Design Con 2015
Breaking News
Blog

28nm: The Last Node of Moore's Law

NO RATINGS
9 saves
View Comments: Threaded | Newest First | Oldest First
Max The Magnificent
User Rank
Blogger
5, 10, 20 years?
Max The Magnificent   3/20/2014 9:33:14 AM
NO RATINGS
Hi Zvi -- I'm convinced that true monolythic 3D ICs will one day be the norm ... I just don't know when they will become mainstream ... but it's fun watching this all evolve...

bec0
User Rank
Rookie
Re: 5, 10, 20 years?
bec0   3/20/2014 4:39:09 PM
NO RATINGS
Hi Max - you are unsure of 5, 10, or 20 years. I agree, the crystal ball is mucho muddied right now. But we as an industry had better figure it out fast. With the economics flat or nearly flatted at 28nm, only a very few will see a need to go lower (for some other reason than cost scaling). I wonder about the slowing of innovation from this bottoming out as well. However, an aspect that may help us do this sooner rather than later: since 28nm will be around for a long long time, being the bottom of the cost bathtub, it is the perfect platform to try out monolithic 3d techniques (and any other ideas out there), especially for SoCs. The NAND folks are already getting there.....

techgc
User Rank
Rookie
Your opinion request...this may be the answer?
techgc   5/3/2014 2:14:49 PM
NO RATINGS
Monolythic GaAS combining optics and electonics on one die.  In development for 30 years and now moving from R+D to commercialization by DR. Talyor (U.Conn Tech Dean, PHD) at POET Technologies.   

Many X clockspeed performance of Si (I am not exaggerating) , up to 70% less power of current SI models.  Can be built with existing Semi fab equipement.   Little known company that is "coming out to the wolrd" exploding with 400% stock performance this year.   

This post may come off as a pump mail (full discolure, I am a shareholder), this said - I highly encourage you to check this out - if only for your interest.    

BAE verified technology. NASA funded (previously), design kits are built..company now turning into a fabless licensing play (see ARM), with first partner being announced within months. 

http://www.midasletter.com/midas-letter-financial-radio-podcast/poet-technologies-tsx-vptk-ceo-peter-copetti/

http://poet-technologies.com/

http://www.vvcnetwork.ca/empireclub/20140428-taylor/

http://poet-technologies.com/category/industry-articles/

Intersted in any feedback you have.  

TechGC    

 

buprestid
User Rank
Freelancer
28nm really lowest cost?
buprestid   3/20/2014 4:47:12 PM
NO RATINGS
I don't think 28nm has the cheapest cost per transistor. Is the author just pushing monolithic 3d?

1. Directly contradicts Intel Slides that show cost per transistor falling @22 and @16

2. You would see huge margin pressure at Intel shipping their 22nm Finfet devices if in fact it was so much more expensive. The die sizes and chip prices are similar to their 32nm products. 

The slides by Nvidia/Broadcom are best guesses and were made long before 20/16 were ready.

TSMC also has slides showing cost per transistor going down @20 and @16.

Or_Bach
User Rank
Rookie
Re: 28nm really lowest cost?
Or_Bach   3/20/2014 5:53:36 PM
NO RATINGS
It will be nice to see TSMC or Intel showing better cost of transistors. What I have seen out there is density chart which is actually do not result in reduce transistor cost due to the higher wafer cost. BUT this is well known by now, the real problem is the SoC costs !, it will be nice if you really read the blog before posting comments. Please look again at the SRAM density chart from the recent ISSCC 2014. If effectively the density increase from 28nm to 16nm is about 10% than the SoC cost will much higher at 16nm. Unless you can show data that is different you need to admit that 28 nm costs would be the cheapest for SoC for a while. And you are advise to read a previous blog published by EE Times " Why 450mm Will Be Pushed-Back Even Further" http://www.eetimes.com/author.asp?section_id=36&doc_id=1321239& , which provide additional support to the problem with embedded SRAM scaling. So regardless if you like monolithic 3D or not the 28nm is the last node of Moore's Law.

Zeev00
User Rank
Rookie
Re: 28nm really lowest cost?
Zeev00   3/20/2014 6:21:03 PM
NO RATINGS
The industry is in a deep dodo -- 450mm (infinitely?) delayed, EUV (infinitely?) delayed, and even pure logic transistors cost is not dropping anymore. One can believe Intel's charts if one chooses to but, clearly, their life is at stake over this question, so they must be optimistic or short their own stock. Everyone else's data shows level cost at best, somewhat up realistically.

