It depends on what those models are used, which also can have a big impact on the cost.
Key takeaways
A model captures aspects of a development flow, but EDA flows have been constrained by the cost of creating, verifying, and maintaining these models. That may change with the rollout of AI, which could dramatically reduce the cost, and thus have a significant impact on tools and methodologies.
Models capture essential information while hiding unnecessary details. They are at the core of EDA flows, and many have been standardized to enable portability between tools and between vendors. Some are ubiquitous, such as an RTL model of the design itself, or a collection of models that form a testbench. Others may exist for only one task, at one point in the flow, such as an architectural performance model.
While some believe it will become possible to jump from specification to tape-out in one step, the likelihood of that, at least in the near future, is slim. No model, including the specification, can contain all the details of a design and have sufficient execution performance for analysis. That is why abstractions, simplifications, or estimations need to be made. In many cases details emerge through the design process, and these details need to be fed back, in some form, to an earlier model for additional analysis.
Traditionally, models are expensive because they have to be created, maintained, and verified for accuracy. Without a high level of confidence in the model, wrong decisions may be made, and that can be expensive. Promising new flows have been created only to be abandoned because the cost of creating the necessary models was too high. One such example was the electronic system-level abstraction that promised to enable early architectural exploration and performance analysis.
Today, there is hope that AI may come to the rescue because it may be able to create the necessary models for a fraction of the cost. New flows and new tooling can quickly be put in place to solve highly specific concerns, which in some cases may be unique to a specific customer or design style.
“Much of the process in silicon engineering, as with many other physical processes, is trying to approximate hugely computationally complex physical processes, mostly by simulations and approximate models,” says Arvind Srinivasan, product engineering lead for Normal Computing. “If you look at the trajectory of AI, and especially reinforcement learning across some of these domains, we have gotten to better learned simulations of something, which gives a better starting point. A behavioral model is a proxy for some final outcome, so you do not have to simulate fully. Eventually, you want either a more accurate or more optimal representation of that physical process.”
Some companies are considering having a range of models, with differing levels of detail. “We are already seeing this start to happen,” says Ramesh Narayanaswamy, member of technical staff at Synopsys. “For example, a company may want to use the expensive model for planning. For implementation, they may use the intermediate model. And if I’m just doing data analysis, I can use a much faster and cheaper model. The agent and sub-agent should be aligned for picking a model that’s appropriate and not blow the budget on the most expensive model.”
An abstract model can often execute an order of magnitude or more faster. “Beyond that, speed has not been that important,” says Tom Demuer, research and development fellow at Keysight EDA. “But that will change as soon as we have a more industrial way to create these models. Then you can start to balance them more. You could say, ‘I want to have a more accurate model, but at a certain cost.’ And then you can pick and choose, but it does require that you have a very well-established pipeline for creating these models.”
That will enable highly specialized models. “The industry is going toward incorporating their own domain-specific knowledge, so that it is much more robust in terms of giving answers,” says Sathishkumar Balasubramanian, head of products at Siemens EDA. “We are seeing it in very specific areas, very domain-specific areas. For example, the back end is behind closed doors. People don’t share the data. That’s the area where customers are trying to fine-tune so that they can use it internally on their own designs.”
The model creation process remains in the early phases of development. “One question is whether the right approach is to train a surrogate model, or to use AI to directly steer the real thing, where it learns its own internal representations,” says Thomas Ahle, head of machine learning at Normal Computing. “In weather prediction, they do not first train a model to act like a weather system and then train another model to predict using it. They try to capture it end-to-end. You want to try and train things end-to-end rather than first creating the intermediate representation.”
The first signs of emerging technology and models are in the analog mixed-signal space. “The adoption of mixed signal, which is a top-down behavioral model-based flow, had a lot of headwinds because people didn’t have good modelers,” says Abhi Kolpekwar, senior vice-president and general manager at Siemens EDA. “They didn’t know how to write good-quality models. With AI coming in the picture, the user entry barrier of models may get reduced. It will become easier for people to take a significant portion of their costly analog simulation and convert that into digital, or a higher level of abstracted model, to speed up the simulator.”
