero]

The big problem of Intel is the exclusiveness of the x86 uarch, making it a brutal disadvantage since they are failing hard on mobiles... They don't have any rocking device to sell and the Asus Zenfone is only a sunlight in the cloudy panorama of them.

 

[VirtualLarry]

The big problem of Intel is the exclusiveness of the x86 uarch, making it a brutal disadvantage since they are failing hard on mobiles... They don't have any rocking device to sell and the Asus Zenfone is only a sunlight in the cloudy panorama of them.

I think that your point is, it's not just Intel that's failing, it's the entire x86 / x64 ecosystem!

 

[cbn]

One thing I noticed about phone SoCs is that they don't appear to need aggressive power reduction on the node to make the smaller (higher density) xtors worthwhile. This I assume due to the integration of various low power IP in addition to the CPU and GPU cores.

This, in contrast, to homogenous high power parts like CPUs (eg, high core count Xeons) and GPUs where it appears the silicon used needs greater power reduction in order to justify the smaller xtors.

 

[ShintaiDK]

I really hope they hit their 7nm targets, I would hate to see the 16nm/14nm nodes last as long as 28nm.

That's going to happen anyway due to cost. GPUs for example in 2020 will be 14/16nm too unless a miracle happens. (EUV for example).

 

[sm625]

I love how people are praising TSMC while I'm still stuck with a gpu on a process that is 5 years old. Yep, TSMC is owning everybody. 28nm forever!

 

[ShintaiDK]

http://www.digitimes.com/news/a20160330PD207.html

Apple is going down fast due to the global economy. The big question is who will pay for these next nodes.

 

[Arachnotronic]

Sorry but you are underestimating the threat of TSMC to Intel's server business.

Didn't say a peep about Intel or TSMC in servers. Strange that you would just randomly bring it up.

TSMC enables Intel's competitors so the threat of TSMC 7nm High performance products competing against Intel 10nm process products in 2019 is real.

There are always competitive threats. Anyway, ask yourself why TSMC hasn't put out a PR claiming 10nm risk production start yet when it was supposed to happen late last year?

10nm is proving to be quite hard for everybody: TSMC, Intel, and Samsung.

TSMC 7nm will have the same transistor density as Intel 10nm and the fact that TSMC is having a 7nm high performance version optimized for enabling 4 Ghz clock trees is a good sign that the ARM ecosystem wants TSMC to enable them in their fight against Intel.

Cool.

TSMC's model of BEOL changes followed by FEOL changes at 20nm/16nm has proved very successful as TSMC has had the steepest ramp for any process node at 16FF+.
10nm/7nm will follow a similar model. The big advantage is Apple will guarantee them the 10nm volumes and the crucial yield learning while the rest like Xilinx, Nvidia, Qualcomm wait for the 7nm process for their FPGA, GPU, mobile and server product needs. 7nm will use the same BEOL as 10nm and will benefit tremendously from the 10nm yield learning as TSMC will only make changes to FEOL at 7nm.

Incorrect. At 7nm TSMC is indeed planning BEOL changes. At 10nm TSMC will be using triple patterning for critical metal layers and will transition to SAQP for critical layers at 7nm.

 

[ksec]

I love how people are praising TSMC while I'm still stuck with a gpu on a process that is 5 years old. Yep, TSMC is owning everybody. 28nm forever!

And not a fault of TSMC. The only foundry that offer 20nm was TSMC, not only it is most consumed by Apple, it is also only a half step to 16nm. GPU decided to wait rather then invest into 20nm. And because of the HUGE ramp in 16nm FFC, you should be enjoying a fast GPU soon.

 

[MrTeal]

That's going to happen anyway due to cost. GPUs for example in 2020 will be 14/16nm too unless a miracle happens. (EUV for example).

Why do you think GPUs will skip the 10nm node, or are you thinking 10nm won't be ready for larger dies until 2020?

 

[ShintaiDK]

Why do you think GPUs will skip the 10nm node, or are you thinking 10nm won't be ready for larger dies until 2020?

Economics.

 

[ShintaiDK]

And not a fault of TSMC. The only foundry that offer 20nm was TSMC, not only it is most consumed by Apple, it is also only a half step to 16nm. GPU decided to wait rather then invest into 20nm. And because of the HUGE ramp in 16nm FFC, you should be enjoying a fast GPU soon.

Samsung got a 20nm node too. It was used for the Exynos 5430 for example.

 

[MrTeal]

And not a fault of TSMC. The only foundry that offer 20nm was TSMC, not only it is most consumed by Apple, it is also only a half step to 16nm. GPU decided to wait rather then invest into 20nm. And because of the HUGE ramp in 16nm FFC, you should be enjoying a fast GPU soon.

