TSMC 16 FINFET Plus in risk production

[raghu78]

http://www.tsmc.com/tsmcdotcom/PRListingNewsAction.do?action=detail&newsid=9001&language=E

"Hsinchu, Taiwan, R.O.C. – November 12, 2014 – TSMC (TWSE: 2330, NYSE: TSM) today announced its 16-nanometer FinFET Plus (16FF+) process is now in risk production. This enhanced version of TSMC’s 16FF process operates 40% faster than the company’s planar 20-nanometer system-on-chip (20SoC) process, or consumes 50% less power at the same speed. It offers customers a new level of performance and power optimization targeted at the next generation of high-end mobile, computing, networking, and consumer applications."

"The 16FF+ process is on track to pass full reliability qualification later in November, and nearly 60 customer designs are currently scheduled to tape out by the end of 2015. Due to rapid progress in yield and performance, TSMC anticipates 16FF+ volume ramp will begin around July in 2015."

"NVIDIA and TSMC have collaborated for more than 15 years to deliver complex GPU architectures on state-of-the-art process nodes,” said Jeff Fisher, Senior Vice President, GeForce Business Unit, NVIDIA. “Our partnership has delivered well over a billion GPUs that are deployed in everything from automobiles to supercomputers. Through working together on the next-generation 16nm FinFET process, we look forward to delivering industry-leading performance and power efficiency with future GPUs and SOCs.”

This process node is going to bring some amazing technology products from Nvidia, AMD, Qualcomm, Apple. 16FF+ is a massive leap from 28HPM which is the current node for mobile chips like Snapdragon 805 and GPUs like GTX 980. I cannot wait to see products built on 16FF+ which will start appearing in late 2015 and early 2016.

At same power (TSMC 16FF+ is roughly 40% faster than 20SOC)
TSMC 16FF+ - 1.60x
TSMC 16FF - 1.40x
TSMC 20SOC - 1.15x
TSMC 28HPM - 1.00x

At same perf

TSMC 16FF+ - 0.35x
TSMC 16FF+ - 0.45x
TSMC 20SOC - 0.70x
TSMC 28HPM - 1.00x

The power/perf relation between 16FF, 20SOC and 28HPM is derived from the below presentation (slide 19)

http://www.eda.org/edps/edp2013/Papers/4-4 FINAL for Tom Quan.pdf

 

[ShintaiDK]

Remember all the numbers are based on a transistor and not how a design will work. So it cant be used for much when compared to actual designs.

TSMC 20nm risk production started in Q1 2013 as compare.

Seems the first 16FF+ products may appear in Q2 2016. But I wouldnt expect more than smartphone/tablet SoCs until Q3 or Q4.

 

[raghu78]

Remember all the numbers are based on a transistor and not how a design will work. So it cant be used for much when compared to actual designs.

Still it gives an idea of what we can expect from new products built at TSMC 16FF+ . For comparison 28 HP was 45% faster than 40G at same leakage.

http://www.nu-vista.com:8080/download/brochures/2011_28 Nanometer Process Technology.pdf

TSMC 20nm risk production started in Q1 2013 as compare.

16FF and 16FF+ share the same BEOL (Back end of line) tools and benefit from the 20SOC yield learning and ramp.

http://www.tsmc.com/uploadfile/ir/quarterly/2014/3Vs3J/E/TSMC 3Q14 transcript.pdf

page 5

"Next, I'll talk about the 16-nanometer ramp and competitive status. In 16-nanometer, we have two versions, 16 FinFET and the 16 FinFET Plus. FinFET Plus has better performance and has been adopted by most of our customers. 16 FinFET we began the risk production in November last year and since then have passed all the reliability qual early this year. For the FinFET Plus, we also passed the first stage of the qualification on October 7 and since then entered the risk production. The full qualification, including the technology and product qual, is expected to be completed next month. So right now we have more than 1,000 engineers working on ramp up for the FinFET Plus. On the yield learning side, the progress is much better than our original plan. This is because the 16-nanometer uses similar process to 20 SOC, except for the transistor. And since 20 SOC has been in mass production with a good yield, our 16 FinFET can leverage the yield learning from 20 SOC and enjoy a good and smooth progress. So we are happy to say that 16-nanometer has achieved the best technology maturity at the same corresponding stage as compared to all TSMC's previous nodes"

 

[witeken]

Remember all the numbers are based on a transistor and not how a design will work. So it cant be used for much when compared to actual designs.