It is not clear if allowing AI to create the models without guidance is advisable. “If you have a filter design, I would first look at physical properties to see if my model — the modeling technique — is sound to begin with,” says Keysight’s Demuer. “A counter-example would be if you naively take a neural network approach and start modeling a filter. It might visually look very good, but you may traverse above the passivity violation criteria. If you use knowledge of electronics and how the systems respond, that would never happen. It would be a better model. So those would be evaluation criteria on which I would pick one class of modeling techniques, and then within that class you can define arbitrary measurements on how accurate you want it. That will be a tradeoff between how much time you want to spend on training versus using it.”
Analog traditionally relies on very accurate simulation. “Full transistor-level or physical simulations, including electromagnetic or thermal, can take weeks or months for complex designs,” says Mehir Arora, head of engineering at ChipAgents. “Behavioral models replicate the input-output behavior of a block without simulating every internal detail, reducing runtime from weeks to hours or minutes. Top-level simulation assembles many blocks together, and without behavioral models, each block must be simulated in full detail — making top-level runs impractically slow. AI-generated behavioral models act as fast stand-ins for verified sub-blocks. Analog/mixed-signal (AMS) blocks like PLLs, ADCs, DACs, and LDOs are notoriously difficult to model behaviorally. Their behavior is nonlinear, process-sensitive, and context-dependent. AI agents negotiate these tradeoffs to produce high-fidelity behavioral models, saving experts hours of work.”
The gains can be significant. “We are trying to accelerate simulations in the analog domain,” says Alexander Petr, senior director at Keysight EDA. “You keep seeing papers that talk about 100X, 1,000X faster simulations. What’s basically happening is we’re trying to use AI technologies, like model order reduction surrogates, to replace some of those time-intensive simulations. Or, we do it the old-fashioned way by simplifying the problem. You’re onboarding accuracy limitations to accelerate, and then you can optimize the problem. When you do a final simulation, that turns into finer verification.”
But the models do not get created without cost. “The AI models would be something where it’s pre-trained,” says Demuer. “You are front-loading some simulations to build a model. The most usage would be where you are constrained on time to do the simulations, such as during the design phase or during an optimization phase. You already have a baseline design, and you want to optimize it, and for that you would use an AI-generated model instead of the full simulation. That gives you a quicker turnaround time.”
These days, it seems as if everything is becoming analog. “As designs push toward atomic-scale complexity, existing EDA tool chains do not do a good job of atomic-level physical design,” says Normal’s Srinivasan. “At the atomic level, the permutation space is too large, and you have to do something closer to probabilistic because you have to learn very good approximations of the physics.”
Similar optimization techniques are being deployed to create surrogate models at the back end of digital designs, as well. “In reinforcement learning-based tools, such as DSO.ai and Cerebrus, the model learns by running thousands of synthesis and implementation trials on real designs, accumulating knowledge of which tool configurations produce better PPA outcomes,” says Simon Davidmann, AI and EDA researcher at the University of Southampton. “The starting point for model training is the design flow itself: the tool generates its own training data by executing the EDA tools it is learning to optimize. This gives RL-based systems a structural advantage, since as design volume through the tool grows, so does the accumulated model knowledge. Vendors with large installed customer bases have a compounding data advantage that newer entrants cannot easily replicate, and every tape-out in their user base strengthens the model further.”
In the digital domain, there are hopes that AI can take a design from specification to implementation without detailed hardware knowledge. “Functional correctness and implementation correctness are not the same property,” continues Davidmann. “RTL that is functionally correct, but written without awareness of timing closure, reset strategy, or clock domain discipline, can pass simulation and still fail to close timing in the back-end flow. RTL that simulates correctly but closes timing at 60% of the target frequency is not a starting point. It is a rewrite. Current LLMs mostly produce functional correctness. Physical-aware coding discipline is systematically under-represented in available training data.”
There are large holes in training data. “For DRAM verification, the failure modes you care about, such as timing violations, protocol corner cases, and conditional interactions, are systematically underrepresented in typical simulation traces because most simulation time is spent on nominal scenarios,” says Hanna Yip, product manager at Normal Computing. “A model trained on that data will be confidently accurate on common cases and silently wrong on the cases that actually cause re-spins. Deliberate sampling strategy in the data is as important as the fitting method, and it requires domain knowledge to get right.”
In addition, specification writing is not precise, which creates ambiguity. “The specification from which you are building the model might contain clauses that are underspecified or open to interpretation,” Yip says. “A behavioral model trained on simulations from one vendor’s interpretation will confidently propagate that interpretation, which may be wrong for a different vendor’s silicon. More training data will not solve this. It requires the model builder to identify ambiguities and make explicit choices about how to handle them.”
