Intel’s Stacked Nanosheet Transistors Could Be the Next Step in Moore’s Law

[BiG StroOnZ]

A process that builds two transistors—one directly atop the other—will boost chip density. NMOS and PMOS devices usually sit side-by-side on chips. Intel has found a way to build them atop one another, compressing circuit sizes.

 

Quote

 

 

 

 

The logic circuits behind just about every digital device today rely on a pairing of two types of transistors—NMOS and PMOS. The same voltage signal that turns one of them on turns the other off. Putting them together means that electricity should flow only when a bit changes, greatly cutting down on power consumption. These pairs have sat beside each other for decades, but if circuits are to continue shrinking they’re going to have to get closer still. This week, at the IEEE International Electron Devices Meeting (IEDM), Intel showed a different way: stacking the pairs so that one is atop the other. The scheme effectively cut the footprint of a simple CMOS circuit in half, meaning a potential doubling of transistor density on future ICs.

 

The scheme starts by using what’s widely agreed to be the next generation transistor structure, called variously nanosheet, nanoribbon, nanowire, or gate-all-around device depending on who’s involved. Instead of the main part of the transistor consisting of a vertical fin of silicon as it does today, the nanosheet’s channel region consists of multiple, horizontal, nanometers-thin sheets stacked atop one another. Intel engineers used these devices to build the simplest CMOS logic circuit, an inverter. It requires two transistors, two connections to power, one input interconnect, and one output. Even when the transistors sit side-by-side, as they do today, the arrangement is very compact. But by stacking the transistors and adjusting the interconnects, the inverter’s area was cut in half. Intel’s recipe for building stacked nanosheets is called a self-aligned process because it builds both devices in essentially the same step. 

 

It starts with repeated layers of silicon and silicon germanium. This is then carved into a tall narrow fin, and the silicon germanium is etched away to leave a suspended set of silicon nanosheets. Ordinarily, all the nanosheets would go to form a single transistor. But here, the top two nanosheets are connected to phosphorous-doped silicon with the aim of forming an NMOS device, and the bottom ones to boron-doped silicon germanium to yield PMOS. Robert Chau, Intel senior fellow and director of components research, says, “Once we get the hang of it, the next step will be to go after performance.” That will likely involve improving the PMOS devices, which right now lag behind the NMOS in their ability to drive current. The answer to that problem is likely the introduction of “strain” to the transistor channel, says Chau. The idea is to distort the silicon crystal’s lattice in such a way that charge carriers (holes in this case) flit through faster. Intel introduced strain into its devices back in 2002. In separate research at IEDM, Intel showed a method of producing both compressive strain and tensile strain in nanoribbon transistors.

 

 

 

Source 1: https://spectrum.ieee.org/nanoclast/semiconductors/devices/intels-stacked-nanosheet-transistors-could-be-the-next-step-in-moores-law

Source 2: https://www.hardwareluxx.de/index.php/news/hardware/prozessoren/54998-iedm-2020-intel-will-durch-gestapelte-transistoren-deren-dichte-verdoppeln.html

 

While new GPU and CPU rumors are a dime a dozen. This is some of the more interesting stuff, IMO, as it tackles an issue that will plague the technology space very soon. The inevitable end of the traditional Moore's law observation, that has essentially been the bread and butter of computer performance improvements the past few decades. Supposedly, there have been some recent results thus far (employing this technology) that haven't proven to be very promising (apparent inconsistencies towards the upper layers) but there are other organizations that are also pursuing the same research under other names (e.g. complementary FETs, or CFETs). Therefore, I expect more positive progress is to eventually happen (especially considering these other organizations are tackling it somewhat different than Intel is, giving variation to the implementation).

[leadeater]

Wonder what the envisioned time to market for this is, sounds like a rather complicated technical challenge to get it mass production ready. That's also after it's fully developed too. 

[Prodigy_Smit]

Intel execs may be stupid but their engineering team is damn smart.

[straight_stewie]

Wait, so Intel is investing more money in trying to beat Moore's law than in competing with AMD?