TSMC 20nm risk production started in Q1 2013 as compare.

Seems the first 16FF+ products may appear in Q2 2016. But I wouldnt expect more than smartphone/tablet SoCs until Q3 or Q4.

The big is news is that HVM will start in July, which is Q3. For those who didn't follow it, it was originally expected to go into HVM within 1 year after 20nm HVM, which started in January. Just a few weeks ago it was Q2, so now it's officially slipped even further.

 

[raghu78]

The big is news is that HVM will start in July, which is Q3. For those who didn't follow it, it was originally expected to go into HVM within 1 year after 20nm HVM, which started in January. Just a few weeks ago it was Q2, so now it's officially slipped even further.

20nm volume production might have started in January but the production volume was low in H1 2014 as TSMC worked yields up. This was followed by a steep ramp in 20nm wafer revenue in Q3 2014 (10% of total wafer revenue).
TSMC now expects a similar steep ramp in Q3 2015 with high single digit revenues realised in Q4 2015 at 16 FINFET. The lead times for 16 FINFET are 1.5 quarters from wafers in to finished chips out. So expect wafers to go in late Q2 / early Q3 and you can expect the first 16FF / 16FF+ products in Dec 2015 or Jan 2016.

 

[ShintaiDK]

TSMC always expect very optimistic predictions and people put too much into foundry roadmaps.

Just look at history, when was the last time they even delivered on track? Not to mention the usual issue about foundry production and actual products.

 

[tviceman]

I think this is another strong indication that Nvidia is going to skip 20nm for everything (with the possibly exception of Tegra Erista). I also believe the 20nm AMD GPU rumors are false too, but due to the heavy amount of rumors around AMD's next gen GPU's, there is still a undeniable chance those GPU's could be on 20nm.

 

[f1sherman]

this is good as done for NV

for AMD I don;t know... they have been saying "20nm in subsequent few quarters" since 2013.
(but never clearly enough so that we know what products they are referring to)

IMHO they should have been smart enough to endure 28nm, such as it is, even with losses if need be
20nm being simply too weak stopgap station, and having to transfer asap to 16FF+

 

[witeken]

20nm volume production might have started in January but the production volume was low in H1 2014 as TSMC worked yields up. This was followed by a steep ramp in 20nm wafer revenue in Q3 2014 (10% of total wafer revenue).
TSMC now expects a similar steep ramp in Q3 2015 with high single digit revenues realised in Q4 2015 at 16 FINFET. The lead times for 16 FINFET are 1.5 quarters from wafers in to finished chips out. So expect wafers to go in late Q2 / early Q3 and you can expect the first 16FF / 16FF+ products in Dec 2015 or Jan 2016.

Your definition of HVM is a lot different than how TSMC views HVM. In any case, if TSMC is consistent with its definition, then it will begin HVM in Q3 followed by your "steep ramp" in Q1'16.

http://www.kitguru.net/components/g...ction-of-chips-using-20nm-process-technology/

 

[III-V]

Qualcomm did a bunch of work with TSMC and managed to reduce the metal pitch of the routing layers from 90nm to down to 80nm, reduced the number of masks needed, and pulled in the ramp of 16FF. Very impressive. The best part is that 20 and 16nm will actually provide cost improvements over 28nm, which was previously expected to not be the case, and that only means great things for GPUs.

http://ieeexplore.ieee.org/xpls/icp.jsp?arnumber=6894414

http://www.techdesignforums.com/blog/2014/06/03/qualcomm-optimises-mask-cost-20nm/

Your definition of HVM is a lot different than how TSMC views HVM. In any case, if TSMC is consistent with its definition, then it will begin HVM in Q3 followed by your "steep ramp" in Q1'16.

http://www.kitguru.net/components/g...ction-of-chips-using-20nm-process-technology/

TSMC went into 20nm HVM in Q1 of this year, no? They managed to feed both Qualcomm's 20nm modem demands and put out Apple's A8... That's pretty high volume.