 

But this blog brings the other whammy to the table -- memory shrink ... that barely happens anymore. Which means that complex chip cost below 28nm will go up. Significanly.

 

No. The sky will not fall. But many companies' stocks will.

krisi
User Rank
CEO
Re: 28nm really lowest cost?
krisi   3/21/2014 11:41:25 AM
NO RATINGS
If the cost below 28nm goes up why would anyone go there???

buprestid
User Rank
Freelancer
Re: 28nm really lowest cost?
buprestid   3/21/2014 12:50:25 PM
NO RATINGS
You can still use below 28nm even if the cost of the transistor is the same.


But you would see die sizes shrinking and total transistor counts staying the same to keep margins the same. Or they could increase prices (not likely!). Nvidia 20nm and Intel 16nm parts are coming soon. They will confirm if 28nm was the last cost effective node.





krisi
User Rank
CEO
Re: 28nm really lowest cost?
krisi   3/21/2014 1:02:25 PM
NO RATINGS
thank you @buprestid...makes sense...on the other note: I hope that Moore's law ends soon, I am just tired of listening to that phrase for the last 30 years ;-)...give me an new law pleeeease!

DMcCunney
User Rank
CEO
Re: 28nm really lowest cost?
DMcCunney   3/24/2014 12:38:05 PM
NO RATINGS
@krisi: If the cost below 28nm goes up why would anyone go there???

For the same reason such decisions are made in any industry.

Because those going there have a use case that they believe requires it, and for which they believe their customers will be willing to pay enough to let them do it profitably.

The question going forward will be what applications will require going below 28nm, and whether customers will be willing to pay what below 28nm parts will cost.

Moore's Law has historically meant semiconductor components would be progressively smaller, faster, and cheaper.  It's increasingly apparent that cheaper will no longer be part of the equation, and the new paradigm will be smaller, faster, and more expensive.  We are all in the process of finding out what that will mean for the semiconductor business.

krisi
User Rank
CEO
Re: 28nm really lowest cost?
krisi   3/24/2014 12:59:57 PM
NO RATINGS
thank you @DMcCunney...looking forward to see who is willing to pay for faster, smaller and more expensive...let's examine those in more details: faster is probably only marginally faster as chip delays are dominated by interconnect lines not intrinisic transistor speed...smaller doesn't really matter much unless the application is constrained in space like smartphone...and noone likes more expensive so I am not sure who will be willing to do faster, smaller and more expensive afterall...Kris

tpfj
User Rank
CEO
Re: 28nm really lowest cost?
tpfj   3/24/2014 1:14:12 PM
NO RATINGS
I agree with @krisi, faster has not been on the cards for almost a decade now. Sure low-power (LP) has got faster, but only by being singificantly less LP, in fact not LP at all these days.

Don't discard power in this trade-off of area, cost and speed. Mobile, by far the largest volume semi market, lives and dies by power consumption - no pun intended.

krisi
User Rank
CEO
Re: 28nm really lowest cost?
krisi   3/24/2014 1:38:46 PM
NO RATINGS
Good point @tpfj...we need to add power dissipation to this consideration...so "faster, smaller, less power hungry and more expensive"...maybe addition of lower power tiltes some applications towards using less than 28nm processing

DMcCunney
User Rank
CEO
Re: 28nm really lowest cost?
DMcCunney   3/24/2014 1:33:58 PM
NO RATINGS
@krisi: I am not sure who will be willing to do faster, smaller and more expensive afterall

I'm not either.

We were already bumping into speed limits as paople trid to push CPUs faster.  (There was an amusing series on a tech site as the editors successfully pushed a Pentium to 5ghz, using liquid nitrogen to cool it.  It was a "Kids, don't try this at home!" story.)  Aside from the technical challenges in pushing the chips faster, you had the parallel challenge of how to handle the heat faster CPUs generated.

Smaller still matters, and not just in handheld platforms like smartphones, with System On a Chip the promised land.

But the question is always "What will the customer pay for?"

The industry is already segmented, and there are "commodity" markets where price is the driver of the purchase decision and higher level niche markets where price is only one factor and more can be charged.  (And an industry challenge is that everything eventually becomes a commodity.)

