Is Intels fabrication process advantage shrinking?

[blackened23]

Graphene transistors will be in use by 2020, Moore's law will not die. Graphene is not financially feasible right now, but it will be a thing by the end of this decade.

 

[Abwx]

FD-SOI has little to do with GloFo. It was developed entirely outside of them, and has only recently (last 12 months) been licensed to them such that GloFo could offer it to their customers.

Claiming that GloFo is trumping TSMC because STM developed FDSOI and cross-licensed it to GloFo is not the same thing, that just means STM developed a better process flow...and so did Intel, nothing new there.

GloFo itself is still stuck with the same internal R&D problems, tied to IBM and guaranteed to lag behind TSMC by a year, if not two, at 20nm and beyond.

If GloFo's future depends on STM continuing to develop FDSOI nodes for them to cross-license then that is a dim future, nobody makes money being the middle man between two other middle men.

There s a pdf doc from GF published at Taiwan s FDSOI tech forum
from 24 april 2013 with process explanations and caracteristics
including the Ids/Vgs curves , that is , transconductance , as well
as gates delays/VDD.

They say that the process is adequate for low power low leakage ,
wich exclude AMD s products apart from kabini , and that frequency
is better by 33% at 1.3V and by as much as 400% at 0.5V in respect
of bulk 28nm.

Test on real silicon of a dualcore A9 show 30% lower TDP at 1.8Ghz,
max frequency being extended from 1.8 to 2.5Ghz , frequency being
45% higher at 0.6V and max being currently at 3GHz as with ST-Ericsson
dual A9 wich was demonstrated at Computex.

 

[raghu78]

Graphene transistors will be in use by 2020, Moore's law will not die. Graphene is not financially feasible right now, but it will be a thing by the end of this decade.

DARPA looks at future technologies more than any other company or organization. So if the director of DARPA is saying there is meaning to it. Colwell also says economics and not just physics will bring the death of Moore's law.

 

[blackened23]

http://www.engadget.com/2011/04/08/ibm-shows-off-155ghz-graphene-transistor-in-the-name-of-darpa-re/

Graphene has already happened but is not economically feasible - yet. That will change by 2020 according to projections. As much as one firm might enjoy a halt in progress (AMD), that sure would benefit them eh? Progress wont' stop, unfortunately. Intel (among MANY universities, IBM, and others) have already heavily invested their R+D into graphene transistors and their projections, oddly enough, show commercialization possible around 2018-2020. Everyone in the silicon industry is heavily invested in graphene, the only question is when. Most signs point toward the turn of this decade.

Until then, Moore's law may slow to 3 years for most silicon producers - But it is still alive and well at intel until the turn of the decade when transition will likely happen.

That being said, economic viability will definitely be tough for 10nm and lower at TSMC and GloFo. Most firms do not have the resources that intel has and progress will slow for many companies. But - again, Moore's law is still alive and well at Intel.

 

[Fjodor2001]

So what will graphene bring? Will it enable going to 7 nm (and how far below that going forward)? Any other advantages like lower power consumption, stackable dies, or similar?

7 nm is estimated to become available in 2020 according to this Wikipedia article by the way:

http://en.wikipedia.org/wiki/Semiconductor_device_fabrication

 

[ShintaiDK]

So what will graphene bring? Will it enable going to 7 nm (and how far below that going forward)? Any other advantages like lower power consumption, stackable dies, or similar?

7 nm is estimated to become available in 2020 according to this Wikipedia article by the way:

http://en.wikipedia.org/wiki/Semiconductor_device_fabrication

Intel will have 5nm production start in 2019.

Graphene will allow higher frequencies and flexible devices.

 

[anexanhume]

Here's a good overview of where the players are placing their bets (FD SOI, FinFET, III-V): http://www.eetimes.com/document.asp?doc_id=1262887&page_number=1

I wouldn't get excited about graphene just yet.

 

[blackened23]

Here's a good overview of where the players are placing their bets (FD SOI, FinFET, III-V): http://www.eetimes.com/document.asp?doc_id=1262887&page_number=1

I wouldn't get excited about graphene just yet.

Intel's FinFET has already proven superior to planar in every respect thus far, I don't think that article has a basis in reality.

 

[Idontcare]

DARPA looks at future technologies more than any other company or organization. So if the director of DARPA is saying there is meaning to it. Colwell also says economics and not just physics will bring the death of Moore's law.

Actually DARPA doesn't do nearly as much in the field of semiconductors as the NSF (national science foundation).

When it came time to do serious research work on ultralow-k dielectrics for BEOL metallization targeting the sub-10nm era we had to seek out a grant through the NSF because DARPA isn't interested in research, they are interested in development (taking pre-existing researched ideas and implementing them into functional field models).

Personally I would not use DARPA as a bellwether or harbinger of anything. They do some interesting stuff but they are by no means experts in the science of semiconductors nor the economics of business.

