|5nm Design Progress|
Improvements in power, performance and area are much more difficult to achieve, but solutions are coming into focus.
July 17th, 2018 - By: Ann Steffora Mutschler
Activity surrounding the 5nm manufacturing process node is quickly ramping, creating a better picture of the myriad and increasingly complex design issues that must be overcome.
Progress at each new node after 28nm has required an increasingly tight partnership between the foundries, which are developing new processes and rule decks, along with EDA and IP vendors, which are adding tools, methodologies, and pre-developed blocks to make all of this work. But 5nm adds some new twists, including the insertion of EUV lithography for more critical layers, and more physical and electrical effects that could affect everything from signal integrity and yield to aging and reliability after manufacturing.
“For logic, the challenge at 5nm is to properly manage the interaction between the standard cells and the power grid,” said Jean-Luc Pelloie, a fellow in Arm’s Physical Design Group. “The days where you could build a power grid without considering the standard cells are over. The architecture of the standard cells must fit with the power grid implementation. Therefore, the power grid must be selected based on the logic architecture.”
At 5nm, IR drop and electromigration issues are almost be impossible to resolve if this interaction has not been properly accounted for from the beginning.
“The proper power grid also will limit the impact of the back-end-of-line(BEOL) effects, primarily the simple fact that via and metal resistances increase as we continue to shrink into 5nm,” Pelloie said. “In addition to considering the logic architecture for the power grid, a regular, evenly distributed power grid helps reduce this impact. For designs using power gates, those gates need to be inserted more frequently to not degrade the performance. This can result in an increase of the block area and can reduce the area gain when shrinking from the previous process node.”
The migration to each new node below 10/7nm is becoming much more difficult, time-consuming and expensive. In addition to the physical issues, there are changes in methodology and even the assumptions that engeers need to make.
“You’ve got a higher-performance system, you’ve got a more accurate system, so you can do more analysis,” said Ankur Gupta, director of product engineering for the semiconductor business unit at ANSYS. “But a lot of engineering teams still have to move away from traditional IR assumptions or margins. They still have to answer the question of whether they can run more corners. And if they can run more corners, which corners do they pick? That’s the industry challenge. When running EM/IR analysis, it’s a strong function of the vectors that the engineering chooses to run. If I could manufacture the right vectors, I would have done it yesterday, but I can’t.”
Choosing the right vectors isn’t always obvious. “Technology is quickly evolving here as a combination of voltage and timing that can intelligently pick or identify the weak points,” Gupta noted. “That’s not just from a grid weakness perspective, but from the perspective of grid weakness plus sensitivity to delay, to process variation, to simultaneous switching—sensitivity to a bunch of things that ultimately can impact the path and cause a failure.”
This changes the entire design approach, he said. “Can the margins be lowered, and can flows be designed so they are convergent throughout the entire process? Could I potentially use statistical voltages instead of a flat guard band IR drop upfront and then potentially go down to these DVD waveforms — really accurate DVD waveforms — and a path to get high levels of accuracy in the signoff space? Could I potentially analyze chip, package and system? Could I potentially do all of this analysis so I don’t waste 5% margin coming from the package into the chip? At 7nm, we were talking about near-threshold compute, as in some corners are at NTC, not the entire chip, because you look at the mobile guys and they’re not always running sub-500. There are some conditions and modes where you’ll be running at sub-500, but at 5nm because of the overall thermal envelope and the overall power consumption budget, the mobile guys are probably going to be running all corners sub-600 millivolts.”
It’s not just mobile. The same is true for networking, GPUs, or AI chips, because a lot of these designs have the same total power envelope restrictions. They are packaging so many transistors into a small space that the total power consumption will dictate the max operating voltage. “You can’t burn enough power if you’re upgrading, you don’t have enough power to burn at 800 millivolts or so if the entire chip now starts to operate at 600 millivolts or lower,” Gupta said. “Then you take tens of sub-500 millivolt corners and that becomes your entire design, which puts you in the land of ‘must-have these [analysis] technologies.’ Next to 7nm, we are seeing the variation impact at 5nm in early versions of spice models is worse.”