20nm wasn't a half node, it was a full node jump from 28nm. With good scaling you would have expected a doubling of density between the two.

 

[Systems analyst]

First Physical IP Licence signed for ARM 7 nm Physical IP, in 2015 Q4.

http://ir.arm.com/phoenix.zhtml?c=197211&p=irol-EventDetails&EventId=5213082
refers.

Apple A11?

 

[astroboy888]

Their nodes are not what they claim relative to Intel. This is not a new thing.

http://www.extremetech.com/computin...-10nm-production-this-year-claims-5nm-by-2020

Further, EUV tech is primarily being driven by ASML, in which Intel, Samsung and TSMC have invested, with IBM also having use of the same tech. TSMC can't magically leap ahead of everyone when they are all basically waiting on ASML's EUV tech to get things working properly at smaller nodes.

TSMC are basically catching up with Intel while Intel are having issues, but the article seems to assume TSMC won't have the same issues. Further, once TSMC hits "7nm" (10nm) they will probably hit a problem if they don't have ASML equipment up and running.

TSMC has demonstrated that they can still do 7nm with triple patterning photo lithography techniques. So EUV is not required until 5nm node in 2019. BTW, the company started working on 10nm and 7nm technology simultaneously starting 2011-2012. Looks like 10nm will be short-life node (like 20nm node) - not many people (except Apple and few mobile chip makers) will use it. Because 7nm node is coming on line earlier than expected, many of their customers might skip 10nm and to directly to 7nm in 2017-2018 time frame.

Regarding EUV, TSMC being the biggest spender in semiconductor equipment basically has an entire department of ASML engineers working on site at TSMC. TSMC has been working on EUV since 2008. Currently, TSMC has demonstrated 500-1000 EUV wafer processing per day, but to sustain consistent production rate is the another story.

If interested, you can read all about it here. This an summary of presentation at SPIE conference of TSMC and Intel's progress.
https://www.semiwiki.com/forum/content/5518-tsmc-intel-long-road-euv.html

 

[astroboy888]

http://www.digitimes.com/news/a20160330PD207.html

Apple is going down fast due to the global economy. The big question is who will pay for these next nodes.

Well Apple has about 220 billion USD on hand and still makes 200+billion revenue per year. (On a down year). Apple's bad year is everyone else's great year. So they be fine. Besides, Apple only pays TSMC $3.5-3.7B per year - about 15-20% of TSMC's $20B per year revenue. So even if Apple orders are down, they got plenty of other customers (particularly orders from China) to pay for their process.

 

[ShintaiDK]

Well Apple has about 220 billion USD on hand and still makes 200+billion revenue per year. (On a down year). Apple's bad year is everyone else's great year. So they be fine. Besides, Apple only pays TSMC $3.5-3.7B per year - about 15-20% of TSMC's $20B per year revenue. So even if Apple orders are down, they got plenty of other customers (particularly orders from China) to pay for their process.

Like who? Who else is fielding that amount of leading edge parts besides Intel?

 

[destrekor]

I think that your point is, it's not just Intel that's failing, it's the entire x86 / x64 ecosystem!

Not really though. Intel has been demonstrating tremendous success taking the Broadwell-DE cores to market in SoC packages that outperform ARM designs. And they really just got started with that effort in the Xeon D series. That will enable Intel to continue to fend off ARM chips for low-power servers for the foreseeable future.

And AMD's Zen could be refreshing.

 

[Systems analyst]

Like who? Who else is fielding that amount of leading edge parts besides Intel?

I would think that Qualcomm are very interested, for new server SoCs and consumer SoCs, and Mediatek for consumer SoCs. The Chinese will emerge as large customers too.

 

[Systems analyst]

Not really though. Intel has been demonstrating tremendous success taking the Broadwell-DE cores to market in SoC packages that outperform ARM designs. And they really just got started with that effort in the Xeon D series. That will enable Intel to continue to fend off ARM chips for low-power servers for the foreseeable future.

My view is that in future the server/network market will demand a whole sea of SoCs, each fit-for-purpose in its own application domain; IntelD: can only cover so much; an example of a specialist application is Paypal's use of ARM based servers to monitor transactions in real-time.

" - Barcelona Supercomputing Center announcing that its "inexpensive" ARM-based High Performance Computers are now fully-operational and made available to scientific community across Europe
- PayPal releasing information demonstrating the benefits of using ARM based servers with a 50% reduction in capex, an 85% reduction in running costs and a 10-fold increase in server density compared to traditional data center equipment"