中文 English

Impact Of GAA Transistors At 3/2nm

Some things will get better from a design perspective, while others will be worse.

popularity

The chip industry is poised for another change in transistor structure as gate-all-around (GAA) FETs replace finFETs at 3nm and below, creating a new set of challenges for design teams that will need to be fully understood and addressed.

GAA FETs are considered an evolutionary step from finFETs, but the impact on design flows and tools is still expected to be significant. GAA FETs will offer additional freedom to design teams to optimize their designs because there is no quantization. With finFETs, quantization in the fins limits the ability to balance drive current, leakage, and performance. As a result, different processes are needed for wider devices, which boost performance, or narrower devices for low-power applications. GAA FETs eliminate that problem.

The new gate structure sharply reduces current leakage. At 7nm and 5nm, leakage in finFETs is beginning to increase because the bottom side — the part connected to the silicon body — is not fully controlled. This was a key reason why finFETs were introduced in 2011. With planar transistors, even when a device was turned off, current would continue to flow between source and drain. As a result, designers were forced to utilize such approaches as power gating and other techniques to minimize wasted power.

But the transition from 2D transistors to 3D transistors created significant modeling issues. There was a spike in the number of parasitics that had to be considered. All told, it took several years to fully work out the implications of this new device structure, requiring significant changes in the development flow – particularly for analog devices.

Now, finFETs are running out of steam. At 5nm, finFETs are reaching the end of their ability to shrink and still provide meaningful scaling benefits. The number of fins has been reduced, and in practice cannot go below two. While fin width can be reduced, fin height has to be increased to compensate. New materials are being considered for the fin so that carrier mobility can be maintained, but the writing on the wall is clear.

So the primary focus of the industry is bringing the gate to the fourth side of the channel, creating a gate-all-around structure. By raising the channel in the transistor and creating a fin, which wraps the gate around the channel on three sides, the surface area between the gate and channel is increased.


Fig. 1: Planar transistors vs. finFETs vs. gate-all-around Source: Lam Research

Many articles have been written that describe these new structures and how to fabricate them (Moving To GAA FETs, New Transistor Structures At 3nm/2nm). The industry remains in a validation phase for the models and design flows, which will be required for these new structures at 3nm and below. Production is expected to begin in the 2022/2023 timeframe.

The impact
The good news is that the basic physics have not changed. The transistors still have all of the same elements as they always have. But their characteristics will improve, and some of the restrictions of the past will be lifted. It all comes down to the channel width. The wider the channel, the more electrons can flow and the faster the device can operate. But that allows for more leakage. A fully surrounded channel (sometimes called a nanowire) makes it difficult for electrons to escape. By stacking multiple nanowires on top of each other, you can have elements of both. Each wire can be tightly controlled, and multiple wires operating in parallel provide superior drive capabilities.

How big a disruption will this be for designers? “FinFET was the first device in the third dimension and there were a lot of parasitics around the Z dimension,” says Dusan Petranovic, principal technologist for Siemens EDA. “That was a big change. But gate-all-around (GAA) is more evolutionary. Even though there are a lot of changes, foundries believe that 90% of the process can be re-used and the BEOL has not been changed much. Nanosheets are also 3D, and there can be 3, 4 or 5 nanosheets. Even though it is a 3D structure, we can approximate that like a planar structure with the variable width of the sheets. People know how to approach this from an extraction point of view.”

Parasitic extraction is one the main areas that is impacted. “Essentially, it’s all about accuracy at this point because smaller transistors mean smaller wires, and the routing of these wires will be compact and crowded, impacting capacitance and the coupling capacitance between the wires,” says Hitendra Divecha, product management director in the Digital & Signoff Group at Cadence. “Smaller transistors have to be modeled accurately — we are talking about attofarad (aF) and almost 3D field solver-like accuracy for these parameters. For MEOL (middle end of line), due the proximity to the device itself, there will be new modeling features that will have to be implemented to accurately capture the impact on the timing of standard cells and EMIR. Besides parasitic RC values, RC topology will also matter for extraction accuracy.”

It is a progression. “They know what questions to ask,” says Carey Robertson, director of product management for Siemens EDA. “We had multiple generations of planar technologies, and as you went from planar to planar you knew what to ask. Now we’ve had a generation of 3D transistors that spawned a whole new group of questions, and so designers know what they need to investigate and to make sure they understand how it’s going to operate.”

With GAA FETs, performance is expected to improve by 25%, with power consumption reduced by 50%. With finFETs, both numbers have been roughly in the 15% to 20% range.

The addition of the gate on the fourth side provides more control. “The electrostatic controls on GAA and Vts are much more controlled,” says Aveek Sarkar, vice president of engineering at Synopsys. “This is important because at the smaller nodes we are seeing significantly more variability, especially for SRAM. So with GAA, we expect some of those to get more contained. But the variability and parasitic effects are going to be significantly higher.”