I'm sure there will be applications for that sort of technologies, and markets willing to pay what it will cost to produce.  The question is whether the markets will be large enough to fund the R&D investments needed.   You get your R&D funds from the revenues you get by selling products to your customers, and the size of your market and the total revenues you can generate will place hard caps on what you can devote to R&D.

The old Chinese curse is "May you live in interesting times".  Well, we are.

buprestid
User Rank
Freelancer
Intel Cost Slides
buprestid   3/20/2014 7:20:00 PM
NO RATINGS
See slide 36. The cost per transistor is fallling. Also note that they are already in production of 22nm so the chart so its not some hypothetical guess.

http://mindspace.ru/wp-content/uploads/2013/12/INTC_investor_meeting_2013_IM_Smith.pdf


Everyone agress we are close to the end of Moore's law, but not there yet. EUV @10nm will make 10nm cost effective.

 

 

 

resistion
User Rank
CEO
Re: Intel Cost Slides
resistion   3/20/2014 9:50:38 PM
NO RATINGS
EUV is now being considered for 7 nm insertion with double patterning (10 nm insertion decision point past). I am not sure if that is an improvement to the cost curve. The source power target is not fixed, has to keep moving up.

resistion
User Rank
CEO
Re: Intel Cost Slides
resistion   3/21/2014 1:43:26 PM
NO RATINGS
Actually late last year, it seems Intel already decided to let quadruple patterning replace EUV at 10 nm (and presumably continue to 7 nm) node.

http://www.euvlitho.com/2013/S3.pdf (see p. 8)

buprestid
User Rank
Freelancer
Intel Processors
buprestid   3/20/2014 7:38:20 PM
NO RATINGS
Here is the link to die sizes from 32 to 22nm. Die sizes haven't decreased to possible compensate for extra die cost. Intel margins have also stayed the same.

http://www.anandtech.com/show/7003/the-haswell-review-intel-core-i74770k-i54560k-tested/5

Broadwell is right around the corner. Lets see what their die sizes are at 16nm. Perhaps SRAM amount will stay the same to compensate for poor SRAM scaling.

Intel is a good example to use since they don't need to take a margin hit in desktops and and servers. If 16nm really costs the same as 22nm they will have to reduce the die size by 50% to give the same margins.

 

 

alex_m1
User Rank
CEO
Re: Intel Processors
alex_m1   3/20/2014 8:56:43 PM
NO RATINGS
22nm vs 28nm is just margin case because no need for double patterning yet(according to intel). Why hasn't TSMC just migrated to 22nm without it ? Or is really hard to make 22nm work with single patterning ?

 

resistion
User Rank
CEO
Re: Intel Processors
resistion   3/20/2014 9:53:39 PM
NO RATINGS
It's a good question, it might be just to differentiate from Intel's 22 nm. But the aggressive DP cost might be too high to be tolerated by customers, maybe they'll go for 16 nm to get FinFET benefit (if any) as well.

Or_Bach
User Rank
Rookie
Re: Intel Processors
Or_Bach   3/20/2014 10:43:20 PM
NO RATINGS
Unlike the foundries we don't really know enough about Intel to analyze the transistor or their SoC costs. We do know that the for the foundries 28nm is the last node of Moore's Law. As to Intel it might help you reach your own conclosing if you read our blog: Intel vs. TSMC: an Update  <http://www.monolithic3d.com/2/post/2014/01/intel-vs-tsmc-an-update.html>

resistion
User Rank
CEO
Re: Intel Processors
resistion   3/20/2014 11:34:04 PM
NO RATINGS
Some evidence for Intel single patterning is provided by Intel:

http://www.intel.com/content/dam/www/public/us/en/documents/presentation/silicon-technology-leadership-presentation.pdf

But their fin pitch I read somewhere else is beyond standard single patterning.

http://www.chipworks.com/en/technical-competitive-analysis/resources/blog/plenty-of-room-at-the-bottom-intel-thinks-so/

So I am not sure if that counts as double patterning. If so then Intel 22 nm is already afflicted by double patterning.

alex_m1
User Rank
CEO
Re: Intel Processors
alex_m1   3/21/2014 9:30:41 AM
NO RATINGS
Resistion , looks like you're right.

"Intel uses self-aligned vias, sacrificial hard masks, and and DOUBLE PATTERNING on M1. The rest of the layers all are single patterned."

http://semiaccurate.com/2012/12/10/intel-talks-about-22nm-soc-transistors/

why on M1 ? how all this description makes for 8nm fins ?

