90 nm? Is smaller possible?!

[MetalStorm]

I have been thinking, the next big improvement in CPU speed will be when they go 90nm, now to do this they use photolithography, and I know there are plans for x-ray lithography in the future and that should also bring an improvement in CPU speed due to a smaller architecture but thats going to be a while... But I have an idea...

I know im not an electrical engineer and to put it simply, I only know a little about how CPUs are made but I still think this is a good idea! I will elaborate:

To me this seems pretty basic, but im sure its not, but I will continue anyway. If you think of the process they use to make 130 nm chips at the moment, the tracks are complete and are the right thickness and consistency, now I was thinking if you were to only put a very small dusting almost of silicon atoms on your to be silicon wafer you obviously wont have a working chip at the end of it... so what if you were to squash the entire chip, so that the atoms you spread on there in a light dusting before, now form tracks? I was thinking you could do the squashing by putting your wafer in a very high pressure chamber, reaching several thousands of atmospheres - and with sufficient cooling and an inert medium in which to convey the pressure to the wafer you could significantly reduce the size of the the architecture as the dusting of atoms at 130 nm would be squashed together forming tracks at say 65nm?

I have already said that I dont make CPUs and my idea seem simple in theory but is probably a lot harder to carry out successfully, but as I said, its just and idea...

Comments anyone?

[SWE_HELIX]

Uhm, didnt a Soviet sattelite designed to land on Venus melt at a 100 atmospheres in a laboratory?

[MetalStorm]

Obviously with pressure there is heat, so that is why I suggested appropriate cooling as well, so it doesnt melt!

[method().man]

Getting silicon oxide the size of "almost atoms" would probably be much harder than 90 nm lithography. Remember, we'd need something to get the material that small in the first place. Also, arranging the atoms to form a matrix would be very difficult as well. You're basically talking about arranging a material at the atomic (molecular) level - as far as I know, something like this is possible on a limited basis now, but is exceedingly difficult and expensive.

[MetalStorm]

Yeah, as I said, I didnt know much about how CPUs are made, it was just meant to be food for thought though... maybe one day they will be able to do what I suggested? If they do, I want some money for it!

[zabomb4163]

The issue right now is finding lasers that can etch architectures that small.

[Overclocker550]

I have heared the limit is about .07 microns then we would need a totally different approach

[SWE_HELIX]

And what kind of cooling would that be? I dont think that such a cooling exists on this planet..

[MetalStorm]

Originally posted by SWE_HELIX
And what kind of cooling would that be? I dont think that such a cooling exists on this planet..

Okay, so are you suggesting watercooling doesn't exsist?? Its not the pressure that determins the heat generated, its how fast you do it... If you were stupid enough to slam the chips in from 1 to 100 or 1000 atmospheres in ten seconds, no you dont have a chance of cooling it, but if you were to do it over a period of hours with a bit of watercooling it would easily be possible...

Think of this - the Marianas trench is about 11000 meters deep - that means there is a LOT of water above you - 11km of it even, and as you probably know, water isn't that light, at the bottom its about 16,000 lbs per square inch, thats a lot of pressure - but it isn't boiling is it! May I suggest you think before you speak?

[OC Detective]

I thought IBM were in talks with AMD on the next generation after 0.09 being 0.065 which would be manufactured in Singapore by UMC?

Edit
Quote from ebnonline.com

At present, few details are know about how AMD and UMC will fabricate 0.065-micron processors on 300-mm wafers, except that it will involve common copper interconnect processes and low-k dielectrics, which will be used by both companies in their own plants outside of the Singapore venture later this decade. UMC and AMD officials said they were not sure yet whether the 0.065-micron technology will employ extreme ultraviolet (EUV) lithography or 157-nm optical tools for minimum feature sizes.

[NookieN]

Originally posted by OC Detective
I thought IBM were in talks with AMD on the next generation after 0.09 being 0.065 which would be manufactured in Singapore by UMC?

They probably are (though they're gonna have to get 90nm working first). Intel announced development on their 65nm process (P1264) earlier this year. 45nm technology is the next likely step after that, and it seems feasible with advanced lithography techniques.

Ultimately, the limit for process size is the size of atoms (approx. 0.1nm). Of course, you can't position a single atom by using lithography. However, there are various quantum techniques that could be used. A few months ago, an independent research team announced a transistor using a single atom for the gate, so it's theoretically possible (but a long ways off in practical terms).