Many of these technology and design issues have been getting worse for several nodes.
“There are more challenging pin access paradigms, more complex placement and routing constraints, more dense power-ground grid support, tighter alignment necessary between library architecture and PG grid, more and tighter electromigration considerations, lower supply voltage corners, more complex library modeling, additional physics detail in extraction modeling, more and new DRC rules,” said Mitch Lowe, vice president of R&D at Cadence. “Obviously, EUV lithography is critical, which does reduce but not eliminate multi-patterning challenges and impacts. While some things are simplified by EUV, there are some new challenges that are being addressed.”
The EDA community has been working on these issues for some time. “We are at the stage to see leading EDA solutions emerge,” Lowe said. “Much more work is ahead of us, but it is clear the 5nm technologies will be successfully deployed.”
The EDA ecosystem is heavily investing in continuous PPA optimization and tightening correlation through the integration of multiple common engines. One example is combining IR drop impacts with static timing analysis (STA) to manage the increasing risks inherent in using traditional margining approaches at 5nm, Lowe said.
Other changes may be required, as well. Mark Richards, marketing manager for the design group at Synopsys, noted that 5nm is still immature, with various foundries at different points in their development plans and execution.
“Outside of the main foundry players, which are aggressively moving to deliver a production ready flow in a very short timeframe, research is being conducted on new architectures for transistors, because to some degree the finFET is being stretched to its limit toward the 5nm node,” Richards said. “This is why there is somewhat of a tailing off in top-line performance benefits, as reported by the foundries themselves. As you deploy fin-depopulation to meet area shrink goals, this necessitates an increase in the height of the fin to mitigate the intrinsic drive reduction. That brings intrinsic capacitance issues and charging and discharging those capacitances is problematic from a performance perspective,” he explained.
Samsung and GlobalFoundries have announced plans to move to nanosheet FETs at 3nm, and TSMC is looking at nanosheet FETs and nanowires at that node. All of those are gate-all-around FETs, which are needed to reduce gate leakage beyond 5nm. There also are a number of nodelets, or stepping-stone nodes along the way, which reduce the impact of migrating to entirely new technologies.
Fig. 1: Gate-all-around FET. Source: Synopsys
At 5nm, a very strong increase in both electrical and thermal parasitics is expected, Dr. Christoph Sohrmann, advanced physical verification at Fraunhofer Institute for Integrated Circuits IIS, said. “First of all, the FinFETdesign will suffer from stronger self-heating. Although this will be taken care of from the technology side, the reduced spacing is a design challenge which cannot entirely be coved by static design rules. The enhanced thermal/electrical coupling across the design will effectively increase to a point where sensitive parts of the chip such as high-performance SerDes may suffer from a limited peak performance. However, this depends strongly on the use case and the isolation strategy. Choosing the right isolation technique — like design-wise and technology — requires more accurate and faster design tools, particularly focused at the parasitics in those very advanced nodes. We expect to see a lot of new physical effects which need to go into those tools. This is not too far away from quantum scale. To get the physics right, a lot of test structures will be required to fit the models of those novel tools. This is a time consuming and expensive challenge. Fewer heuristical models are also expected, with more real physical approaches in the models. On top of that, the foundries will be very cautious about those parameters and models. All future standards in this area need to account for this, too.”
Then, for 3nm and beyond, there will have to be a move to new transistor structures to continue to achieve the performance benefits that are expected at new nodes, Richards said. “With the increased introduction of stepping-stone nodes, you’re basically borrowing from the next node to some degree. When you throw a node in the middle, you kind of borrow from the next node as far as what the projected benefits will be. That’s what we’re seeing in some of these boutique nodes in between, but they are important given end-customer demand, and they do enable our customers to hit aggressive product-delivery windows.”
For any new process node, tremendous investment is required by the EDA and IP community to make sure tools, libraries and IP are aligned with the new technical specifications and capabilities. Part of this is the process design kit that design teams must adhere to for that new node.
