How much smaller can chips go? [article]

[Campy]

in ten years with 15 layers and hundreds of billions of transistors taping out a chip will costs well over 100 million dollars.

Why?

 

[bryanW1995]

because the physical properties of the silicon are what allow us to use it to make a transistor. Hence you need atoms.

uh, yeah, so you're saying that the only possible way to build a computer is by using silicon?

 

[SHAQ]

Exactly. Carbon nanotubes will be used after silicon.

 

[extra]

There's other materials that are supposedly suited to making faster cpus, like gallium arsenide, but I don't know much about that...perhaps one of our experts can give some info...

 

[frostedflakes]

GaAs devices are more expensive I believe, which has relegated it to more niche applications.

edit: Looking over the Wikipedia article it lays out the advantage and disadvantages pretty well. Apparently GaAs is also poorly suited for CMOS.

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

 

[aphorism]

Why?

the cost of masks is on a per transistor basis, unlike ic's where all transistors are made simultaneously. the cost of masks has been doubling for many years. a 65nm mask set is ~3million dollars, a 45nm mask set is ~6million, 32nm is 12million, etc.

also the number of masks is increasing because another metal layer is added every other process or so.

 

[cbn]

the cost of masks is on a per transistor basis, unlike ic's where all transistors are made simultaneously. the cost of masks has been doubling for many years. a 65nm mask set is ~3million dollars, a 45nm mask set is ~6million, 32nm is 12million, etc.

also the number of masks is increasing because another metal layer is added every other process or so.

So the masks aren't really getting more expensive at smaller nodes. It is just that these smaller nodes often have more xtors?

I am getting this right?

For example, 3 million xtors on 65nm process would cost the same to mask as 3 million xtors on 32nm process?

 

[Kakkoii]

As partly said by someone before, when we reach the limit of lithography and silicon we will have to transition to carbon nanotube interconnects and graphene gates, which have amazing electrical properties that themselves will be able to increase processing power vastly, apart from the fact of the fab size shrink. Much headway has been made in the past few years in making chips with these.

Random articles on it:
http://en.wikipedia.org/wiki/Graphene#Electronic_properties
http://en.wikipedia.org/wiki/Graphene#Graphene_transistors

100Ghz graphene transistor made by IBM
http://www.technologyreview.com/computing/24482/?a=f

http://www.hplusmagazine.com/articles/toys-tools/graphene-next
http://gadgetophilia.com/graphene-carbon-nanotubes-promise-cheaper-more-powerful-electronic-devices/


But again, then we reach the physical limit of atoms themselves, so we must then work on fabricating in cheaper ways, developing better architectures and coding ways, and building chips that have layers upon layers of cores all interconnected. Until the day comes when quantum computing is capable at room temperature at an acceptable hardware size and cost. How fast that day comes is obviously unknown though. But we are making fast strides in that area also.

 

[Campy]

We definitely still have a long long way to go in computing. When one paradigm is exhausted we'll simply move on to another.

 

[Kakkoii]

There will never be a single atom transistor, and you can quote me on that. There are few places in engineering where I can say never, this is one of them.

To have a transistor as we use them, AT LEAST 3 atoms are necessary (Base, collector, emitter). And frankly, it is highly unlikely that we will get to the state of a tri-atom transistor (unless, somehow we create a chemical compound that transfers electrons like a transistor.)

Or how about make a transistor that operates in a completely different manner? Gotta think outside the box.

http://en.wikipedia.org/wiki/Ballistic_transistor
http://www.rochester.edu/news/show.php?id=2585

"Due to its high electronic quality, graphene has also attracted the interest of technologists who see them as a way of constructing ballistic transistors."
http://en.wikipedia.org/wiki/Graphene#Graphene_transistors

But yeah, of course no single-atom transistor.

 

[Cogman]

Or how about make a transistor that operates in a completely different manner? Gotta think outside the box.

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

"Due to its high electronic quality, graphene has also attracted the interest of technologists who see them as a way of constructing ballistic transistors."
http://en.wikipedia.org/wiki/Graphene#Graphene_transistors

But yeah, of course no single-atom transistor.

Even these involve at least 2 atoms. While they may be different, they rely on having the two or more atoms pinch/unpinch the flow of electrons.

Again, no single atom transistor will ever exist.

 

[Kakkoii]

Even these involve at least 2 atoms. While they may be different, they rely on having the two or more atoms pinch/unpinch the flow of electrons.

Again, no single atom transistor will ever exist.

Uh, I don't think you read what I said at the bottom of the post..

"But yeah, of course no single-atom transistor."


and I wouldn't say they pinch/unpinch the flow. They merely direct the flow down one path or another using magnetism.

 

[Cogman]

Uh, I don't think you read what I said at the bottom of the post..

"But yeah, of course no single-atom transistor."

🙂 Yep, My mistake. I tend to focus on pretty blue links first.

 

[aphorism]

So the masks aren't really getting more expensive at smaller nodes. It is just that these smaller nodes often have more xtors?

I am getting this right?

For example, 3 million xtors on 65nm process would cost the same to mask as 3 million xtors on 32nm process?

yes, processing/repairing the mask design is a lot of the cost but there are some costs related to materials and other issues.

for example intel's 22nm node uses computational lithography. according to ACML it requires a sustained 10tflops for 24hours to process one layer. on 22nm this could be for only critical layers but in future processes it might be all of them. someone here could probably tell you the costs of operating/buying/maintaining several 10tflop machines, but that's not my thing.

also there are other ways of enhancing resolution. such as optical proximity correction or phase shift masks. all of these cost more but continue scaling.