Simon7382
User Rank
Manager
Re: Intel Processors
Simon7382   3/25/2014 4:31:35 AM
NO RATINGS
This whole article and discussion is prettty menaingless without Intel. There is no "wall" for Intel at 28nm whatsoever. The problem is with the other foundries who do not have the capital needed for follow Intel not to mention to ecatch up with their 2 node advanatge in the foreseeable future. It seesm to me that for now Moor's law is alive and well at Intel, but not so much at TSMC, GloFo, STM, tec.

resistion
User Rank
CEO
Re: Intel Processors
resistion   3/25/2014 7:42:51 AM
NO RATINGS
Intel does have volume to help pay off, but still there are some mixed signals regarding 14 nm. It's the first node to have such signals, particularly from its former 14 nm designated fabs. Maybe not a wall, but more like an obstacle in the course.

alex_m1
User Rank
CEO
Re: Intel Processors
alex_m1   3/25/2014 8:30:05 AM
NO RATINGS
Currently ,cost at 22nm at intel is more expensive than 28nm(at least according to fpga compnies who sell such silicon).

Also intel will lose a lot of stock market value if it appears that moore's law is dead, since better manufacturing is their only or at least most of it's core value today in a world that is shifting to ARM.

beinglass
User Rank
Rookie
Re: Intel Processors
beinglass   3/21/2014 2:28:38 PM
NO RATINGS
The reason is that 28nm is a planar transistor, 22 (20) can't be panar it has to be FinFET for many reason. TSMC decided t ojump to 16(14) with FinFET. There will b esome 20nm -planar but doesn't have an advantage.

Or_Bach
User Rank
Rookie
Re: Intel Processors
Or_Bach   3/23/2014 4:32:20 PM
NO RATINGS
The data for this blog is based on public information of the leading vendors (Intel, Samsung, TSMC, GF)  as was presented in previous blogs:

Why 450mm Will be Pushed-Back Even Further

Intel vs. TSMC: An Update

Why SOI is the Future Technology of Semiconductors

 


 

The two key points:

A. Below 28 nm the 2x transistor scaling cost benefits are neutralized by the escalating lithography costs.

B. Below 28 nm the embedded SRAM bit cell scale very poorly.  Just to keep with 2X density improvement per node, at 14 nm bitcell area should be less than 0.04 µm². The published data from TSMC and Samsung shows ~0.07 µm². Accordingly SoC build at 14nm would cost much more than the one at 28 nm.

 

Hence - 28 nm - The Last Node of Moore's Law

Brutus_II
User Rank
Manager
Monolythic 3D ICs
Brutus_II   3/21/2014 12:47:43 PM
NO RATINGS
@Max.  I agree with you about monolythic 3D IC's becoming mainstream.  It certainly seems like the logical choice and the ~28nm "wall, if you will, will drive this technology relatively quickly.  My crystal ball indicates that Fab 42, currently on hold, will be completely retooled to eventually accommodate carbon nanotube technology (likely a hybrid).   It won't happen in this decade and maybe not until well into the next.  Meanwhile the industry is in for seismic changes.  I'm glaad to be nearing the twilight of my career in this industry. I see lots of pain ahead.

DougInRB
User Rank
Manager
Re: Monolythic 3D ICs
DougInRB   3/24/2014 2:02:43 PM
NO RATINGS
Why all the doom and gloom?

Nothing except semiconductor based technology has gotten cheaper and better for so long a period of time.  We have just gotten used to it.

If semiconductors start getting more expensive per transistor, innovation will have to take a new direction.  Why does loading an OS take multiple GB of memory?  Because memory is cheap and bloated software is the path of least resistance.

Never underestimate the ability of really smart people to figure out a way to make more from less - as long as there is money to be made in doing so...

Or_Bach
User Rank
Rookie
Why all the doom and gloom?
Or_Bach   3/24/2014 5:50:56 PM
NO RATINGS
Moore's Law stopping at 28 nm will cause some shock waves in our industry but it is far from 'doom and gloom'. The collapse of number of vendors from 50 to less than 5 due to the escalating costs associated with dimension scaling was not a happy trend. Now we hopefully see the playing field get broaden again with growing number of vendors and growing diversified technology innovations.  

resistion
User Rank
CEO
China (SMIC) reaches 28 nm
resistion   3/25/2014 9:21:27 AM
Just read that a couple of months ago SMIC already had its first 28 nm MPW.