Intel's been in HVM as well, but they won't really be ramping things up until... Well, right about now, if they plan on having Cherry Trail and the rest of Broadwell out by BTS 2015.

 

[witeken]

Qualcomm did a bunch of work with TSMC and managed to reduce the metal pitch of the routing layers from 90nm to down to 80nm, reduced the number of masks needed,

So now they're on par with Samsung.

and pulled in the ramp of 16FF.

I've only seen delays. They were able to pull it in by more than a year simply because it is the same node.

Very impressive.

Intel's 14nm transistor looks a lot more impressive, in my opinion. And: it's available right now. http://www.chipworks.com/en/technic...ces/blog/intels-14-nm-parts-are-finally-here/

The best part is that 20 and 16nm will actually provide cost improvements over 28nm, which was previously expected to not be the case, and that only means great things for GPUs.

http://ieeexplore.ieee.org/xpls/icp.jsp?arnumber=6894414

http://www.techdesignforums.com/blog/2014/06/03/qualcomm-optimises-mask-cost-20nm/

Well, apparently wafer costs didn't rise that much, the process is up to sniff and yields are okay. To me, that sounds more like the bare minimum.

TSMC went into 20nm HVM in Q1 of this year, no? They managed to feed both Qualcomm's 20nm modem demands and put out Apple's A8... That's pretty high volume.

Sure, but it took them 8 months to do so and when they started HVM, 28nm was already 2 years available. July + 8 months is March. 8 months is to pessimistic, so take 2 months. For comparison, Ivy Bridge ramp up started in October, which is also about 6-7 months.

Intel's been in HVM as well, but they won't really be ramping things up until... Well, right about now, if they plan on having Cherry Trail and the rest of Broadwell out by BTS 2015.

I guess they only produced Core M due to low yields.

 

[Silverforce11]

this is good as done for NV

for AMD I don;t know... they have been saying "20nm in subsequent few quarters" since 2013.
(but never clearly enough so that we know what products they are referring to)

IMHO they should have been smart enough to endure 28nm, such as it is, even with losses if need be
20nm being simply too weak stopgap station, and having to transfer asap to 16FF+

As was expected by some of us a long time ago, NV has clearly chosen to ride the cheap and mature 28nm all the way to 16nm FF, because the 20nm process at TSMC is not suitable in terms of HP as it was designed for mobile SoCs, meaning it was too expensive for the minor gains.

I still believe AMD is not going to waste their time with TSMC 20nm and go straight to 16FF also. Why? Because Tonga is here, Tonga XT will drop soon and after that, Fiji XT on a large 28nm die. That is enough performance until 16FF.

 

[KaRLiToS]

I am so disapointed.

On the other side, my wallet won't. This is one of my oldest system in term of longevity.

Rampage IV extreme bought in 2011 , updated my 3930k for a 4930k for few to no improvement except the better IMC. And 4 x 290x bought in Oct 2013 that will last another year. Sigh. ( I hope I can really hold myself of upgrading)

 

[96Firebird]

Anyone gonna use GloFo's 14nm LPP?

 

[raghu78]

Anyone gonna use GloFo's 14nm LPP?

btw thats Samsung's 14 LPE/ 14LPP. GF has licensed it. The latest news is TSMC has overtaken Samsung in FINFET progress due to better yields. Definitely Samsung will have customers for their FINFET process. But the key here is can they get to good yields. If they do they could win even the Apple business

http://www.eetimes.com/document.asp?doc_id=1324211

http://www.fool.com/investing/gener...rs-on-b.aspx?source=isesitlnk0000001&mrr=1.00

 

[Madpacket]

Well I just ordered the Galaxy Note 4 with the 805 processor in it. Hopefully by the time I'm ready to replace it in a few years phones with the 16FF+ process are available. Nice to finally see some progress being made, feels like we've been on 28nm for too long.

 

[Mondozei]

We won't see 16 FF+ until 2016 at the earliest for desktop GPUs.
Mobile SoCs get to jump the queue, there's simply way too much money involved in that area compared to desktop GPUs for it to be any other way.

