TSMC begins 20 nm mass production early 2014 - goes for 16 nm FinFET in 2015.

[Mondozei]

In the latest quarterly report from TSMC, we had this nugget:

20-nanometer SOC and 16-nanometer FinFET are both progressing well. On 20 nanometers, we see little competition. The risk production has started in the first quarter[...]equipment are streaming in and are being installed, are going to be tried out and the volume production will start in early 2014

Morris Chang, the chairman, is saying that they've been testing the technology for many months now in small batches and that the plans are going smoothly.

In other worlds, we're talking about mass production in early 2014. Pretty interesting considering Intel's continous delays. Intel says they'll introduce 14 nm next year but that could be limited production only, we don't know yet, and at any rate, Samsung has promised the same thing.

Remember all that hype about how Intel simply couldn't be caught up due the sheer size of their head-start?

 

[blackened23]

What's hilarious is that Intel's 22nm is still superior to TSMC's 20nm, TSMC is not using FinFET.

 

[gevorg]

"superior" performance is irrelevant for most of the phone/tablet market if the price is too high

 

[videogames101]

In the latest quarterly report from TSMC, we had this nugget:



Morris Chang, the chairman, is saying that they've been testing the technology for many months now in small batches and that the plans are going smoothly.

In other worlds, we're talking about mass production in early 2014. Pretty interesting considering Intel's continous delays. Intel says they'll introduce 14 nm next year but that could be limited production only, we don't know yet, and at any rate, Samsung has promised the same thing.

Remember all that hype about how Intel simply couldn't be caught up due the sheer size of their head-start?

Intel has delivered, Samsung and TSMC say they will.


My company Semiconductors 'R' Us is entering mass production of 7nm wafers in 2 months. Trust me, we'll deliver.


This thread exists one or two years too soon.


BUT HEY - all aboard the Intel hate machine!

 

[NostaSeronx]

TSMC and GlobalFoundries are both technically early for 20nm. For TSMC if this was 28nm, TSMC could be producing Tahiti XT for a November launch. GlobalFoundries is only slightly behind but will have a larger initial ramp in January/February.

 

[bullzz]

@Mondozei
"Pretty interesting considering Intel's continous delays. Intel says they'll introduce 14 nm next year but that could be limited production only, we don't know yet, and at any rate, Samsung has promised the same thing"
only difference here is that Intel has announced 2 products on 14nm that will be released in 2014, TSMC hasnt

 

[Haserath]

Could we get relevant numbers to compare processes with?

 

[blackened23]

edit: I have reading comprehension issues, nevermind

 

[NostaSeronx]

Could we get relevant numbers to compare processes with?

20-nm TSMC and GlobalFoundries:
-- 20nm LPM --
Vddnom = 0.9V with OD support
4Vts with multi Lg
1.8V IO (1.5V UD, stacked IO, LDMOS)
Comprehensive SRAM menu (UHD, HD, HC, Low Vmin)
90-nm CPP -> LgateMin: 22.5 nm
64-nm M1

I'm not spending money to get NMOS/PMOS Idsat/Ioff for 20nm for TSMC/GloFo. For GlobalFoundries, the Idsat/Ioff are supposedly 2x the metrics of 28nm but it is very vague which model is in the comparison.
http://www.techdesignforums.com/pra...tel-SoC-finFET-characteristics-table-lrg1.jpg

I'm not sure if TSMC and GlobalFoundries teamed up or they are expecting to one up each other on prices.

 

[ShintaiDK]

So as expected. First chips will hit retail around summer 2014. If we are lucky we might see 20nm GPUs before xmas 2014. And knowing its TSMC, a good deal of time needs to be added. They have to be more than lucky to get 16nm in 2015.

 

[NostaSeronx]

Xilinx UltraScale FPGAs on 20nm will be out for customers by Q4 2013.
Altera Arria 10 FPGAs on 20nm will be out for customers by Q1 2014. (TSMC is delaying them because Stratix 10 is on Intel's 14nm)

 

[BUnit1701]

So as expected. First chips will hit retail around summer 2014. If we are lucky we might see 20nm GPUs before xmas 2014. And knowing its TSMC, a good deal of time needs to be added. They have to be more than lucky to get 16nm in 2015.

I can tell you there are TSMC customers who are fully confident that 16nm FinFET is going to be on time, so confident in fact that they are skipping 20nm designs.

 

[sontin]

nVidia could be one of them who are skipping 20nm for Tegra. 16nm is expected one year later and TSMC said that the lifetime of 20nm would be shorter than it was for 28nm.

 

[ShintaiDK]

We know 40nm was delayed. We know 28nm was delayed. And we do know 20nm is delayed as well:

I wouldnt bet on TSMC shipping 16nm products in 2015.

 

[Idontcare]

I can tell you there are TSMC customers who are fully confident that 16nm FinFET is going to be on time, so confident in fact that they are skipping 20nm designs.

Given that 16nm will be using what will be, at the time, an essentially mature 20nm BEOL and ex-fab (packaging, testing, etc) process, it definitely lowers the risk of TSMC missing the delivery date.