Across the industry, there is a lot of development work ongoing for cell and IP development. “In real terms, the biggest amount of change and development work materializes in or before the 0.5-level PDK,” Richards said. “Generally, from 0.5 onward, there is a reduced delta to what the PDK would be expected change. So normally everything’s done. Between pathfinding, 0.1 and 0.5, the big majority is done, then the rest tapers off because by that point you’ve had numerous customers do test chips, so the amount of change required is reduced. Beyond that point it’s really about building out and maturing the reference flows, building out methodologies, and really bolstering those in that 0.5 to 1.0 timeframe to make sure the promise from the scaling and the performance perspective are going to be realizable in real chips.”
Fig. 2: 5nm nanosheet. Source: IBM
To move or not to move
Another consideration many semiconductor companies are currently facing is not to migrate to the next node, or at least not so quickly, or whether to move in completely different directions.
“New architectures are going to be accepted,” said Wally Rhines, president and CEO of Mentor, a Siemens Business. “They’re going to be designed in. They will have machine learning in many or most cases, because your brain has the ability to learn from experience. I visited 20 or more companies doing their own special-purpose AI processor of one sort or another, and they each have their own little angle. But you’re going to see them in specific applications increasingly, and they will complement the traditional von Neumann architecture. Neuromorphic computing will become mainstream, and it’s a big piece of how we take the next step in efficiency of computation, reducing the cost, doing things in both mobile and connected environments that today we have to go to a big server farm to solve.”
Others are expected to stay the course, at least for now.
“Many of our customers are already engaged in 5nm work,” Richards said. “They’re trying to work out what this node shift brings for them because obviously the scaling benefits on paper are very different to the scaling benefits that they can realize in a real design — their own designs with their own specific challenges — and so they’re trying to work out what is a real scaling, what are the real performance benefits, is this tractable, is it a good methodology to use, and a good plan from a product perspective.”
Today, the expectation for early adoption of 5nm will be mobile applications, he said. “TSMC itself quoted a 20% bump from N7, and, to my knowledge, an unknown bump from 7++ . Realistically, mobile is a good application, where area – slated to be 45% vs. N7 – is really going to provide a big differentiation. You’ll get the power and performance benefits that are also important but with the latest IP cores growing in complexity and area, you need to have the freedom to develop a differentiated cluster and aggressive area shrinks will allow for that.”
The key metrics are always performance, power and area, and the tradeoffs between all of those are becoming more difficult. Increasing performance brings a subsequent increase in dynamic power, which makes IR drop more challenging. That requires more time to be spent tuning the power grid so designs can deliver enough power, but not kill the design routability along the way.
“The key thing with power really is how to get power down to the standard cells,” said Richards. “You just can’t put the cells close enough together because it spoils the resources with power grid. This means working early in the flow with power and its implications. On an SoC design you might see very different power grids, depending on the performance requirements of each of the blocks on the SoC. It’s not just a one size fits all. It must be tuned per block, and that’s challenging in itself. Having the analysis and the sign-off ability within the design platform is now going to become more and more important as you make those tradeoffs.”
At the same time, the margin between the threshold and the operating voltages is now so small at 5nm that extra analysis is a must.
TSMC and Samsung both have mentioned extreme low-Vt cells, which are paramount for really pushing performance at 5nm, where the threshold and operating voltage very close together.
“The nonlinearities and the strange behaviors that happen when you’re in that phase need to be modeled and captured to be able to drop it as low as possible,” he said. “Obviously LVF (Liberty Variation Format) was required at 7nm, for when the operating voltage was getting very, very low and very close to the threshold, but now even when you’re running what you would not consider a extremely low power design with extremely low voltage Vt cells effectively, you’re back in the same position. You’ve closed that gap again, and now LVF and modeling those things is very important.”
Inductance, electromagnetic effects
Indeed, with the move to 7nm and 5nm, the trends are clear: increasing frequencies, tighter margins, denser integrated circuits, and new devices and materials, stressed Magdy Ababir, vice president of marketing at Helic.