http://www.prnewswire.com/news-releases/smic-unveils-28nm-readiness-and-mpw-milestone-242026971.html

SHANGHAI, Jan. 26, 2014 /PRNewswire/ -- Semiconductor Manufacturing International Corporation ("SMIC",NYSE: SMI; SEHK: 981) , China's largest and most advanced semiconductor foundry, announced today that its 28nm technology has been process frozen and the company has successfully entered Multi Project Wafer (MPW) stage to support customer's requirements on both 28nm PolySiON (PS) and 28nm high-k dielectrics metal gate (HKMG) processes. Over 100 IPs from multiple third party IP partners as well as SMIC's internal IP team are prepared to serve various projects from worldwide design houses that have been showing interest in SMIC 28nm processes.

28nm process technologies primarily target mobile computing and consumer electronics related applications, such as Smartphone, Tablets, TV, Set-top Boxes and networking. It provides customers high performance application processors, cellular baseband, wireless connectivity etc. According to IHS' forecasts, the pure-play foundry revenue potential for 28nm will continue to rise with a CAGR of 19.4% from 2012 to 2017.

"I am pleased to announce the successful 28nm process milestone, which enables SMIC to better position itself in engaging and serving mobile computing related customers," said Dr. Tzu-Yin Chiu, Chief Executive Officer & Executive Director of SMIC. "As the first foundry in mainland China to offer 28nm process technologies, this significant milestone demonstrates SMIC's continuous growing capabilities in offering leading foundry technologies to worldwide IC designers."

"The first SMIC 28nm MPW shuttle included both 28PS and 28HKMG related customer products for verification, which was already launched at the end of 2013 as planned," said Dr. Shiuh-Wuu Lee, Executive Vice President of Technology Development of SMIC. "By taking more MPW shuttles in 2014, we will continue to take more positive steps to strengthen and diversify our technology offerings and meet customers' growing demands on both advanced and differentiated technologies."

AKH0
User Rank
Manager
Re: Intel Cost Slides
AKH0   3/25/2014 9:50:26 AM
NO RATINGS
Sang Kim, I think you are quoting the cost of bare wafer which is far less than what it costs to fabricate. And for your reference 14nm FDSOI uses a Si thickness of 6nm. That's all you need for a gate length of about 20nm that fits the pitch and by the way is far less than what FinFETs are using. Even at 10nm you do not need to scale the gate length. All needed is self-aligned contact, a card Intel has already played at 22nm. At the end Moore's low has nothing to do - at least to the first order - with the gate length. Regardless of transistor choice, whether bulk planar, FinFET, or FDSOI, each transistors need three contacts and routing metals. Pattering of those contacts and metal lines with enough design flexibility is what forced the industry to use multi- pattering (4 masks just for contact vias at 20nm). Soon we'll need up to 4 masks to print active regions, at least 3 to define gate, plus one more for each gate length, 5-6 for contacts, 2-4 for each via, and 3 for dense metals. FinFETs complexity just adds to the cost (both development and design). And its cost adder is by no means close to the 2% number that Mr Bohr quoted.

michigan0
User Rank
CEO
Re: Intel Cost Slides
michigan0   3/25/2014 11:41:39 PM

Sang Kim

 

Moore's Law predicted in 1965 was fulfilled because 

Moore's prediction in my opinion was based on the bulk 

silicon technology that ends at 28nm and not beyond, 

definitely not 7nm.

 

ST claims the FDSOI technology will continue Moore's law. 

However, look at some of the facts on FD-SOI. The 28nm

bulk silicon is in mass production over several years by 

major semiconductor companies such as Intel, TSMC, 

Samsung and others. 12 inch bulk silicon wafers currently 

are used in mass production today in order to maximize 

28nm transistor yields and minimize transistor process costs

at the same time. Therefore, even if 28nm FD-SOI were 

manufactured today, it would not be competitive with 28nm 

bulk silicon because 12 inch SOI wafer costs more than 

2 x 12 inch bulk silicon wafer. 

