Chip industry leverages massive compute resources and AI for virtual sandboxes.
Key Takeaways
Digital twin technology is drawing significant attention across the chip industry, even though it’s still unclear who will own or manage them, what the optimal levels of abstraction will be, and how they will be connected.
Nevertheless, investments in digital twins are rising. Early versions are beginning to roll out, and various segments of the chip industry are experimenting with them, in part to see how they can be used and in part so they don’t get broadsided by new or existing competitors. If digital twins live up to their promise, they will:
That’s just for starters, too. The long-term vision across a broad swath of companies, both within and outside the chip industry, is to provide pre-silicon insights from initial architecture to time zero in manufacturing, and then post-silicon from the field back to the fab or packaging house. This involves an enormous amount of data, which in turn will require a massive number of compute cycles and storage.
A digital twin can go big or small. It can represent an entire car or a city, or a subsystem, chip, or even an individual process under a combination of workloads or conditions, and all of this before a chip or system ever reaches production.
“A digital twin is like a system of systems, where you can stitch pieces together,” said Eli Roth, smart manufacturing product manager at Teradyne. “It has evolved out of largely isolated simulation and emulation environments, and now digital twins are being stitched together all the way from design through packaging and test. They’re like SOLIDWORKS models on steroids that can fit into the NVIDIA Omniverse 3D model. Where this gets really cool is with chiplets and heterogeneous integration, where the coupling is getting tighter and tighter between dies, substrates, materials, and thermals, and any kind of little warpages that cause expensive failures. So if you can figure out what’s going on before they go into those expensive, tightly coupled systems, that would be very helpful.”
Others agree. “These are game-changing ideas,” said Vincent Chu, senior consulting manager for Advantest Cloud Solutions. “People talk about digital twins of the fab processes, but they tend to limit this only to the front-end process. If you combine it with the back end — because ATE has the golden standard of the device — then you can make the simulation of the front-end process more accurate. And by feeding forward across insertions within this digital twin environment, we can develop models and do many kinds of what-if scenarios to make the back-end testing process more efficient.”
This could include multi-variant screening to improve throughput and quality, which is particularly important in multi-die assemblies. “We are making AI chips, and those chips typically include heterogeneous integration,” Chu said. “The challenge is how we can ensure the quality of the dies before they are assembled. These are very big packages, combining several to many dies. If you don’t screen out the one die with a problem, it will cause you to scrap the whole package.”
While investment is higher prior to manufacturing, the savings in terms of yield and fewer RMAs (return merchandise authorizations) due to higher reliability and more domain- and workload-specific designs can be significant. Still, orchestrating how all the pieces go together is non-trivial.
“When a wafer is tested at the fab, they don’t know where those chips are going to go to be packaged, assembled, and tested, so they’ve got to know where to send the material. That’s not something that’s gone on,” said John Kibarian, CEO of PDF Solutions. “This is like a low-code or no-code way of connecting major apps. You need PLM systems, ERP systems, MES systems, engineering data, manufacturing yield data, and ultimately design automation information. You need to be able to get that information in order to take the human out of the loop so AI can make the appropriate decisions. That’s really what you’re trying to do.”
Table 1: Uses and benefits of digital twins. Source: PDF Solutions/SEMICON WEST
Mixing disciplines
Where digital twins are today is still a long way from the vision of where they can go in the future. It will take multiple steps to get there, starting with the initial goal of reducing variability in manufacturing processes. That, in turn, will help improve reliability from time zero throughout the expected lifetime of a chip.
“With test we’ve been seeing this shift left for years, but it’s now getting to the extreme point,” said Lee Harrison, director of automotive IC solutions at Siemens Digital Industries Software. “Becoming a DFT engineer, you’re also becoming a functional architect, as well. We’re using PCIe, we have processes controlling test, we’ve got security enclaves. So as a DFT engineer, you’re having to do full subsystem architecture design with functional buses, embedded software, and understanding how to train PCI and UCIe. One of the things we’re ramping internally is having the DFT engineers become functional designers, because when you move into these bigger designs, such as those with chiplets, the control and monitoring and overall setup of the test infrastructure is now done through a kind of small functional domain, rather than it being controlled by random pins on the outside of the device.”
Digital twins can be used to test the DFT infrastructure and constructs, as well. “For the first time ever we had DFT guys at DVCon (Design And Verification Conference),” Harrison said. “We can dry run the whole test infrastructure on the ATE and prove it works before you go to silicon.”
This is roughly equivalent to starting with a Google Earth view, then drilling down to the impact of a workload-specific thermal gradient on a 2nm transistor. It requires connecting a lot of pieces that never went together in the past, and manipulating those pieces to obtain an optimal, or at least workable, solution under different stresses. This already has been done on a limited basis with large simulations, but the ability to manipulate various processes and swap in and out different components, such as chiplets developed at different process nodes or different interconnects or memories, is new.