He noted during the recent Design Automation Conference, a panel discussed and debated such concepts as: where and when should full electromagnetic (EM) verification be included; whether ignoring magnetic effects leads to more silicon failures during development; whether the methodology of applying best practices to avoid EM coupling and skipping the tedious EM verification part should still be a valid practice; if this methodology is scalable to 5nm integrated circuits and below; if the dense matricies resulting from inductive coupling and difficulty of simulations are the main reason why industry did not widely adopt full EM simulations; and what can be done in-terms of tool development, education, and research to lower the barrier for industry to adopt full EM simulation.
“The panel members all agreed strongly that the full EM analysis is becoming fundamental in at least some key parts of any cutting-edge chip. A panelist from Synopsys was of the opinion that is needed in some key places in a chip such as clocking, wide data busses, and power distribution, but not yet in mainstream digital design. An Intel panelist was of the opinion that for current chips, applying best practices and skipping using full EM simulations still works, however this methodology will not scale into the future. A panelist from Nvidia simply stated that EM simulations is a must with his very high frequency SERDES designs, and a panelist from Helic agreed strongly here, and showed examples of unexpected EM coupling causing failures in key chips. The moderator was of the opinion that magnetic effects are already there strongly and have been very significant in integrated circuits for a while, but the difficulty of including magnetic effects into simulation, and manipulating very large and dense matrices resulting from inductive coupling is the main reason full EM verification is not mainstream yet. Everyone agreed that not including EM effects in verification results in overdesign at best and potential failures,” Abadir offered.
In the end, the panel agreed that there is a need for significant improvement of tools that handle EM verification, better understanding of magnetic effects, and significant research on how to protect against EM failures or even adopt designs that benefit from magnetic effects. The panel also agreed that current trends of higher frequencies, denser circuits, and scaling of devices combined with the exploding penalty on a chip failure, makes including full EM verification imperative, he added.
An additional challenge at 5nm is the accuracy of waveform propagation. Waveform propagation is notoriously expensive from a runtime perspective, and as a result needs to be captured throughout the entire design flow. Otherwise, the surprise at sign-off would be that the design is too big to close.
The typical way to solve these problems is by adding margin into the design. But margining has become an increasingly thorny issue ever since the advent of finFETs, because dimensions are so small that extra circuitry reduces the PPA benefits of scaling. So rather than just adding margin, design teams are being forced to adhere to foundry models and rules much more closely.
“Foundries do provide models of devices that represent corner models,” said Deepak Sabharwal, vice president of IP engineering at eSilicon. “In the past, you were told the corner models capture the extremes of what would be manufactured, but that is no longer the case. Today, there are still corner models, but there are also variation models, both global and local. Global variation capture the global means of manufacturing, such as when multiple lots are run at a foundry, each lot is going to behave in a certain manner and that is captured as part of my global variation. Local variation models represent when I’m on a die and my die has a Gig of elements. Then I have the middle point of my distribution, and what the outliers are on that distribution.”
At 5nm, both the global plus the local variation must be considered, because they are incremental.
“At the same time, these kinds of analysis are experience-driven,” Sabharwal said. “How much margin do you add, and also make sure you do not go overboard? If you design for too much of your sigma, you ended up being uncompetitive. That’s what you have to watch out for, and that’s really where the experience comes in. You have to make sure you put in enough margin that you can sleep at night, but not kill your product by putting in too much extra area that you don’t need to put in.”
More than ever, 5nm brings together a range of new challenges. “When you think about the billions of components sitting on that chip, it explains why the size of the teams needed to build these chips is now increasing as you flip from one generation to the next. It’s all these challenges that are coming our way. These problems are going to remain, where people will come up with techniques to resolve them and just continue business as usual. Engineering is really that art of building stuff that will work reliably all the time,” Sabharwal said.
7/5nm Timing Closure Intensifies
The issues may be familiar, but they’re more difficult to solve and can affect everything from performance to yield.
Tech Talk: How to improve confidence that designs will work properly in less time.
Design Rule Complexity Rising
Total number of rules is exploding, but that’s only part of the problem.
Quantum Effects At 7/5nm And Beyond
At future nodes there are some unexpected behaviors. What to do about them isn’t always clear.