That's an interesting strategy.

[AndreiArgeanu]

Does that mean that cpu's will become THICC?

[Prodigy_Smit]

1 minute ago, AndreiArgeanu said:

Does that mean that cpu's will become THICC?

THICC CPU's is better than non-thicc CPU.

 

I doubt they will be much thiccer than they are now since that would cause many cooler incompatibilities.

[AndreiArgeanu]

1 minute ago, Prodigy_Smit said:

THICC CPU's is better than non-thicc CPU.

 

I doubt they will be much thiccer than they are now since that would cause many cooler incompatibilities.

I doubt compatibility will be an issue, if cpu's were thick enough that current mounting mechanisms aren't compatible, they can just include a new bracket in the box for this.

[f22luke]

I have been expecting this for a while now, but the issues are more than just consistency between layers.  The heat density is bound to be an issue unless we are also going to include some integrated cooling channels.  It will be interesting to see the different approaches taken by the different fabs. 

[DoTr]

I don't think that this will become a reality till the end of this decade, as Intel haven't even made a prototype chip yet, like photonics

[porina]

4 hours ago, leadeater said:

Wonder what the envisioned time to market for this is, sounds like a rather complicated technical challenge to get it mass production ready. That's also after it's fully developed too. 

Random speculation: maybe the generation after 7nm. Presumably 7nm would be well on its way being implemented into production for such early research to fit in. 

 

4 hours ago, straight_stewie said:

Wait, so Intel is investing more money in trying to beat Moore's law than in competing with AMD?

I don't see a difference in that. As long as Intel has fabs, they will need to continue improving their process to both catch up again with AMD and try to beat them. They have to spend money for the future for that to happen.

 

If you're talking about beating AMD today, I don't see any legal ways to really do that. Intel's problems of today are not something that can be solved in the short term by throwing money at it. They have their plans to recover, but they take time. I'll be looking at Rocket Lake and Alder Lake performance very closely.

[Lightwreather]

2 hours ago, porina said:

I'll be looking at Rocket Lake and Alder Lake performance very closely.

I dunno about rocket lake tho, I kinda feel like Rocket lake is essentially a stopgap for Intel and they're putting more of their efforts into Alder Lake. They litterally announced it via a Blogpost (If my memory serves me right)......A BLOGPOST! But anyway, I'll be still be looking at it hopeful

[porina]

21 minutes ago, J-from-Nucleon said:

I dunno about rocket lake tho, I kinda feel like Rocket lake is essentially a stopgap for Intel and they're putting more of their efforts into Alder Lake. They litterally announced it via a Blogpost (If my memory serves me right)......A BLOGPOST! But anyway, I'll be still be looking at it hopeful

I don't recall exactly what happened at the time. I guess you're referring to Intel sometimes posting info first on Medium than their own website.

 

Still, it remains the first mainstream desktop microarchitecture change since Skylake. There are things that are kinda known, such as it is based off Sunny Cove which you can already get in Ice Lake mobile, and we're all familiar with the 14nm process. So at the same time it is both known and yet unknown exactly what it will do due to the combination. I think it will be the first time people will get to try such architecture outside of power limits. It might still be a variation of 14nm we first saw on Broadwell over 5 years ago, but it is 14nm going out with a bang. Hopefully not a thermal surplus based one.

 

Alder Lake is more unknown. Finally 10nm makes it to desktop. It will be at least one more microarchitecture on (depending on how you count them) and will likely bring with it DDR5 support into the mainstream. The big/little core thing... I hope they're well advanced working with Microsoft for decent OS support as soon as it comes out. DDR5 and mixed cores wont just be new for Intel, but they're new for x86 in general.

[WereCatf]

7 hours ago, leadeater said:

Wonder what the envisioned time to market for this is, sounds like a rather complicated technical challenge to get it mass production ready. That's also after it's fully developed too. 

The fact that stacked designs are also harder to cool than unstacked ones will also affect what it's good for and thus what market they'll aim it at.