[Caffinehog]

Actually, the limit is molecules, not atoms. Once you get down to the atomic level, the motion of the atoms (as well as the positional uncertainty from quantum mechanics) makes it too unpredictable to work.

One of the hottest areas of chemistry now is making new molecules that act as individual electronics components. Resistors and diodes have been created recently. The biggest barriers to the use of these are self-assembly and connection to more traditional electrical components. The latter has been solved, at least technically, in several ways. Though real-world experiments have not been done yet. The former is a bit more challenging, but things are heading in that direction.

Honestly, I think we're looking at at least 10 years before these become a reality, though. Maybe 20...

[NookieN]

Originally posted by Caffinehog
Actually, the limit is molecules, not atoms. Once you get down to the atomic level, the motion of the atoms (as well as the positional uncertainty from quantum mechanics) makes it too unpredictable to work.

True, although since quantum effects are seen even in modern semiconductor manufacturing, they would definitely affect a singe molecule. If you can position a molecule, you can position an atom.

For anyone who's interested, here's an article that discusses atomic-scale transistors: http://www.sciencenews.org/20020810/bob9.asp

[FunkDaMonkMan]

Chips with extremely small traces would have to be cooled very well, b/c when atoms heat up, they will move around, and lose connection. This will be the biggest obstacle in producing chips with <1nm traces.

[Maxvla]

i think this is why we are seeing these mainstream computer companies starting their foray into watercooling. i believe for 65nm and smaller watercooling will almost be mandatory. i don't think dell wants a case with 5 vantec tornados in it which is what would probably be required for air cooling something that small. it would be way too loud for consumers when watercooling is more expensive yet nearly silent if properly done. i think the computer manufacturers are seeing this increase in temp and concentration of heat and realizing they will need to invest in watercooling soon or be left in the dust.

its very interesting though.

[MetalStorm]

I thought with each step down in size of the cpu, the less voltage it requires to run at a certain speed, so, they will still be running on 1.5v to 1.8v and the temperatures will be reduced surely - think of the thunderbird chips - 180nm 1.85 v for 1.4GHz - now a 1700+ DLT3C 1.5v for 1.463GHz...

[Overclocker550]

if they watercool, where will the water go? in a cd rom bay? a giant case with a bottom chamber for water? why not use chilled air instead? something like air conditioner

[verbatim]

I dunno if this is a silly idea or not but instead of spending all this money atm and having probs with 90nm (or doing this while 90nm is sorted) why dont they use the same technology but make the chips a little bigger?

If u made the intel chip the size of a AMD chip (or AMD a little bigger) wouldnt u be able to fit alot more circuitry onto the chip? This would of course be a new platform (new sockets etc) but I hear there is all sorts of problems getting 90nm off the ground. In the meantime this seems like a quick-fix solution with alot less R&D involved and quicker results for the companys.

I know nothing about how chips work tho...

[Caffinehog]

Originally posted by NookieN


True, although since quantum effects are seen even in modern semiconductor manufacturing, they would definitely affect a singe molecule. If you can position a molecule, you can position an atom.

True, but metal atoms are essentially free-moving and only weakly bound. While moleclules stretch, bend, twist, and move, they are bound together, and could be bound on the ends.

In short, with molecules, you don't know their position, but you do know approximately their position in relation to other components, since they are bound in a specific way. Metal atoms, on the other hand, don't have a specific structure, and you therefore lose that certainty.
Here (http://www.physics.vanderbilt.edu/pantelides/APL76-3448.pdf) is a link on some of the earlier research on this. I don't have the appropriate tools to find some of the more recent research, but by replacing a couple of the hydrogens on the benzene ring with other groups has created FAR better transistors.

[FunkDaMonkMan]

Originally posted by verbatim
I dunno if this is a silly idea or not but instead of spending all this money atm and having probs with 90nm (or doing this while 90nm is sorted) why dont they use the same technology but make the chips a little bigger?

If u made the intel chip the size of a AMD chip (or AMD a little bigger) wouldnt u be able to fit alot more circuitry onto the chip? This would of course be a new platform (new sockets etc) but I hear there is all sorts of problems getting 90nm off the ground. In the meantime this seems like a quick-fix solution with alot less R&D involved and quicker results for the companys.

I know nothing about how chips work tho...