 

[Exophase]

Intel's FinFET has already proven superior to planar in every respect thus far, I don't think that article has a basis in reality.

In that article the only one hedging anything on planar is IBM, and that's clearly for much smaller geometries. No one is saying that Intel's 22nm finFETs aren't superior to the planar processes currently out.

 

[anexanhume]

In that article the only one hedging anything on planar is IBM, and that's clearly for much smaller geometries. No one is saying that Intel's 22nm finFETs aren't superior to the planar processes currently out.

And Ivy Bridge was noted for its inferior overclockability, although that's a very niche segment.

 

[Exophase]

And Ivy Bridge was noted for its inferior overclockability, although that's a very niche segment.

And Haswell for even worse overclockability - it's hard to say how much is intrinsic to the process and how much is due to different design goals by Intel.

 

[blackened23]

And Ivy Bridge was noted for its inferior overclockability, although that's a very niche segment.

The 22nm with FinFET process is designed for low leakage and in that respect it is far better than planar. These devices are going into low power mobile devices, so low leakage (= low power consumption) is obviously the proper path forward for mobility and efficiency. This is what the market for mobile products and enterprise products demand. Efficiency. Intel's process delivers in spades in that respect thanks to the low-leakage characteristics.

Overclocking has nothing to do with, anything, really. I don't know what has to do with this. The fact of the matter is Intel's 22nm with FinFETs will still be better than TSMCs 20nm.

 

[anexanhume]

The 22nm with FinFET process is designed for low leakage and in that respect it is far better than planar. These devices are going into low power mobile devices, so low leakage (= low power consumption) is obviously the proper path forward for mobility and efficiency. This is what the market for mobile products and enterprise products demand. Efficiency. Intel's process delivers in spades in that respect thanks to the low-leakage characteristics.

Overclocking has nothing to do with, anything, really. I don't know what has to do with this. The fact of the matter is Intel's 22nm with FinFETs will still be better than TSMCs 20nm.

I don't doubt that it's better positioned for mobile. The fact remains it's inferior to their planar processes in overclocking. There's always trade-offs. It's not some magical panacea of processes. They probably felt safe losing their overclockability given their dominant position in the market over AMD. I guarantee they still enjoy selling processors to enthusiasts for 400 to 1000 a pop though.

 

[blackened23]

I don't agree with "bad" overclocking. Everyone is just spoiled by Sandy Bridge, chips like that come once every 10-15 years. My 3770k OC'ed to 4.7ghz fairly effortlessly which is a huge increase over stock speeds. To say that overclocking is "poor", that really isn't the case whatsoever - people just have inflated visions of 1.5GHz overclocks being sustainable over the course of many production generations. Probably not going to happen. It didn't happen in the past, either - once in a while a great chip would be released for overclocking to not be replicated for many years. The celeron 300A comes to mind, among others.

Regardless, intel is designing products for markets that sell. Enterprise, data centers, corporations, and mobile. Desktop sales dip by 20% per quarter, while the other areas are still growing, so it's obvious that their uarch focus will be for efficiency. It's the smart business move and the proper path forward as mentioned.

 

[ehume]

I would agree that Haswell is not a bad overclocker. It seems that its nominal temp is closer to its ceiling.

Also, the change from solder to TIM under the IHS means that we are more likely to run up against irremediable thermal limits before we hit stability limits. People who believe otherwise need to see how well their chips do when using AVX2.

 

[N-A-N-0]

As remarkably intelligent as DARPA and Colwell are, far more intelligent than any of us, perhaps especially a newbie like me, I've got to add this:

Mr. Colwell is insisting on either 7 or 5 nm being the smallest possible shrink... Both of those will require EUV it seems:

As anyone who has been following the chip-baking industry knows, getting down to 7nm without the advent of commercially usable extreme ultraviolet (EUV) lithography will be a daunting task – Intel CTO Justin Rattner admitted as much to The Reg just this May.
But EUV is proving to be an elusive technology, as GlobalFoundries CEO Ajit Manocha told SEMICON 2013 this July. "We all know that EUV is late," he said. "We desperately need EUV, and EUV is still not ready."
Another chip-process researcher, speaking at that same conference, has gone so far as to have given up on EUV. "I'm not working on EUV at all," said Laurent Miller, CEO of Leti, the nanotechnologies arm of the French research-and-technology organization CEA. "Absolutely not, because I don't believe in it."
Exactly what inspires James to have such faith in the continuance of Moore's Law, she didn't say.

Source: http://www.theregister.co.uk/2013/09/10/intel_reveals_14nm_pc_declares_moores_law_alive_and_well/

She then goes on to talk about how she feels graphene is exciting but won't be possible for several generations or so is the thinking.

See how many different informed opinions there are about this? If graphene is possible then there's your further die shrinks below 5 nm, and if EUV is possible there's your 7/5 nm. No one person or organization can know for sure just yet how small we can really go and what's doable.