In addition, some of the problems introduced by finFETS will be relaxed. “You have the possibility to continuously vary the width of the nanosheet,” says Siemens’ Petranovic. “They can now be sized appropriate for different applications. If you need high switching speed, you might want to use wider nanosheets for more current. If you are designing SRAM cells, you may be more concerned about the area footprint. Libraries will be developed to explore this new degree of freedom. With finFET, we had discrete steps — 1,2,3 fin scaling. Now we can vary it continuously, and that new degree of freedom has to be exported into various tools like synthesis and place-and-route. There may be some parameterization of library cells to be able to optimize designs better.”

New challenges
With change comes uncertainty. There is even more potential for variability in these new devices. “This will be a bigger concern than in the past,” says Petranovic. “Part of it is just because you have smaller dimensions, and you have to deal with the effect of line-edge roughness and thickness variation. There will probably be new equipment that is used for this purpose. We will use EUV for edge roughness control, but still it will be a challenge.”

Line edge roughness is a factor because this can impede the flow of electrons. A new source of variability will be nanosheet thickness variation (STV). This causes variation in quantum confinement that impacts performance.

There are other changes, that while not directly to the GAA transistors, could be considered collateral damage. “We see the decreasing supply and threshold voltages, and the non-availability of thick oxide devices that result in transistors with lower breakdown voltages,” says Andy Heinig, group leader for advanced system integration and department head for efficient electronics at Fraunhofer IIS’ Engineering of Adaptive Systems Division. “That means transistors for classical output or driver cells aren’t available in such technologies. So chiplet approaches become necessary, where the GAA part is only responsible for the digital part, and other components in older technology nodes that can realize the input/output interface.”

Some analog components may still be necessary. “The industry will have to figure out how to do analog design in these processes, because anything interesting is going to have some analog content,” says Robertson. “That will probably have to be at higher voltage. The digital VDD of these chips will certainly come down, but there will be different voltage regions to accommodate other design styles.”

Challenges remain, though. “The finFET forced quantization, and that highly impacted analog circuits,” says Synopsys’ Sarkar. “That flexibility will become a lot more helpful in terms of what they can and cannot do. But there are things that are going to become more challenging. With the 3D topology, in terms of the capacitive and resistive models, are the scalable rules that we have historically used going to be sufficient and accurate for analog circuits? Do you need to have a different approach to getting the parasitic, especially at the local interconnect level? And how many RC parameters did you get?”

Other things are impacted just by scaling. “The cross section of the wires is smaller,” says Petranovic. “That means a significant RC delay increase, which is a potential bottleneck, and there are a lot of techniques trying to avoid this. One is to introduce new material for BEOL and even for MEOL. There is the introduction of air gaps for intermediate layers. There are schemes for reducing VIA resistance. Source/drain contacts are seeing increasing resistance. They have a notion of self-aligned gate where they are trying to put contacts directly on top of active devices.”

These changes will drive new kinds of analysis. “Thinner wires coupled with stronger drive strengths means we have to think about EMIR drop for MEOL — those wires that are very close to the transistor,” says Robertson. “Traditionally this was only done at the full-chip level and for power distribution.”

Again, these are incremental concerns. “There is no indication that additional layers will be introduced like we had when we made the jump to finFET with local interconnects and additional vias, which then translated into an explosion in parasitics,” says Cadence’s Divecha. “There will always be third-, fourth- or even fifth-order manufacturing effects that parasitic tools will have to model for accuracy purposes, so there will be more BEOL modeling that will have to be enhanced to ensure the impact on both timing and EMIR is minimized. There may be additional routing rules that will have to be done for place-and-route. However, from an extraction standpoint, the extraction of metal layers will continue to exist as they do today for finFET designs, but the focus will be more on accuracy and capacity.”

Power delivery network
Another area that almost certainly will be impacted is the power delivery network. Traditionally, this has been located in the metal stack built on top of the substrate.

PDN problems are growing. “The biggest problem with the PDN is the RC effect — Ohm’s Law degradation,” said Sarkar. “Then, there is the inductance effect. When you brought the chip and the package together, the Ldi/dt effect started to become high impact. Foundries started to provide more advanced decoupling capacitors, in addition to providing device level capacitors to suppress some of the noise and get a smoother supply noise profile. The challenge, especially with GAA, is you’re going to be packing a lot more devices in a square millimeter, and they are going to be switching more frequently. So is there some way you can short circuit and provide the current to the devices in an alternative manner?”

There are other power-related challenges, as well. “The decreased supply voltage can only be realized by an extremely stabilized supply network,” says Fraunhofer’s Heinig. “Different approaches are being discussed, such as on-chip regulators, backside supply using TSVs or different stacking options.”