The problem would be heat. Even a good watercooling system would have trouble keeping a double sized xp3000 cool. You'd get twice the power, but you would also get twice the heat from that small of an area. You would also have to run more volts through the thing because you are using twice as much material... raising heat more. Right now computer companies want cpu's that they can cool nearly silently, because the mainsteam public would perfer a silent computer over an extremely fast one.

[Krowa 02]

This is all interesting stuff, I know that one day we will reach the end of the line and we will be unable to scale any further with chips, but this is a long way down the road still. I think at that point we will just go on with multi core chips to make processing better, since speed will be maxed out.

[NookieN]

Originally posted by MetalStorm
I thought with each step down in size of the cpu, the less voltage it requires to run at a certain speed, so, they will still be running on 1.5v to 1.8v and the temperatures will be reduced surely

Yes a lower process size means a lower voltage. If all other things were equal, the heat dissipation will be lower. But when you start adding more transistors and/or increasing clock speeds, the current draw goes up... and heat along with it.


Originally posted by verbatim
I dunno if this is a silly idea or not but instead of spending all this money atm and having probs with 90nm (or doing this while 90nm is sorted) why dont they use the same technology but make the chips a little bigger?

As FunkDaMonkMan said, heat would be an issue. But it's not the only one. Bigger chips are also more expensive to make, because you can fit fewer dice on a wafer and yield drops as chip size increases. If chip companies are to keep increasing performance and decreasing cost, they must continue to improve their processes.

And the die on a P4 Northwood is actually slightly bigger (131 mm^2) than the die on an XP Barton (101 mm^2).

[MetalStorm]

Originally posted by Krowa 02
This is all interesting stuff, I know that one day we will reach the end of the line and we will be unable to scale any further with chips, but this is a long way down the road still. I think at that point we will just go on with multi core chips to make processing better, since speed will be maxed out.

Then comes quantum computing which will be nothing short of impressive to say the least - I just hope they can do it soon!

[The_Jizzler]

i was always under the impression that .08 is the limit for silicon. after that the gates/transistors are so small that there physically isnt enough material to keep elctrons in order. i remember reading a article on quantum and bio computing a few years back, and basically said .08 is the pinnacle for silicon. this may be different for other substrates like gallium arsenide, but the real future is like caffeine hog said. getting cells to compute, aka bio-computing. it will make everything now current look like a total joke supposedly.

[The_Jizzler]

double post please delete

[Caffinehog]

Bio computing... circuits that work like the brain... they're definately the future.

I see it like this, though:
Silicon hits its limits. Exotic materials, like molecules, find their way in, and light data, or photonics, starts to get integrated into these. Molecules and light are fundamentally affected by quantum mechanics, so that's where quantum mechanics comes in. Biocomputing will take quite a while... we have no clue how it works, so it will take quite a few new ideas for it to become a reality. I suspect a temporary surge of biocomputing, but when the principals are understood, they will be applied to the quantum/photonic computers. Cells may be extremely efficient, but molecules are cheaper, more environmentally tolerant, and you can fit a lot more into the same area.

[AS3]

QUESTION:90 nm? Is smaller possible?!
PART OF QUESTION:Comments anyone?

ANSWER: YES
ANSWER TO PART: WHEN MODERN HUMANS FIRST EVOLVED DID THEY THINK PC'S LET ALONE THE CONCEPT OF PC'S LET ALONE ELECTRICITY etc. "was possible?"

NO
:D

[btw, nice idea] nono...no comments on your theory

[Mr. Fri]

A few comments on the orignial question:

I know there are plans for x-ray lithography in the future...

1) x-ray litho has been off the semiconductor roadmap for years. There are new technologies like phase shifting, EUV, and (water) imersion lithography that are being evaluated and prove to be better (from a cost-of-ownership viewpoint) than x-ray.

QUESTION:90 nm? Is smaller possible?!

2) Offical lithrography roadmaps show nodes at 130nm, 90nm, 65nm, 45nm, and 22nm. So, It doesn't look like 90nm will be the wall.

3) BTW, some companies are developing a 1X printing techlology where they etch sub-90nm patterens on quartz plates then physically squish photoresist between the etched plate and a smooth wafer in order to transfer a pattern to the dice. At first glance this sounds crazy, but there are large programs working on this technology. It's jokingly called the squish-and-wish process!:D

[Overclocker550]

"2) Offical lithrography roadmaps show nodes at 130nm, 90nm, 65nm, 45nm, and 22nm. So, It doesn't look like 90nm will be the wall."


wow man, 45 to 22nm is less than half the size, why such a huge jump

[Mr. Fri]

I don't know but that's the roadmap. I'll see if I can find out why today.