AMD is going 20 nm next year, that has been confirmed multiple times by AMD's executives.

Nvidia is still a mystery. I'm guessing they are waiting for R9-390 series to come out and see how it does. Big Maxwell can go on 28 nm and it could potentially be competitive with R9-390. If it isn't, they might go 20 nm. Either way, they are likely not releasing Big Maxwell for some time to come.

 

[III-V]

So now they're on par with Samsung.

Not sure. I haven't seen Samsung's 14nm specifications. They're probably pretty similar, though, given that they're both gate-last.

I've only seen delays. They were able to pull it in by more than a year simply because it is the same node.

Yes, there have been delays. Everyone is suffering from them, right now, unfortunately.

Things would have been far, far worse if they hadn't taken the time to do some further cost optimizations, though. I am unsure how much of that was responsible for the delay, but if 20 and 16 nm did not improve the economics over 28nm, the processes would be virtually worthless, outside of some niche designs with really fat margins. Thankfully, that's not the case, and we'll start seeing the market being flooded with 20nm devices from other companies that aren't Apple or Qualcomm.

Another thing to keep in mind is that FinFETs are really, really hard. It took over 12 years for them to become a reality, from conception in 1999 (perhaps even earlier) at Berkeley, to their debut in 2012 with Ivy Bridge. Everyone had been playing with them from at least ~2004/2005, and only Intel so far has managed to get viable yields for HVM.

I feel like you're being too dismissive of TSMC... the fact that fab tech has gotten as far as it has is incredible as is. It's tough, as an enthusiast, to not see things progressing as fast as they used to, but as a scholar, I've got a lot of respect for the people that come up with ways of constructing devices only a handful of atoms thick.

Intel's 14nm transistor looks a lot more impressive, in my opinion. And: it's available right now. http://www.chipworks.com/en/technic...ces/blog/intels-14-nm-parts-are-finally-here/

Well, yeah. Intel's the industry leader, by a wide margin. But what TSMC, et al., do is still very respectable. All of these companies have such a tremendous hand in making the world a better place. Delays suck, but everyone's doing the best they can do. The industry is too competitive for slacking off -- even if 16FF shows up a year later than it was initially slated for, you can bet your lucky stars that everyone at TSMC (and Samsung) was working to bring it to market as fast as possible. Especially with Intel now dangling its fab-carrots to fabless companies... it doesn't take much for a long-time, exclusive, high-volume customer to jump ship.

Well, apparently wafer costs didn't rise that much, the process is up to snuff and yields are okay. To me, that sounds more like the bare minimum.

I don't know what you're expecting, to be honest. Things are getting tougher and tougher to progress, and they're already difficult as it is. 20 nm is doing very well, the main issue is that picking up Apple meant a lot of other customers had to wait in line; whereas before they did not have to. 16 nm is coming along nicely as well, I'm sure.

Sure, but it took them 8 months to do so and when they started HVM, 28nm was already 2 years available. July + 8 months is March. 8 months is to pessimistic, so take 2 months. For comparison, Ivy Bridge ramp up started in October, which is also about 6-7 months.

Really, of all the things to criticize, TSMC's 20nm ramp should certainly not be one of them. They're probably pumping out around 50K wafers a month, at this point in time.

I guess they only produced Core M due to low yields.

I think the TSX bug had a lot to do with Broadwell's delay. Yields are definitely a part of it, though.

 

[witeken]

Another thing to keep in mind is that FinFETs are really, really hard. It took over 12 years for them to become a reality, from conception in 1999 (perhaps even earlier) at Berkeley, to their debut in 2012 with Ivy Bridge. Everyone had been playing with them from at least ~2004/2005, and only Intel so far has managed to get viable yields for HVM.

FinFETs were invented way before 1999, but that doesn't mean they get a lot of attention from the start. Intel did a lot of pathfinding from 2001 on FinFEts and around 2008, Intel had to choose which technology they'd implement at 22nm and their engineers were confident with Tri-Gates.

Source: https://www.youtube.com/watch?v=r28Z4mjLUDk