What are back-side power supplies? “The idea is to move the power and ground wires beneath the transistors – on the back side,” says Petranovic. “Then through-silicon vias are used to supply power to the active layers. This is to reduce IR drop and noise on the signal wires, and to reduce congestion.”

That potentially adds a new form of analysis. “You now have back-side metal,” says Robertson. “Previously, you put the transistors on the substrate, and you pretty much ignored electrical effects between transistor and substrate. You did some rudimentary modeling. Now you essentially have transistors in the middle of lots of wires, instead of just at the bottom. That should reduce noise overall, but if you have a noisy power grid, you now have significant power grid interaction to the transistor. You will probably need analysis tools to validate the noise contribution of the power grid to the transistor, whereas previously the power grid would be on metal layers 13 and above, quite a bit of separation from those devices.”

This adds a new issues. “What kind of stress does this create?” asks Sarkar. “You have to supply the power on a regular interval over to the devices. You will create additional layers of stress in the silicon, and how you model some of that will become pretty critical.”

New models
Getting the right models is important. “Every new node gets more complicated and adds new technology effects have to be modeled,” says Petranovic. “EMIR, thermal, reliability, electomigration – all of these will get more complicated, but that would happen anyway just with scaling. For the device itself, it depends how accurately we need to model it. You have nanosheets vertically stacked, so the question is – can we approximate that to be something similar to planar with some vertical effects, or do we need to go inside the structure and extract some components? The right answer is to find the minimum detail necessary to accurately analyze the impact on performance.”

Getting it right is often an iterative process. “It’s not just the model themselves,” says Sarkar. “It’s also process development and device creation, which is where you have the transistor architects, the process integrators feeding into the people who are doing the first libraries, who are creating the first ring oscillator to see this is coming together and getting an early preview of how a block is going to look like. Finding out if there are certain things that we should be doing. The notion of Design Technology Co-Optimization is becoming even more important. How are we able to influence the various pieces that have resided in different teams within organization? If they are in different organizations, it’s even more challenging. How are we able to bring them together to have an early preview of these effects, and provide feedback to the process engineers and the architects on the left hand side of the equation, to help them help the right hand side in a more efficient manner.”

Without the appropriate level of accuracy, engineers have to over-margin their designs. “Today’s designer might need an extra 2-4 months to close the signoff loop,” says Divecha. “Extraction is a critical step in the signoff loop and we’ve heard from designers that while the extraction runtime varies based on design sizes and types, full flat extraction at these advanced nodes can take up to three days with some extraction tools. This puts an enormous amount of pressure on designers to achieve design closure in a timely manner to meet time-to-market pressures.”

The industry is currently trying to validate those models. “There’s two parts to this: one is developing the model, and then there is the analysis around it,” says Robertson. “Going from planar to finFET, to gate-all-around, there are new effects that need to be modeled, and I don’t know that we have quantified all of them. Using an example from the past – we didn’t care about proximity of planar transistors to wells. Around the 20-nanometer node, that became an important physical effect. I think we have an overall understanding of what needs to be modeled, but we need more test chips, more experiments to make sure that we’re capturing all of the physical effects in the model and once we do that, we can have analysis tools in place. The industry is going through a validation exercise.”

There is more to learn. “This has to happen as foundries and EDA vendors focus on making these types of devices mainstream,” says Divecha. “Having said that, whether you are doing digital designs or custom/analog designs, most of these requirements will be taken care of by EDA software, specifically extraction tools, and all the effects will be captured in foundry-certified techfiles.”

Conclusion
At this point in time, each of the foundries is looking at a range of possibilities. But based on early announcements, it would appear that there may not be a lot of commonalities between them. Each will have to work out which methodology works best for them and what provides the best yield.

Time will tell what will be the most successful. But the “good” news is that scaling is likely going to be the larger cause of pain, and not the change in structure of the transistors.

Related
Inside Intel’s Ambitious Roadmap
Five process nodes in four years, high-NA EUV, 3D-ICs, chiplets, hybrid bonding, and more.
GAA Knowledge Center
Top stories, special reports, blogs and white paper on GAAFETs
The Increasingly Uneven Race To 3nm/2nm
An emphasis on customization, many more packaging options, and rising costs of scaling are changing dynamics across the industry.
Moving To GAA FETs
Why finFETs are running out of steam, and what happens next.
New Transistor Structures At 3nm/2nm
Gate-all-around FETs will replace finFETs, but the transition will be costly and difficult.



2 comments

Jay says:

I wonder how would substrate play its role in GAAFets as the nanosheets do not seem to have any connection to the substrate in comparison to finfets/planar.

2R says:

You need to isolate it, to cut the parasitic channel (e.g. with Bottom Dielectric Isolation. Added parasitic capacitances have to be considered for AC performance evaluation). Regarding back-bias biasing to modulate threshold voltage, it is of course non existing (that said, in FinFET it was already practically nothing).

Leave a Reply


(Note: This name will be displayed publicly)