16nm is a non-standard affair for a number of reasons, but TSMC is actually in pretty good shape for getting their finfets off the ground.

In using a 20nm BEOL they don't have to worry about dealing with aggressive gate pitch which can further muck up the variability.

Compared to what TSMC is having to do to get to 10nm, 16nm is going to be a walk in the park. Customers justifiably have more confidence in TSMC's 16nm timeline than they had in their 40nm timeline.

 

[sontin]

In this slide the 28HPM and 20G node are on par with the reality.

 

[Khato]

I can tell you there are TSMC customers who are fully confident that 16nm FinFET is going to be on time, so confident in fact that they are skipping 20nm designs.

Though is that a vote of confidence in TSMC's "16nm" process? Or simply a complete disinterest in what TSMC is promising for 20nm - an anemic improvement in transistor performance characteristics and scaling that only might get to the break-even point in cost per chip compared to their 28nm.

Even if TSMC delivers their "16nm" process on schedule in 2015 they're still going to be ~3 years behind Intel. TSMC is only promising a 1.1x gate density improvement for their "16nm" process compared to their 20nm - Source, page 19. Realistically TSMC's "16nm" process is likely going to be a bit more dense than Intel's 22nm while being comparable in other metrics. (Similar to how TSMC's 28nm HPM is a bit more dense than Intel's 32nm, but otherwise comparable.)

 

[lefty2]

The million dollar question is though, who will get 20nm first, TMSC, GlobalFoundaries, or Samsung?
I'd say TMSC got to be the favourites, since they got to 28nm first.

 

[NostaSeronx]

Samsung because of NANDs. For logic, GlobalFoundries is producing Cortex A57s on 20nm, I don't know for who.

 

[LegSWAT]

Though is that a vote of confidence in TSMC's "16nm" process? Or simply a complete disinterest in what TSMC is promising for 20nm - an anemic improvement in transistor performance characteristics and scaling that only might get to the break-even point in cost per chip compared to their 28nm.

Even if TSMC delivers their "16nm" process on schedule in 2015 they're still going to be ~3 years behind Intel. TSMC is only promising a 1.1x gate density improvement for their "16nm" process compared to their 20nm - Source, page 19. Realistically TSMC's "16nm" process is likely going to be a bit more dense than Intel's 22nm while being comparable in other metrics. (Similar to how TSMC's 28nm HPM is a bit more dense than Intel's 32nm, but otherwise comparable.)

Get used to the process names by GF and TSMC being stamped by marketing:
TSMC's 16nm is about their 20nm BEOL combined with FinFET, what GF calls "14XM" is about their respective 20nm BEOL using FinFET. There's two more things to note:
1. Despite profiting mostly on the density side and less on the transistor performance, their basic 20nm processes can be equipped with some nice little extras each.
2. It seems harder to increase transistor performance in the same proportion as density; in consequence, processors with many execution units designed for parallel workloads will profit more than those designed for single-threaded performance.

 

[BUnit1701]

Though is that a vote of confidence in TSMC's "16nm" process? Or simply a complete disinterest in what TSMC is promising for 20nm - an anemic improvement in transistor performance characteristics and scaling that only might get to the break-even point in cost per chip compared to their 28nm.

Even if TSMC delivers their "16nm" process on schedule in 2015 they're still going to be ~3 years behind Intel. TSMC is only promising a 1.1x gate density improvement for their "16nm" process compared to their 20nm - Source, page 19. Realistically TSMC's "16nm" process is likely going to be a bit more dense than Intel's 22nm while being comparable in other metrics. (Similar to how TSMC's 28nm HPM is a bit more dense than Intel's 32nm, but otherwise comparable.)

It would have to contain some level of confidence in terms of timeline, if you dont design a 20nm chip and 16nm ends up late to the party, you have NO product instead of an under-performing, underwhelming 20nm product.

 

[Khato]

It would have to contain some level of confidence in terms of timeline, if you dont design a 20nm chip and 16nm ends up late to the party, you have NO product instead of an under-performing, underwhelming 20nm product.

Nope - if you don't design a 20nm chip and 16nm ends up late to the party you just happily continue producing your product on TSMC's 28nm HPM process. It's all a question of wafer pricing. If 20nm wafers cost twice as much as 28nm then the only reason to use the 20nm process is if it offers superior performance. From the TSMC presentation I linked previously one can extrapolate that 20nm offers 14% reduced power at same speed or a 15% increase in speed at same power... which isn't terribly impressive compared to numbers around 45% for 40nm to 28nm HPM. (Note that TSMC doesn't specifically claim that, they simply state that HPM provides better speed than 28nm HP at 28nm LP leakage levels, and 28nm HP was a 45% increase in speed over 40nm G.)

So as said, it's all about the wafer pricing. And my impression from various articles is that it's going to take awhile for 20nm to reach the same cost per chip as 28nm. (Makes sense when you consider that they have to go to double patterning for many more of the wafer processing steps on top of the normal cost increases that accompany a new process node.) Now that's not to say that the process won't be used, just that it's likely going to have a much more limited market than normal.