“We can simulate from atoms to airplanes,” said Adam Cron, distinguished architect at Synopsys. “A lot of digital twinning is simulation and performing things without actually building them. We already simulate functional patterns before we tape out. That’s all digital twinning. If things don’t work, then either don’t build it, or at least you know that you’re building something that’s broken and you know why. Fabrication is an expensive and time-consuming endeavor. It’s much faster to digital twin it before pulling the trigger on manufacturing.”
Toward better results
Achieving the full promise of digital twins will require significant coordination across different industries, and in some cases, numerous segments within those industries.
“We’re in a kind of point solutions zone right now,” said Cron. “We have the knowledge assistants today. That’s pretty much rolled out. We’re at the generative place where scripts are being cranked out automatically, too. If you need a particular piece of collateral, we’re at a point where we’re starting to put the features in the tool so you can ask something on the screen to create that for you and put it into your flow. And we’re starting to move into agentic areas where a DRC violation could be recognized, and then tooling could fix it for you. Pretty soon, agents will talk to agents. That’s already being done in captive environments.”
Digital twins also can be used to improve individual manufacturing processes. “They can help derive greater value from the things that we measure, both incoming and outgoing, as well as help our tools perform better,” said Sean King, product manager in Onto Innovation’s Enterprise Software business unit. “With our own inspection and metrology tools, we are collecting more data, getting it in one place, finding inconsistencies, and making everything jive together in a central platform so you can understand how they’re performing and matching with each other. So you can think about a digital twin for predictive maintenance and making sure our tools are operational. On the customer side of that same problem, our tools are the inputs for what their twins and models might be for quality measurements and understanding the actual material they’re producing. If they can’t trust the output of our tools, how can they trust their models?”
Big challenges here include the integration and organization so that it can be used effectively and efficiently to identify, prevent, or solve a problem. “Time to result has to be part of this,” said King. “That comes down to what’s the problem we’re trying to solve. How much value is tied to it? Is it in-house? Is it on a big, expensive server somewhere else or in the cloud? Are you trying to build something that’s too big and over-fitting the problem? Then it may be hard to even identify where problems come up. It’s easier to pinpoint a problem, like if drift is happening, and how best to take action. But what happens in a multi-supplier ecosystem? What if we don’t talk and share the same way?”
Ironing out those differences will be a challenge. “There’s a big European automotive project called CHASSIS (Chiplet-based Hardware Architectures for Software-Defined Vehicles), and a significant part of the project is to build a digital twin of the entire platform so that all the different vendors that are providing chiplets into the system can virtually plug in their chiplet and see if it works,” said Siemens’ Harrison. “Does all of the test infrastructure connect up and work? Can we run all the DFT infrastructure once it’s inserted into the overall system? The last thing you want is to have this great stack of chiplets that may work perfectly fine functionally, but you find that something’s wrong on the DFT. It didn’t get hooked up properly, so I can only test half of it. So you have this great product, but it’s complete junk. You can’t put it in that car because it’s only half-testable.”
There also is a problem of connecting different types of data, which is where AI comes into the picture. “What we’re finding is that people are taking an agentic approach,” said Jon Holt, worldwide fab applications solutions manager at PDF Solutions. “So you have legacy systems and siloed systems out there, and each one is communicating through a PLC or an EDA controller, and another through MES systems. You may have something as simple as a data sheet delivered with the product that you load into the system and it becomes digitized. We started treating each of these as an agent, and then you don’t have to communicate all the information for the IP. You only need to communicate from these agents that sit in place at the point of use for the information you need.”
Holt said the goal is to bring all of the data from these agents together and to automate them in an agentic workflow. “It’s an ability to build on all this legacy infrastructure that we put in place, and then start utilizing it with the capabilities that GenAI or LLMs can now bring,” he said. “One of the keys is securing that data pipeline. Whatever form that takes, you have to have sensors in place that represent the physical world at the granularity needed to support your results or conclusions. So that may be sampling your environment once an hour, or it may be sampling every millisecond on an RF pulsed deposition tool.”
Standards will help. SEMI has been sharing information on digital twins for the past half-dozen years, running workshops about their role in manufacturing and supply chain resilience. What’s changed since the initial concept is the availability of nearly unlimited compute resources in the cloud, and the addition of AI to glue different data types together. That allows digital twins to be deployed on a large scale, or a tiny but extremely deep scale, and to be able to drill down or out.