[Mr. Fri]

My mistake. The litho roadmap goes from 65nm to 45nm to 32nm. At this point in time no one knows how we will do 32nm gates but the same could be said of current technology 10 years ago!

[Janissary]

I don't think that the production process will be the only limiting factor for better performance. How about going from 64 bit to something like 1024 bit. CPU's are extremely small and I am sure there is space for much bigger ones.

[futura2001]

Ummm, 1024 bit processors would be pointless for the everyday user. Not to mention slower.
All that a 1024 bit processor would do is cause the CPU to have to calculate integers as 1024 bit instructions. Instructions are built to be 32 and 64 bit instructions, and there is no need for the precision that 1024 would offer in virtually every scenario outside of a lab. Which means that an instruction that only needs 32 bit processing would take 32 times as long.
Yeah, that's better performance...

[OC Detective]

Originally posted by Mr. Fri
My mistake. The litho roadmap goes from 65nm to 45nm to 32nm. At this point in time no one knows how we will do 32nm gates but the same could be said of current technology 10 years ago!
That makes more sense as typically new generation chips are 0.7x the previous version - something you see when applying Moore's Law.

[NookieN]

Originally posted by OC Detective

That makes more sense as typically new generation chips are 0.7x the previous version - something you see when applying Moore's Law.

Just guess, but that's probably because doubling the number of transistors that you can fit into the same area requires shrinking the process dimensions by sqrt(2) (or 1.414 -- i.e. multiply by 0.707 as you say).

But I don't think there's any physical property that requires the process sizes to follow those numbers. They could be just about anything. A lot of memory chips are made on .15u, for example.

[Janissary]

Originally posted by futura2001
Ummm, 1024 bit processors would be pointless for the everyday user. Not to mention slower.
All that a 1024 bit processor would do is cause the CPU to have to calculate integers as 1024 bit instructions. Instructions are built to be 32 and 64 bit instructions, and there is no need for the precision that 1024 would offer in virtually every scenario outside of a lab. Which means that an instruction that only needs 32 bit processing would take 32 times as long.
Yeah, that's better performance...

Hmmm, then what about, lets say lots of 32 bit processors working within one big proc rather than todays multi processor workstations. It would not rise the bandwith physically, but it should give us a huge calculation power resulting in similar performance increase.

[OC Detective]

Originally posted by NookieN


Just guess, but that's probably because doubling the number of transistors that you can fit into the same area requires shrinking the process dimensions by sqrt(2) (or 1.414 -- i.e. multiply by 0.707 as you say).

But I don't think there's any physical property that requires the process sizes to follow those numbers. They could be just about anything. A lot of memory chips are made on .15u, for example.
Yes doubling through either smaller transistors, or better design process and originally bigger chip area!
No of course they dont need to but manufacturers will tend to follow this rule within a specific sector as a self fulfilling prophecy(after all Moore did start Fairchild and Intel!) but now in terms of scaling down through improved design and process and smaller transistors rather than as originally intended for scaling up.

[OC Detective]

Originally posted by Mr. Fri
My mistake. The litho roadmap goes from 65nm to 45nm to 32nm. At this point in time no one knows how we will do 32nm gates but the same could be said of current technology 10 years ago!

Extract from PC Magazine from last month.

"Companies like AMD, IBM, and Intel will continue using silicon to build smaller and faster microprocessors for at least another ten years, but not without the help of extreme ultraviolet (EUV) lithography, a new way of printing circuit patterns onto silicon that eschews lasers and lenses in favor of xenon gas and microscopic reflectors.

Tried-and-true optical lithography techniques that print patterns with features as narrow as 65 nanometers will extend Moore's Law into 2007. Only EUV can stretch it into the next decade, shaving feature widths to 32 nanometers.

When Moore made his seminal prediction in 1965, microprocessors were built with essentially the same optical lithography techniques used today, which rely on lasers and lenses to print circuit patterns onto silicon wafers. A laser shines ultraviolet light onto a mask — a tiny cutout of the pattern being printed — and as the light shines through the mask, it conforms to the pattern. Tiny glass lenses then reduce its wavelength.

