AI Plays Multiple Roles Within EDA

AI’s infusion into our world may seem sudden and unexpected, but EDA has been quietly adopting it for more than a decade. What’s changed is that it’s now becoming more visible, thanks to increasingly powerful large language models (LLMs) and the need to apply them to increasingly challenging multi-physics problems.

Two fundamental shifts underlie AI’s increasing prominence. First, heat is becoming a barrier for larger package sizes and higher integration levels. As a result, power, thermal, and mechanical are now bound tightly to the physical aspects of the materials used for construction. Second, shift left — the desire to have better information available earlier in the flow in order to make more informed decisions — is being coupled with top-down design practices. Both of these changes bring the influences of physics into architectural decisions involving floor planning, partitioning, and integration choices for 2.5D and 3D assemblies.

This is where AI can help, and it does so from two directions. One is inside of EDA tools, making those tools more capable and versatile. The other is outside of tools, providing useful data to engineers. While they can coexist, these approaches use different forms of AI. In addition, AI can be used to help link all of this together by building suitable executable design and verification models to use as abstractions of the physical devices. Those use cases may contribute to enhanced methodologies, enabling better designs more productively.

[Editor’s note: When discussing AI within the context of EDA, the term “model” becomes confusing in that it is used for different things. The first is the AI model, such as a large language model. It is used by an inference engine to provide an answer based on a trained model. The second is an EDA model, which could be part of the design or verification (DV) framework. This uses an execution engine that responds to stimulus and may or may not utilize AI. It can become especially confusing when an AI model is used to create an abstraction of an EDA model. Attempts have been made to clarify this as much as possible.]

Inside or outside?
There are several factors that influence where AI is utilized. It could be used to replace or enhance an existing algorithm within the tool, or it could be more agentic and used to drive the tool as if it were an engineer. “One of the advantages of AI being outside is we can be tool-agnostic,” says Amit Gupta, vice president and general manager of Siemens EDA’s Custom IC Division. “Then the AI can call internally created tools, or third-party tools. The AI layer is doing intelligent calling of tools and iteration with the tools. So why do we put AI inside the tools? The advantage we get when we couple them at the solver level is there is often a lot more knowledge and information that you can train the AI with. You can get performance improvements as a result of that.”

Others agree. “We think of AI as a companion to our tools and as a way to augment, as opposed to replace or substitute,” says Anand Thiruvengadam, product management, senior director at Synopsys. “We do not see tools being out of the loop. Tools will always be in the loop, especially for signoff. We see the need for tools being there as a guardrail. You can have AI do a lot of the front-end stuff. You can have AI do a lot of prediction, optimization, etc., but ultimately, you will always need a tool in the loop to sign off.”

Agentic AI takes this in a different direction. “The nice thing about agentic AI, as a concept, is that you can consider it to be humans — just artificial,” says Michal Siwinski, chief marketing officer at Arteris. “You basically give it a higher degree of freedom to do specific tasks and then orchestrate those tasks to work in unison based on different levels of tradeoffs. How much freedom you want to give it is where the human in the loop is still going to remain part of the process. You don’t have a department of engineers where you can ask one to focus on performance and another to focus on area, or focus on throughput, or focus on overall schedule. You basically have this little team — it just happens to be AI agents versus human agents, and they all have some domain of expertise.”

Rapid progress is being made on this front. “There is a lot of tool iteration that needs to happen,” says Vamshi Balanaga, co-founder and CTO for Partcl. “For example, tool settings and updating RTL, based on testbench or PPA results, are best done via an agent that calls the tools akin to how a human behaves today. For now, the agents aren’t great at this, but they will get better, particularly as tools are modified to enhance their reasoning abilities.”

That can make it easier to build flows that are usable by a larger number of engineers. “It has huge value at the front end, where you might want to go from specification to RTL, or SystemC, or looking at the specification of a previous generation design going to a next generation design,” says Rob Knoth, group director for strategy and new ventures at Cadence. “How could I create a derivative product, to satisfy this new market? That is a rich and important industry where LLM-based AI, generative AI, agentic AI, has a tremendous opportunity to show value. That is a swimming pool that it is natively good at, and something that adds a huge amount of value.”

Integrating AI inside the tool also can provide a lot of value in some cases. “Historically, we were using AI externally for one application,” says Siemens’ Gupta. “Then we did some benchmarking to find the advantage that we could get by coupling AI with the solver itself. We were able to get a 2X to 30X performance advantage by coupling it with the solver and having more information coming from the solver directly to the AI, which wouldn’t have been available otherwise through the API.”

The internal argument is often based on the problem you are trying to solve. “The deeper the connection to physical processes, the more tightly integrated the neural network needs to be within the tool,” says Partcl’s Balanaga. “Since physical design tools are solving these difficult 3D optimization problems, it makes sense to integrate our trained models deeply into the optimizer to help us find better solutions.”