 

Here are some other facts. The bulk silicon technology ends 

at 28nm node because of the excessive transistor leakage 

current or short channel effects. That is why new FinTETs

are introduced by Intel at 20/22nm nodes. Intel is the only 

one today in mass production of 22nm bulk FinFETs over 

two years. TSMC 16nm bulk FinFETs will be volume 

manufactured possibly in 2015. The beauty of FinFETs over 

FD-SOI is its scalability. According to transistor device 

physics the transistor channel thickness of 5nm is 

considered to be the practical end of the roadmap. For 14nm 

FD-SOI the SOI thickness of 3.5nm is required to suppress 

transistor leakage current. It means that the 3.5nm is not 

manufacturable. For 14nm FinFETs, on the other hand, the 

Fin width that is equivalent to SOI thickness will require 

14nm. That is why FinFETs can be extended to the end of 

the roadmap, but FD-SOI can not be extended even to 

14nm FD-SOI.

 

IBM invented FD-SOI and ET(extremely thin) SOI, and

created International SOI Consortium in order to develop 

and manufacture them, but couldn't manufacture them

at any technology node yet. IBM exited FD-SOI and 

ET-SOI long ago! That is why I seriously doubt that the 

heavily SOI based S3S can compete with Intel's 22nm 

FinFETs. At what technology node can the S3S technology 

demonstrate its manufactrability, 28nm? 

 



AKH0
User Rank
Manager
Re: Intel Cost Slides
AKH0   3/26/2014 1:03:46 AM
NO RATINGS
Dear Sang Kim, I am afraid you are mixing too many items here.

1) Moore's law was based on the observation made on bipolar IC's. By your argument, it was fulfilled the day CMOS was born.

2) Cost of the bare Si wafer is just a fraction of the total fabrication cost. A cost difference of about $400 between Si and SOI wafers is well absorbed by reduction in the number of masks. I'll asure you, STM is one of the most cost-sensitive companies and would not do FDSOI if it would cost more. On the other hand, Intel is known for large margines. So, when Mr. Bohr claims bulk FinFET adds only 2% to process cost, I put a big question mark there.

3) Bulk Si technology did not stop at 28nm. There is bulk planar technologies by both Samsung and TSMC at 20nm. IMO, fabless companies where better off going to 20nm than getting distracted by the lure of FinFET, just to waste more than a year of design and finally figuring out things are not as rosy as the powerpoints show.

4) Contrary to all claims, I do not see big advantage at product level by FinFET products that are out there. The original claim was FinFET allows reduction in operating voltage of 200mV at the same performance, a 40% drop in active power just from FinFET (let alone natural ~30% reduction in power from node to node). This was never seen in products.

5) There is absolutely nothing in semiconductor physics that says 5nm is the minimum channel thickness.

6) Technology node is note represented by gate length. A 14nm node does not mean 14nm gate length. For all I see from publications, Intel and TSMC FinFET are using a gate length of 30nm or longer (up to 50nm for very low leakage devices), while FDSOI is at 25nm or smaller.

7) IBM is a member of the SOI Industry Consortium: http://www.soiconsortium.org/about-us/list-of-members.php

Promoting FDSOI is only one aspect of the consortium. SOI market spans far more, including RF, MEMS, etc.

8) IBM did not exit FDSOI. What will be 14nm FDSOI offering by ST was a joint development that started at IBM facility and was transfered to ST's site. This is very similar to other joint developments IBM had done with ISDA partners in the past (including the 28nm bulk planar that powers iphones). The fact that IBM decided to use SOI FinFET for its own server products is also very similar to IBM's use of PDSOI at 45nm, 32nm, and 22nm, despite the fact that IBM developed both bulk and PDSOI technolgies at each node.

9) S3S is an IEEE conference. IEEE decided to merge SOI conference, sub-Vt confernce, and 3D, because the committee beleived there is a strong harmony between these topics, despite covering seemingly different fields.

http://s3sconference.org/

 

 

michigan0
User Rank
CEO
Re: Intel Cost Slides
michigan0   3/31/2014 1:23:18 AM
NO RATINGS

Sang Kim

 

Thanks for your comments.

 

Two key issues with FD-SOI are its manufacturability and  

scalerbility as I pointed out. 

 

1) I said Moore's law ends at 28nm bulk Si. Moore in 1965

 couldn't predict a new FinFETs era beyond the 28nm bulk Si,

 and most of all, the new FinFETs can be extended to the end

 of the roadmap. 

 

2) Regardless the total fabrication costs, ST is not able to

 manufacture 28nm FDSOI even today. Even if 28nm FDSOI

 were manufactured today, it would not be competitive with 

 the bulk Si 28nm as I described.