To build smaller and smaller circuits, manufacturers have improved the precision of the laser and lenses, reducing the wavelength of the light hitting the wafer. Equipment used to build the Intel Pentium 4 and the AMD Athlon produces an ultraviolet light with a wavelength of 248 nm, printing circuit patterns with features around 130 nm wide. Later this year, Intel will move to a 193-nm optical system.

But optical lithography will soon reach its limit. "You run into severe materials problems when you drop below 193-nm wavelengths," says Gregg Gallatin, an IBM researcher. In order to develop a 157-nm optical system, which will debut in 2007, scientists had to construct lenses from entirely new materials. Glass wouldn't work. "When you get down to 157 nm, you have to use a single-crystal material called calcium fluoride," says Gallatin. "And it was a lot harder and took a lot longer to grow calcium fluoride with the required optical quality than people expected." Building lasers and lenses capable of wavelengths below 157 nm proved impossible.

Researchers sought out alternative forms of lithography, eventually settling on EUV. Rather than using a laser as a light source, an EUV system produces ultraviolet light by electrically exciting xenon gas. To hone the light, it uses specialized mirrors instead of lenses. By reflecting the light off these microscopic mirrors, the system narrows wavelengths to about 13 nm.

The EUV LLC Consortium, an Intel-led group that includes AMD, IBM, Infineon, Micron Technology, and Motorola, hopes to debut EUV around 2009, shrinking CPU feature widths to around 32 nm. But the technology needs fine-tuning. "It's still not clear that this will be a cost-effective solution," says Gallatin. "EUV has the technical capability, but it may cost a horrendous amount of money to put into production."

Intel Fellow Peter Silverman is confident that the technology will launch as scheduled. "EUV will be affordable for leading-edge companies," he says. "You don't need a lot of tools for the first generation, and there's time to get the cost down for the second generation." Chances are, Moore's Law will reach its golden anniversary."

[Caffinehog]

I don't think they have 10 years. Look at Prescott. Air-cooling can barely handle what intel is about to release. .65 will probably require water, as will .45. Anything under .45 will likely need exotic cooling solutions. The total power consumption is decreasing or staying about constant, but the power density is increasing rapidly.

[Overclocker550]

thats sad, maybe we should all slown down and try to make cpus run cooler first then faster

[L337 M33P]

Originally posted by OC Detective


Extract from PC Magazine from last month.

"
But optical lithography will soon reach its limit. "You run into severe materials problems when you drop below 193-nm wavelengths," says Gregg Gallatin, an IBM researcher. In order to develop a 157-nm optical system, which will debut in 2007, scientists had to construct lenses from entirely new materials. Glass wouldn't work. "When you get down to 157 nm, you have to use a single-crystal material called calcium fluoride," says Gallatin. "And it was a lot harder and took a lot longer to grow calcium fluoride with the required optical quality than people expected." Building lasers and lenses capable of wavelengths below 157 nm proved impossible.

Researchers sought out alternative forms of lithography, eventually settling on EUV. Rather than using a laser as a light source, an EUV system produces ultraviolet light by electrically exciting xenon gas. To hone the light, it uses specialized mirrors instead of lenses. By reflecting the light off these microscopic mirrors, the system narrows wavelengths to about 13 nm.



Just FYI: Some extremely high-quality refraction telecsopes use 2 or more elements (lenses in the front) with one or more being composed of flourite - a specially grown crystal with superior refraction properties and less chromatic distortion than any type of glass. These lenses are produced up to about 6" in diameter, but cost in the region of $4000 to produce. The capability is there but the cost of refining and implementing in a chip fab is probably going to be astronomical ( ;) pun intended)

I still don't get how the "wavelength" of the light is reduced by merely reflecting it - I smell something fishy, unless they are using some sort of photomuliplier that increases the energy of the photons somehow.

[OC Detective]

Originally posted by L337 M33P


Just FYI: Some extremely high-quality refraction telecsopes use 2 or more elements (lenses in the front) with one or more being composed of flourite - a specially grown crystal with superior refraction properties and less chromatic distortion than any type of glass. These lenses are produced up to about 6" in diameter, but cost in the region of $4000 to produce. The capability is there but the cost of refining and implementing in a chip fab is probably going to be astronomical ( ;) pun intended)

I still don't get how the "wavelength" of the light is reduced by merely reflecting it - I smell something fishy, unless they are using some sort of photomuliplier that increases the energy of the photons somehow.
A Japanese company is currently experimenting on this with a stepping system using a fluorine laser as the light source to create 45 nanometre circuit lines. Lens material is calcium fluoride.