The need for abstraction
New models are required to shift left. These need to have the right amount of detail and execution performance. “We can certainly handle these physics and multi-physics problems at the transistor level,” says Gupta. “The industry has been doing this for a long time. It starts with the process design kits (PDKs.) These give you information about how transistors are going to perform, such as the IV curves. Then we use machine learning and reinforcement learning to build numerical models from that data, to do prediction and handle these multi-physics problems.”

Those abstractions are beginning to influence many aspects of the flow. “When it comes to optimization, decision-making is critical, and some sort of model is typically used,” says Benjamin Prautsch, group manager for advanced mixed-signal automation in Fraunhofer IIS’ Engineering of Adaptive Systems Division. “For instance, there is work on new optimization approaches that bypass expensive simulations involving 3D EM solvers. They use a network trained by a reasonable amount of initial — and expensive — simulation runs. Optimization is then significantly accelerated by using those trained design models, thus yielding a better overall result. However, it must be stressed that the investment in the initial simulations and training should have a promising ratio with respect to the savings.”

Over the past couple of years, LLMs have stolen the show. “Directly applying LLMs won’t get you very far,” says Balanaga. “Language models are not well-suited to handle detailed physical information that needs to be encoded into specialized AI models. Using AI for problems close to the physics requires careful handling, as well as injecting domain knowledge into the model design and training.”

EDA-specific domain models are still evolving. “The models are fairly capable today, and there’s a lot of advancements happening when it comes to foundation models and the knowledge that they possess,” says Synopsys’ Thiruvengadam. “The foundation models have certain capabilities, and there is a way to customize these foundation models. That is where, in the context of EDA, they will have to customize the LLMs for their particular domain and context. And that’s how you improve the accuracy of the response for that particular domain.”

That is not as easy as it may sound. “When people talk about physics AI, they are talking about it from a generative and agentic AI perspective,” says Gupta. “Historically, it started with large language models that understood language really well. Now there are large investments going into broadening that from language models to physics models, and being able to be able to handle other domains besides language.”

There has been some success. “Multi-physics means problems that span multiple physical domains and LLMs aren’t well-equipped to handle these today,” says Arvind Srinivasan, product engineering lead at Normal Computing. “Historically, we solved those problems with simulation. Now people are trying to embed physics into AI training. AlphaFold is the gold standard: it combines reinforcement learning, physical constraints, and simulation. That worked because you could quickly reject invalid protein folds. In hardware, running a full simulation takes hours, so you don’t get that quick reward signal. That’s a big problem.”

What are the relevant design and verification models that EDA can use? “AI is enabling a link between language and a mathematical structure,” says Fraunhofer’s Prautsch. “This could be a constraint graph, which is commonly found within EDA. AI will probably act as a converter from requirements into structure, or preliminary parts or constraints of it, and vice versa for verification.”

As existing algorithms need to become more aware of multi-physics issues, there is an opportunity to assess the benefits of AI assistance. “Placement-aware gate resizing is a problem very close to physics, but also exactly the type of problem deep neural networks have shown promise at solving,” says Balanaga. “This model is doing the equivalent of the heuristics that have been hand-tuned in existing tools over decades. Our philosophy mirrors the approach of AlphaGo, which surpassed hand-tuned expert systems by discovering better decision processes. The solution is a combination of synthetic data and reinforcement learning.”

AI-generated models
A cornerstone of semiconductor design uses an abstraction of transistors to provide gate-level models. These models have adequate accuracy with much higher simulation speeds. Shift left requires abstractions that enable informed decision-making at all stages of the development flow, and that requires a plethora of new models. But models are expensive to create and validate. That has many people looking at AI to create the necessary models, which either could be executed by traditional engines or inferred like an AI model.

Analog verification needs fast, but accurate models. “Verification will be supported by accelerated generation of surrogate models as required for system simulation,” says Prautsch. “This is important to enable system-level modeling, especially for AMS, where tedious simulations can be replaced by an AI-based surrogate model.”

EDA always has used data-driven methods, such as simulated annealing. “The real difference with generative AI, is that you can use past simulations as training data for very large models to improve downstream simulations, or tailored LLMs to translate huge specs into physical constraints,” says Normal’s Srinivasan. “But if validating AI-generated RTL or test plans takes just as long as building them yourself, you haven’t saved any time. AI won’t replace areas where reviewing the material it generates doesn’t actually reduce the workload.”

The technique is gaining ground. “You can take a design and train a model on that design,” says Rich Goldman, director, electronics and semiconductors business unit at Ansys, now part of Synopsys. “But you need tons of data. Thousands of data points that form part of a sweep. You take your design, you change it a little bit and run that to get the results, and then change a little more, one more parameter, a little bit, and you sweep through. That is your training data set. Once you train that AI model, and that takes a long time, you apply it to your design and you get back a result almost instantly. All these runs that take weeks and months, you can get instantly. You can apply that to many different designs, many different iterations, and get the best design. That is not going to get you 100% accurate results, but it will get you 90% to 95%.”

Small language models (SLMs) can help here because they can be targeted at specific problems, which reduces cost and the complexity of the models. But scaling down from FP32 to FP16 and below also raises some questions about the accuracy of results.