[williamcll]

Cooling's going to be a pain though.

[straight_stewie]

8 hours ago, porina said:

I don't see a difference in that. As long as Intel has fabs, they will need to continue improving their process to both catch up again with AMD and try to beat them. They have to spend money for the future for that to happen.

 

If you're talking about beating AMD today, I don't see any legal ways to really do that. Intel's problems of today are not something that can be solved in the short term by throwing money at it. They have their plans to recover, but they take time. I'll be looking at Rocket Lake and Alder Lake performance very closely.

Well, you're right. It's six in one, half a dozen in the other.

What I meant to say was that it might be the case that Intel is, and has been, intentionally willing to lose the short game in order to compete better on the long game:

If you look at what AMD is doing, they aren't investing in themselves. Intel is till way better off in the meantime. Intel has been making (including in 2020), about a 25% profit margin on roughly $80B of revenue. In the meantime, AMD has been making a roughly 16% margin on roughly $7B of revenue, and that's only for the last two years, before that they were consistently operating at a loss. As a result, Intel has about $13B on hand, while AMD has about $1.5B on hand.

AMD has been using primarily TSMC to produce their processors for them, which means they now have to compete with Apple (who is wayyyyyyy wealthier than AMD. Like not even in the same league.) for TSMC 5nm production capacity.

All I'm saying is that, if I'm Intel and I'm betting on x86 staying more popular than ARM for the next two decades, I'm keeping my FABs and heading straight for 5nm.

[inteli7.Ti]

Samsung is doing the nanoribbon with their MBCFETs, but this stacked CMOS technology is quite interesting. Heat dissipation is still probably going to be a challenge when designing these though. 

 

I can't wait to see which emerging technology is going to be the future: quantum computing, graphene transistors, carbon-nanotubes, or this. At least the consumer will always benefit

[Sauron]

14 hours ago, Prodigy_Smit said:

THICC CPU's is better than non-thicc CPU.

 

I doubt they will be much thiccer than they are now since that would cause many cooler incompatibilities.

we're still talking about sizes on the scale of a few nanometers so you won't perceive any difference in thiccness

[Kisai]

On 1/3/2021 at 11:05 AM, straight_stewie said:

Well, you're right. It's six in one, half a dozen in the other.

What I meant to say was that it might be the case that Intel is, and has been, intentionally willing to lose the short game in order to compete better on the long game:
 

Doubtful. It's likely that this has been on the backburner, given how much lead time is required to build a new fab. Intel just likely botched 10nm and are now behind by three process nodes relative to TSMC and Samsung. If the smart people decided "screw it, it's aim ahead of TSMC and start building 3nm and don't do any further work on 8/7/6/5nm" that would make sense, so they can put out the new fab tech.

 

Though it still seems like some people running the company aren't recognizing the position they put the company in when all these meltdown/spectre bugs started showing up, and didn't fix them at all on the 14nm chips.

 

 

[porina]

1 hour ago, Kisai said:

Doubtful. It's likely that this has been on the backburner, given how much lead time is required to build a new fab. Intel just likely botched 10nm and are now behind by three process nodes relative to TSMC and Samsung. If the smart people decided "screw it, it's aim ahead of TSMC and start building 3nm and don't do any further work on 8/7/6/5nm" that would make sense, so they can put out the new fab tech.

Again it depends on how you count the nodes but I'd argue Intel are at most 2 nodes behind. Given Intel 10SF seems to be loosely competitive with TSMC 7nm, TSMC shipping 5nm now is one node advantage, and I'd give them another node worth due to how much longer they had been shipping 7nm. We'll have to see when Intel is able to ship 7nm, given they said that was delayed and may be about 2 years out from now. They will surely be working on their 5nm too but I'm not aware of any news on that front. Process nodes are worked in parallel and independent of each other, so delays on one don't directly delay others.

 

1 hour ago, Kisai said:

Though it still seems like some people running the company aren't recognizing the position they put the company in when all these meltdown/spectre bugs started showing up, and didn't fix them at all on the 14nm chips.