 

[Idontcare]

See how many different informed opinions there are about this? If graphene is possible then there's your further die shrinks below 5 nm, and if EUV is possible there's your 7/5 nm. No one person or organization can know for sure just yet how small we can really go and what's doable.

That is the reality with research. Research is not to be confused with development.

What you have captured in your post is commentary from lots of experienced people who are experienced in the area of development, not so much in the area of research.

Consider this - the original development of the transistor was not done by a group of business people looking to develop a new product by aggregating a bunch of existing ideas from research. It was a small team of 3 lead scientists who were simply doing basic research.

Likewise with the invention of the integrated circuit, again just a couple of people (two, to be precise, and they weren't even working together or for the same company) who decided to create something that no one else seemed interested in working on.

Look at the Finfet, sure Intel refined it and put it into mass production but it wasn't developed or even researched internally, it is the brainchild of one person working with a very small team of equally bright scientists nearly 15 yrs ago.

The world is full of people who can follow someone else's example. Show them what must be done and they can execute, but they can't do anything if no one is leading them or telling them what to do.

This is the problem with listening to experts in the area of development, they can't think outside the box they are given by the researchers. So to them everything is a glass ceiling with defined limitations (limitations defined by the edge of their knowledge of what researchers have apprised them of).

If you really are curious about scaling and where things are going you have to engage a researcher, not a developer, and pick their brains over a pint or two.

 

[Abwx]

The fact of the matter is Intel's 22nm with FinFETs will still be better than TSMCs 20nm.

Others are not waiting iddling...

Taiwan Semiconductor Manufacturing Co. has released three silicon-validated reference flows within the open innovation platform (OIP) that enable 16 FinFET systems-on-chip (SoC) designs and 3D chip stacking packages.

http://www.xbitlabs.com/news/other/...ivers_16FinFET_and_3D_IC_Reference_Flows.html

 

[blackened23]

Others are not waiting iddling...



http://www.xbitlabs.com/news/other/...ivers_16FinFET_and_3D_IC_Reference_Flows.html

Yeah, TSMC's so called "16nm" is really their 20nm with FinFETs. And it won't be ready for a long, long time. Now if you don't call that anything other than market deception, then I don't know what to say. It's TSMC 20nm with FinFETs. But for some reason TSMC calls it "16nm". Pretty funny eh?

Oh yeah. I'd imagine paying for that 16nm will be difficult when only giants like Apple will be able to buy in initially. Each new TSMC node will be incredibly cost prohibitive, and not everyone will be able to handle it. Even nvidia's CEO made mention of how incredibly cost prohibitive 20nm and beyond will be. What i'm getting at is, even by the best estimates - most corporations cannot simply flick a switch and say "Voila! We'll do 20nm". Nope. Won't work like that. The cost is absurdly high, and will be even higher for FinFETS - that R+D costs billions IIRC. And TSMC won't eat that cost, they pass it on to willing buyers which I would imagine to be primarily Apple and Qualcomm initially. AMD? Their profit situation has to change and change fast if they want first dibs on new nodes. Nvidia? They could probably handle it, but they sure won't enjoy buying into 20nm and beyond. Using GPUs as an example, initial trajectories show that GPUs "die shrunk" to 20nm will cost nearly double just by virtue of being 20nm. I don't think consumers will be ready and willing for that to happen.

http://www.extremetech.com/computin...y-with-tsmc-claims-22nm-essentially-worthless

Nvidia deeply unhappy with TSMC, claims 20nm essentially worthless

 

[N-A-N-0]

I just realized...

Didn't we reach commercial 22 nm last year. Now Intel's shipping a 14 nm laptop by the end of the year?!

That was....fast.

 

[redhotiron2004]

I just realized...

Didn't we reach commercial 22 nm last year. Now Intel's shipping a 14 nm laptop by the end of the year?!

That was....fast.

Any source?

 

[Arkaign]

Any source?

I think he's confusing limited ES units with units available for sale :

http://www.anandtech.com/show/7309/intel-14nm-progress-update-broadwell-airmont-on-schedule

We *should* see some 14nm stuff floating around very very limited this year (most likely to see them internally at Dell/HP/etc for validation/testing) but not buyable/final until '14. Still fast, but not 1 year fast.

 

[tweakboy]

Yes in time we will have 2nm cpus lol

 

[N-A-N-0]

I think he's confusing limited ES units with units available for sale :

http://www.anandtech.com/show/7309/intel-14nm-progress-update-broadwell-airmont-on-schedule

We *should* see some 14nm stuff floating around very very limited this year (most likely to see them internally at Dell/HP/etc for validation/testing) but not buyable/final until '14. Still fast, but not 1 year fast.

Yes, Intel has said that they'll be shipping their first 14 nm chips ship this year.

I assume for sales this Ultrabook and further 14 nm sales will occur next year but CEO Brian Krzanich was vague.