 

3) Bulk Si technology stoped at 28nm! Samsung and TSMC

 indeed fabricated the planar bulk at 20nm, but the 20nm

 planar bulk transistor leakage current was very high due 

 to the short channel effects. The 20nm planar bulk was

 abandoned. TSMC adopted to manufacture 16nm FinFETs. 

 

4) The first FinFETs are introduced by Intel at 22nm node,

 and is in mass production over two years. Since at transistor

 level FinFETs show the lowest leakage current and highest

 performance. I expect the same at the product level. There

 are no rivals. 

 

5) Absolutely there is a physical limit. I said that 5nm node

 is considered to be practical end of the roadmap. The next

 3nm FinFETs will require a maximum 3nm channel thickness

 in order to suppress the transistor leakage current. At the

 3nm channel thickness, however,a new quantum mechanical

 phenomenon called quantum confinement occurs, resulting in

 an increase in Vt and large statistical variabilities in transistor

 transfer characteristics. This is because in the quantum

 confinement region the channel electrons don't behave like

 electron particles any more, instead, behaving like electron

 waves. As a result, the electron particle based classical

 Maxwell- Boltzmann statistics is no longer applicable, instead

 subject to Heisenberg uncertainty.

 

6) TSMC's 16nm node does really mean 16nm channel length

 or gate length, Lg. Intel and TSMC FinFETs do not require a

 gate length of 30nm or 50nm to obtain very low transistor

 leakage current. In order to suppress transistor leakage current

 or short channel effects for 16nm FinFETs the Fin-width at the

 bottom of transistor has to be only equal to or smaller than the

 channel length or gate length. That is, 16nm or smaller

 Fin-width is only required to suppress the transistor leakage

 current. The Fin-width of FinFETs is equivalent to the channel

 thickness of FDSOI. 16nm FDSOI for comparison 4nm

 channel thickness is required to suppress transistor leakage

 current. What a large difference! 16nm channel thickness

 for 16nm FinFETs vs 4nm channel thickness for 16nm FDSOI.

 

7) Promoting FDSOI is the major aspect of SOI Consortium.

 

8) IBM FDSOI skipped 28nm, 22nm, and will skip also 16nm

 and 14nm nodes. At what technology node IBM FD-SOI 

 will be manufactured? At 10nm node? IBM exited FDSOI long

 ago.I doubt that 14nm FDSOI will be manufacturable by ST

 because 14nm FDSOI requires SOI thickness of 3.5nm that

 is very close to the quantum confinement limit(3nm) as I

 described above. 

 

9) I don't understand why IEEE decided to merge SOI, Sub-Vt and 3D, and create a new S3S IEEE conference. I don't see

 such a strong harmony between these topics. Why these topics

 can't be presented in the well known and reputable, already

 existing IEEE conferences such as IEDM, VLSI and others?

 It seems too early to talk about creating IEEE conference for

 S3S. Who will be attracted by SOI, sub-Vt and and 3D at this

 14nm FinFETs era?    



AKH0
User Rank
Manager
Re: Intel Cost Slides
AKH0   4/3/2014 5:11:29 PM
NO RATINGS
Dear Sang Kim,


I'd apprciate if you would read my previous comments.

1) Moore's law is a bout cost, not about gate length, nor performance, nor leakage. Those are historically governed by Dennard's scaling and we can argue whether it still holds or not. But the discussion here is about the cost. As I pointed out earlier, Moore's law was made in the bipolar days. It was extended through NMOS and later CMOS days because they integrated chips continued to made at a lower cost per component.

2) ST demonstrated U8540 chips two years ago and it was shown to drop total power by 35% copared to 28nm bulk. Ironically ST-Ericson was disintegrated shortly after. So there is no product in the market. ST had continued making test chips for a handful of customers and all those that I am aware of had seen the benefit. However, ST does not have the fab capicity and said they are announcing the fab partner soon. So there is not such thing as ST "is not able to manufacture". There is a bog difference between "not being able to" and "not doing".

3) I have not seen any data suggesting 20nm bulk had leakage issue. Several companies in fact made the announcement that they will design in 20nm as their next node after 28nm. Others complained that there is not enough performance gain and would like to go directly to FinFET.