There were various updates as each generation of 14nm offering was released, although you could argue they were minor. A bigger fix would require a bigger change, which is what the already existing new cores are for. Desktop certainly isn't their priority for pushing the technology so we only get to see something much different from Skylake with the impending Rocket Lake.

[rcmaehl]

This will just 4x transistor count instead 2x it. I call it rcmaehl's law.

[bcredeur97]

I could see this causing problems in terms of heat output (the transistors at the bottom layer won't get much cooling). Will probably only be suitable for high core count, low clock speed situations.

[straight_stewie]

4 hours ago, Kisai said:

If the smart people decided "screw it, it's aim ahead of TSMC and start building 3nm and don't do any further work on 8/7/6/5nm" that would make sense, so they can put out the new fab tech.

 

Though it still seems like some people running the company aren't recognizing the position they put the company in when all these meltdown/spectre bugs started showing up, and didn't fix them at all on the 14nm chips.

Just to point this out, node size is a meaningless term. TSMC 7nm node size brings a transistor density of 96.5 mtr/mm² while Intels current 10nm process (that they use for some mobile chips) gives them 100.76 mtr/mm² . A higher number is better here. What I mean to say, is that Intels 3nm behind process produces a transistor density competitive with the current state of the art 7nm design. Or, in other words, TSMC transistor node size names do not, in any way, relate to actual feature sizes on die. 

TL;DR on the above paragraph: Node sizes between different manufacturers should not be considered to be directly comparable when an allegedly 3 generation behind process is producing a higher transistor density.

Beyond that, Intel isn't actually worried about having the smallest processor node. Intel is, and should be, worried about producing the most profit. Which they are. They are operating at an order of magnitude more revenue than AMD, and they are also operating at an order of magnitude higher profit margin than AMD. Believe it or not, 2020 was one of Intels highest profit years ever. This is easily fact checkable as both Intel and AMD are publicly traded companies.

AMD is barely getting along, even if the general market feeling is that they are the most popular right now. Financial data shows that AMD runs an absolutely terrible business that can't possibly perform well in the long run. Seriously, look at their financial data, then look at Intels. Then realize that AMD now has to compete with Apple (the richest company ever), for manufacturing capability, all while their 10 times richer competitor doesn't have to compete for manufacturing capacity.

My argument is not about whether Intel processors are currently worse/better than AMD processors, my argument is about whether Intel is running a better business than AMD, and that they might actually know how to wield that business. Statistics really talk here, and Intel still has wayyyyyyy higher financial performance than AMD.

[Kisai]

11 hours ago, straight_stewie said:

Just to point this out, node size is a meaningless term. TSMC 7nm node size brings a transistor density of 96.5 mtr/mm² while Intels current 10nm process (that they use for some mobile chips) gives them 100.76 mtr/mm² .

Does that even matter. It's not like this topic isn't brought up every single time Intel releases another 14nm chip. There are no 10nm chips to buy. What is the point of going "oh Intel's 10nm..." when they aren't selling them. Desktop and Servers aren't getting 10nm.

 

And it doesn't matter if AMD or Intel are doing good or poorly in this regard, Intel is behind, and because their fingers are in a lot more diverse things (eg storage, networking, FPGA's, etc) they can actually afford to weather the cluckstertruck that their 10nm process is. But at this point if they're not putting desktop cpu's on 10nm, it doesn't matter, the marketing has spoken for Intel. AMD's chip shortages proves just how far Intel is behind when AMD is charging more for a chip in the same performance tier and has been entirely unavailable the entire year.

 

 

[porina]

1 hour ago, Kisai said:

Desktop and Servers aren't getting 10nm.

Yes they are. End of year Alder Lake for desktop and mobile. Ice Lake server is delayed, no idea when it is now supposed to be ready but largely irrelevant to this forum audience anyway.

[Astronautical]

surely the marketing team will call this quantum wire technology.

I would