“A digital twin can simulate on several levels,” said Advantest’s Chu. “The test cell itself can have a virtual identity. For example, we can have a virtual representation of the tester and the test. The virtual tester can run a test program offline, and it can manipulate the historical data logs in a simulation. If you change the conditions, what will the outcome be? Another level is the device itself. For example, we can have this virtual silicon so that you can run an initial test plan before the silicon is out. If you have the twins of other testers on the test floor, you can have the model predict, based on the current test conditions with adaptive testing, the yield or the throughput. So you can build another model to simulate the operation of the fleet of testers and determine how it compares to the production plan. After the silicon is out, the real testing begins, but there is a trend toward tuning the device. You can use a Verilog model to simulate the device, and then you run a test program to simulate production situations. The data can be synthetic, based on your knowledge of the device from the same family, where you have some historical production data, but not from this new device.”
The same concepts also can be extended to understanding and tracking how circuits age, which is essential in safety- and mission-critical applications. If a particular workload increases the utilization of a chip, or parts of a chip, there may be a higher incidence of electromigration over a shorter time span than if that same chip is used for a different workload.
“You might have a correctible model that shows some kind of degradation in an automobile,” said Synopsys’ Cron. “A model would say that your performance or your leakage will degrade at some particular levels, and that’s the model that the chip itself knows. And then, using SLM and DFT techniques in a kind of holistic way, you could validate that model every time you turn the car on or turn it off, or every microsecond in the field, and then pre-emptively figure out that this chip will last as long as it’s supposed to, or that it’s not lasting as well as the rest of these chips. That moves the digital twin out into the field and makes it very relevant.”
With good data and a well-defined digital twin model, these kinds of predictions can become significantly more precise. “If you can measure, you can predict. And if you can predict, you can prevent,” said Nir Sever, senior director of business development at proteanTecs. “Measurement over time allows you to understand the rate of degradation, and therefore you can extrapolate on the time point of failure. Once you do that, you can predict and send an alert ahead of time because you don’t want to wait for the last moment.”
This can happen in real-time with the right models and data. “If you have a workload with 100% test coverage and you subject your devices to a reliability test under this absolute maximum workload, then assuming your models are correct, you should not be in a situation where a perfectly good device will fail ahead of time,” Sever said. “But there are too many assumptions. In reality, you can test that your models are as accurate as they could be, and therefore you should spend more on testing. But you cannot avoid lifetime monitoring to detect developing or emerging errors before they cause problems.”
Digital twins also can be used to determine if on-chip monitors are actually detecting anomalous performance degradation, said Harrison. “If I add all these monitors on the chip, is it actually giving me useful information? One of the things you can do with the digital twin is model some of the effects you’re trying to detect. You can model some forms of degradation and over-voltage. On the security side, we’re looking at using our digital twin to understand what kind of side channel attacks we can model, as well. It’s really nice to be able to try out a whole bunch of stuff before we actually commit to silicon.”
Looking forward and back
The possibilities for how digital twins will be used are almost unlimited, which is why they’ve been on the industry’s radar for the better part of a decade. This merging of data also helps explain why last year PDF Solutions acquired secureWise and Synopsys bought Ansys. Companies are positioning themselves for the next phase of this technology, and they are investing heavily to get there.
What’s changed is that enough pieces of technology are in place — sufficient compute resources, data management and data mining tools, and AI/ML infrastructure — that digital can finally start to be used in meaningful ways.
“Imagine you’ve got an RMA,” said Teradyne’s Roth. “Here’s a group of bad devices. What happened? What were they doing in production three months ago. You can go look at all the production data, the test results that described it, and what was actually going on in the flow. My flow is really complex. What did I burn in and what did I not burn in? How did I get to these decisions? And the guy doing that RMA isn’t the test engineer who built the test. He’s trying to discern what was going on the test. These test programs are big and complex. There may be 20 or 30 people working on them, and for them to understand what’s going in is like a mystery. But if I can take that device, the SKU, the character I need, I can go back and rerun it in a virtual environment and debug it. And that would be pretty awesome.”
The ability to sift through data more quickly and apply it in unique ways at any step of the design through manufacturing flow is a big leap forward. “You want to feed forward from the fab manufacturing process from the front end to the back end, because if you want to fine-tune the processing at the front end you would like to reference the performance at the back end,” said Advantest’s Chu. “That depends on how comprehensive your model is. You can have very comprehensive twins at the front end, and you also can have twins at the back end. Then you have the data flowing across all stages, from the front end to the back end, and do the simulation.”
Much of this is in the early stages today, and many challenges still need to be addressed. But the possibilities for improved processes, more reliable systems, and cost reductions at multiple levels are very real and significant.
Related Reading
Secure Data Sharing Becoming Critical For Chip Manufacturing
Driven by a plethora of benefits, data sharing is gradually becoming a “must have” for advanced device nodes and multi-die assemblies.
AI In Test Analytics: Promise Vs. Reality
Turning big data into reliable yield insight requires more than good models.
New Frontiers In Fault Detection And Classification
AI is enabling timelier and more accurate data, but AI-based command and control has yet to appear.
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
Great topic for an article. Addressing the applications of a newly surfacing technology.