“You want to distill knowledge from bigger models and inject that into smaller models, but also be able to prune the weights,” said William Wang, CEO of ChipAgents. “So instead of using 16 bits, you may use 8 bits to store the weights. Everything is compressed. There are all these well-known algorithms to compress the weights and make it small, but you also need to get to a certain level of accuracy.”

This is made worse by data scarcity in EDA. “One of the most significant issues associated with AI is obtaining enough training data to get the quality of LLMs to the point where they can be reliable,” says Dave Kelf, CEO for Breker Verification Systems. “We are now seeing the application of SLMs and other techniques to get around the sparse data issue, and it will be this thinking that drives practical AI application. Combining AI expertise with EDA experience seems critical to optimize the use of these new technologies.”

Reinforcement learning can start with smaller data sets. “In order to do reinforcement learning, you need two things,” says Balanaga. “For the placement-aware gate resizing problem, you need a model that makes a prediction — for example, a better set of gate sizes, and an environment such as a timing analysis engine that runs very quickly. The training happens in this loop with the model exploring a bunch of different gate sizes over a large number of synthetic netlists. Admittedly, this is less efficient than having a large number of netlists with optimized gate sizing, but without that data existing, we have to spend compute to generate the data.”

That can affect accuracy. “As you go up additional layers of abstraction, and where natural language becomes the next level of abstraction, having a firm grounding in the underlying computation is incredibly important,” says Cadence’s Knoth. “You can’t go to another layer of abstraction without having faith in the accuracy of what is actually happening underneath you. Otherwise, that abstraction loses accuracy, you become less effective, and the end result is not as compelling as it could be.”

Questions about accuracy are typical with AI models. “People get confused between estimation and correlation,” says Arteris’ Siwinski. “It would be great to have things that fully correlate, but many things change with workload. You do something at the system level and hope it’s going to be a one-to-one mapping to how you implement that. We have built a notion of physical awareness in the architecture-level tradeoffs, where we basically glean enough about placement, congestion, blockages from place-and-route, and wire length calculations while understanding the impact of power. Typically, a front-end team will do something, and the back-end team will say, ‘This doesn’t work.’ We are trying to build enough of the care abouts and the KPIs that you cannot violate early on. It’s an approximation, but it’s a good-enough approximation. The front part of the process is not just creating something that’s fantasy, that is not implementable.”

Trust, but verify. “Nvidia showed that they could use surrogate models to massively accelerate graphics rendering,” says Knoth. “These are similar challenges that we face with sign-off. When you see this image, do you trust the image? That’s all following physics-based principles, and so I don’t think we can rule out the role of AI in the calculations, but it does not displace the importance of first principles. Physics-based algorithms, actually computing these things, will be complementary. Even those surrogate models, where they’re inferring many pixels, they still require the calculation of a seed to do that inference on.”

AI throughout the flow
In the software industry, AI is now expected to create code. “If the tribal knowledge in terms of designer expertise and tool expertise can be documented, then you have the ability to learn from it,” says Thiruvengadam. “That is the power of the LLMs. All we need is this knowledge to be captured in a way that can be ingested into the LLMs. Then the LLM can do a lot of the work. You could ask a model or generative application, ‘What is the right analog architecture or topology for this particular spec?’ The LLM would provide a recommendation for an architecture with an example and a whole bunch of other collateral that can help.”

Other people would like to see AI help with specification and requirements. “The challenge is you have thousands of requirements, and it’s hard to identify logical errors within them,” says Prautsch. “This task is typically language to logic, or language to some mathematical representation, and this is a sweet spot of AI.”

The key is that it must talk to engineers. “When people ask for reasoning, what they usually mean is interpretability — outputs that humans can verify,” says Srinivasan. “LLMs don’t reason like humans. They do transformer decoding. But if you constrain them to output a plan with steps, it looks like reasoning and becomes workable for humans. That is not really reasoning. It is a constraint on how we display outputs so people can interact with them. The point is designing human-AI collaboration, not making machines think like us.”

Rapid Evolution
Given the pace of innovation in all kinds of AI, it is clear that the industry can expect to see rapid improvements both within the EDA tools, and in agentic AI tools that can help to relieve engineers from some of the burdensome tasks involving tool iteration. “New technologies and new capabilities are getting released every day,” says Gupta. “We are constantly evaluating the opportunity to use these new technologies under the hood, to be able to speed things up. EDA is unique. We need to move faster, at the pace that the industry is moving with the foundational technology. We’ll continue to see the pace of AI capabilities coming into our tools keep growing.”

With it, the role of engineers will change. “We expect AI that will enable teams to build a much larger number of chips,” says Balanaga. “We anticipate that the combination of advanced agents and improved tools will dramatically expand design capabilities, and this will shift human effort to meticulously defining architectural specifications upfront. From there, AI will handle most downstream tasks, with humans acting as orchestrators and resolving any ambiguities in designer intent.”

Related Reading
The Future For Formal Verification
Will a formal specification be part of the future, or will we continue to see natural language specifications? Will formal successfully break out of just functional verification?