4) When talking about the FinFET performance what is the reference points? As for the leakage, their SoC paper used devices at 108nm pitch to show ULP leakage -- as opposed to 90nm pitch for other devices, which means these are roughly at Lg = 30+18nm = 48nm! I am confident you can get the same level of the leakage in a bulk planar device with the same gate length. As for the performance, both Intel and TSMC published data normalized to the footprint and not the actual device width. This is actually how TSMC claims performance better than Intel. While we can argue if this is thr right choice or not, first of all none of them were transparent in reporting numbers. Second, the higher current comes at the expense of higher capacitance, so it is not clear if there is net AC performance. Intel showed a RO delay of 13ps at VLSI'12 at 0.7V. That device would be about 10ps at 1V. Their 45nm node at the same leakage was 5ps! TSMC even did not show this. They showed a TCAD simulation of aggragate of what they claim to be representative of circuit performance.

5) It's been a while that node number has nothing to do with gate length. All discussions predicting requirements of Lg vs thickness are irrelevant here. BTW, quantum confinemnet is nothing to be scared of. It's already in play even in buld devices (you have a triangular quantum well vs the square shaped in FinFET). The inversion layer thickness is only 2-3nm in any advanced device. That's all you need to make a transitor. Yes, variability is a problem but it kicks in  at much thicker films, and unfortunately FinFET does not have any magic to offer.


6) Please read TSMC's paper again. Minimum gate length is 30nm -- and so is in Intel's although they decided to call it 26nm at a later point. Their low leakage device is also at 50nm gate length. As I mentioned above, technology node has nothing to do with gate length.

7) As I mentioned in my previous comment, IBM developed FDSOI along with its partners and transferred it to partner sites. There is no skipping here. This is very much same as the 28nm LP devices that powers iphones today. It was develped as a part of common platform and transferred to parners sites. BTW, ST has a paper in the upcoming VLSI. Channel length is not 14nm and channel thcikness is 6nm. It has the same gate pitch and metal pitch of what the FinFET camp calls 14nm.

8) I invted you to attend the conference in October in the bay area and you can see if there is harmoney or not. I beleive the organizing commettees of the three conferenced agreeded there is enough overlap to merge those together. But you are welcome to attend and comment if you think otherwise.

Thank you!

 

David Ashton
User Rank
Blogger
Re: Intel Cost Slides
David Ashton   4/3/2014 5:59:47 PM
NO RATINGS
@AHK0.....you said

"1) Moore's law is about cost, not about gate length, nor performance, nor leakage"

From http://en.wikipedia.org/wiki/Moore's_law

"Moore's law is the observation that, over the history of computing hardware, the number of transistors on integrated circuits doubles approximately every two years. The law is named after Intel co-founder Gordon E. Moore, who described the trend in his 1965 paper."

I think performance would be a better yardstick, but we have to stick to number of transistors.  3D ICs will keep it going a bit longer, maybe.

AKH0
User Rank
Manager
Re: Intel Cost Slides
AKH0   4/3/2014 8:26:31 PM
NO RATINGS
Moore's original paper says "the number of components per integrated circuit with minimum cost". The Wikipedia article seem to forget the minimum cost.

David Ashton
User Rank
Blogger
Re: Intel Cost Slides
David Ashton   4/3/2014 8:32:47 PM
NO RATINGS
@AHK0....  point taken....but using two variables in your yardstick does complicate matters a bit?

Susan Rambo
User Rank
Blogger
Keynoter approved
Susan Rambo   4/2/2014 7:22:45 PM
NO RATINGS
Bunnie Huang, in his keynote today, quoted from this article and said it was a really interesting read. That's why it's back on the EE Times home page. Great keynote today. Now that Moore's Law is slowing down, it's better for smaller companies that want to spend more time on hardware and software design and optimization rather than having to plan for the next few months when Moore's Law was going to make their product obsolete. He said Moore's Law favored bigger companies that could have several teams working on next generation simultaneously.

resistion
User Rank
CEO
ASML selling more multipatterning tools than EUV
resistion   6/26/2014 9:57:48 AM
NO RATINGS
The last earnings meeting they were saying 24 NXT-1970 multipatterning tools in Q1 alone, while hoping to squeak out a few more NXE:3300 EUV this year. Two things boggle me: 1) why continue with EUV if multipatterning is already a good business for ASML, and customers are paying (including the same as paying for EUV)? and 2) why continue the traditional Moore's Law shrink if the traditional cost reduction model is not viable anymore (due to new lithographies, high-k, fins, future channels, etc.)?

Flash Poll
Like Us on Facebook

Datasheets.com Parts Search

185 million searchable parts
(please enter a part number or hit search to begin)
EE Times on Twitter
EE Times Twitter Feed
Top Comments of the Week