Another place where AI is expected to have a significant impact is in the verification flow. “One area where AI is showing some promise is the creation of models to describe a verification plan derived, in turn, from a specification or scenario description all using AI,” says Dave Kelf, CEO for Breker Verification Systems. “These models may be implemented in an abstract language, which then can be fed into a synthesis flow to create the actual verification tests. By using an abstract model, the AI process is broken up into manageable phases while also leveraging existing EDA technologies to insert the ‘tribal knowledge’ so important in these next-generation flows.”
There are other AI influences on verification. “People are starting to create test benches using AI, creating assertions,” says Siemens’ Balasubramanian. “They’re really getting into the design intent and design verification portion of things.”
The portable stimulus language (PSS) was defined to be a model that represents the verification problem in an abstract manner. “In order to get the benefits, you had to learn a new language,” says Shelly Henry, founder and CEO of Moores Lab AI. “AI can now do that. You just use AI to create the PSS, then stitch them together at the top level, and use other tools to get the realization of the test cases. We are taking away the barrier of learning a new language, and they can get the end result very quickly.”
But will AI really narrow the verification gap? “The non-negotiable requirement for production deployment of AI-generated design content is verification in the loop,” says Davidmann. “AI generation without a deterministic verification step is not a production flow. It is a research prototype. Every AI-generated artefact, whether RTL module, UVM testbench, SVA, or coverage model, must pass through a deterministic check before it enters the design database — a simulator, an FPV engine, a lint tool, or a combination of all three.”
Regardless of whether a model is created by a human or AI, it needs to be maintained. Part of that is knowing when the model is out of date and must be refreshed. “A model trained on last year’s process data starts degrading in relevance as the PDK updates,” says Davidmann. “The cost of retraining, the schedule on which it occurs, and the status of designs built with the previous model version are not addressed in current vendor documentation. Teams evaluating AI tool adoption should account for this cost explicitly, since it is an ongoing engineering obligation, not a one-time procurement. Domain-adapted models deliver meaningfully better results for hardware tasks, but the training runs on GPU clusters, the engineering time to curate the data, and the cost of keeping the model current add up to a serious investment that only makes commercial sense above a certain design volume.”
In many cases, AI cannot create the model on its own. “Building up the model can be both time-consuming in terms of simulation time, but also in engineering time,” says Keysight’s Demuer. “If you use an almost blind technique, like neural networks, you typically need a lot of data. What you’re then exchanging is engineering time for simulation time. Do you want a modeling engineer to spend one week thinking through the best modeling architecture, or are you are happy to say, ‘I don’t want to spend that time. I’ll just take a neural network and brute force my way through it with 100,000 simulations.’ I’m pretty sure that if you have a more traditional engineering approach, you might find you can get away with just doing 5,000 simulations. But then you have spent a week of an engineering or modeling team’s time thinking through that. How much do you want to simulate? How much modeling effort do you want to do? And then, how much will you use it?”
But is the creation of a model fundamentally the wrong approach for an optimization loop? “There is no mathematical or physical law that says you can always reduce a system to a simple set of parameters,” says Normal’s Ahle. “You would expect a general system to be impossible to describe more compactly than giving the full thing. It is a debate in the field about whether surrogate models and model control are the right way to go, or whether you should just try and make a good enough black box model that can understand the full thing. End-to-end training might be the solution in the analog design space. The bitter lesson.”
While we have to consider the time and compute costs, we also must consider the carbon footprint. “For an industry that is under pressure to reduce the power consumption of its products, the carbon footprint of AI-assisted design is a question that has received little systematic attention,” says Davidmann. “One important consideration is whether the energy cost of AI-assisted design is justified by the efficiency gains in the chips produced. If AI-optimized designs consume meaningfully less power in the field, the net energy balance over the lifetime of deployed devices may favor AI-assisted flows even accounting for training costs. However, this calculation has not been performed rigorously, and there is still a lot of work to be done on quantifying both sides of the energy equation before the industry can make an informed assessment.”
[Editor’s Note: A second part of this story will delve into more detail about how models are created and verified. It will also consider the limitations of those models.]
 |
 |
 |
 |
 |  |
 |  |
 |
Leave a Reply