The Impact Of AI Automation On Chip Design

Key Takeaways:

• Tools are currently built with a defined notion of how they will be used. For agentic solutions they will be transformed into collections of callable engines.
• Agents may be built by EDA companies for design houses or larger design houses may build their own, potentially transforming the EDA business model.
• The role of engineers will change, but exactly how is not certain.

Agentic solutions are more than a wrapper on top of existing tools. They will transform tools into sets of callable engines, deployed by AI and managed by engineers. This will impact the tools, create new flows, and transform the role of engineers.

No discussion these days is complete without talking about the impact AI will have, whether the topic is EDA tools, methodologies, or the team players. For all of them, the goal is to find out how to adopt AI, and how to ensure the participants aren’t left behind. The race to adoption creates both opportunity and uncertainty.

Some of the claims, perhaps not fully realistic today, would eliminate the entire EDA industry. “With AI there’s a lot of talk about taking a specification and having it create a test plan for you, or taking a specification and have it create code for you,” says Abhi Kolpekwar, senior vice-president and general manager at Siemens EDA. “‘Give me the spec and I can create formal properties for you.’ They claim this aspect of reading a specification and converting it into some tangible artifact is something that AI is well suited to solve.”

The likelihood of that being adopted for any realistic semiconductor device is slim to none today. The risk would be high, and the cost of failure means that everyone will continue to rely on the existing methodology to some extent. “Maybe you have a sub-agent doing spec analysis and deriving design decomposition on one hand, and let’s say a test plan on the other,” says Ramesh Narayanaswamy, member of technical staff at Synopsys. “The user wants to review it, but they may ask for an alternate model and have that model critique the first. When you are closer to the spec, the human needs to be in the loop, at least for now, because the spec is not easy to cross-check. If you misinterpreted a paragraph, you misinterpreted a paragraph. But you can use some cross-checks on the verification side, keeping this independence. You hope that the same mistake is not in two chains of work, which is somewhat true even in human interpretation. Those subtle miscommunication-based bugs are nasty and hard to find, but you could minimize some of it by using models as reviewers and effectively give humans more bandwidth to review.”

That does not mean they cannot look at ways to optimize the design or flow. “My management told me I need to apply AI to my problem,” is something Alexander Petr, senior director at Keysight EDA, hears a lot when talking to his customers. “I always say that’s the wrong way around. You should start by asking, ‘What problem am I trying to solve?’ Then, look for an AI technology that can be used to help you overcome that. AI is the next evolution of automation. We’ve been connecting our tools. Every company has a CAD enablement team, and they have been building workflows by connecting tools. They have been creating glue code to make them more efficient and effective. The question now becomes, ‘How can I turn this into an optimization problem? Where should that optimization loop sit? Where do we want human interaction? Where do we want sign-off by people? What do we want to be done by machines?’ Once you can describe your flow and the loop you’re trying to build, then we can talk about the challenges and where AI can help.”

To reap maximum benefits from AI, other changes may be required. “One of the clearest lessons is that agentic verification works best in structured and traceable engineering environments,” says Andy Nightingale, vice-president of product management and marketing at Arteris. “AI systems perform far better when specifications, coverage goals, architectural intent, and verification data are machine-consumable and well-connected.”

The specification is often not well crafted and changes during the development flow. “If you give a spec to a tool or chat-bot, I don’t care how great that chat-bot is. If the spec isn’t good, you’ve got a problem,” says Paul Graykowski, director of product marketing at Cadence. “What is really central is the mental model, and that is where an engineer working with it can help when the tool says, ‘This is an issue.’ The engineer can work with the tool and explain, and then the tool answers that. Or, you as the engineer can go and look at this and go explain to me how this works.”

Traditionally, the semiconductor industry has been slow to change. “The future is interesting, and we are brainstorming internally about how to change the future of EDA as the industry is changing,” says Hamid Shojaei, distinguished engineer at Cadence. “But I don’t think it will happen soon. People don’t trust a dashboard that shows 100% coverage with all tests passing. They need a certain level of trust before they will do a tape-out. It takes a lot of time for people to change their mindset. But people need to be flexible, and a lot of things can change in the future.”

Tool evolution
The core tools are unlikely to be replaced. “Some people claim they no longer need tools. ‘My agent will tell me if this design is accurate or not,'” says Siemens’ Kolpekwar. “The industry EDA tools have been there for a long time. They are highly computationally intensive. They have been certified. They have been tried and tested. If somebody says I’m going to ask my LLM if this design is correct, using a probabilistic model, I have a hard time believing in that. Is there anyone who wants to tape out a chip using just an LLM?”

While core tools, such as simulation, may remain for the foreseeable future, they may evolve to better fit into agentic flows. “Verification products will increasingly evolve toward contextual, workflow-aware platforms that expose intent, metadata, traceability, and debugging knowledge in machine-consumable form,” says Arteris’ Nightingale. “Future tools will need to connect requirements, implementation, verification evidence, and system behavior across the lifecycle. The realistic value of agentic verification is not autonomous design signoff, but amplification of strong engineering methodology and improved leverage for increasingly complex system verification problems.”

Today’s tools attempt to perform a complete operation, and the required level of granularity may change. “Flows require the ability to bring together fit-for-purpose computational engines to serve what the customers require,” says Kolpekwar. “Let’s say you’re creating a design and you have a number of concerns, such as functional verification, safety compliance, and security compliance. You may also want early performance estimation, and you need a verification solution that can invoke these engines on the fly, on a fit-to-purpose basis, and then exchange the data that they create with each other to provide the end result. I really believe that engines will come together and basically interact with each other to give you what you want.”

The tools already are changing. “We are trying to create more structured information that models can interpret,” says Synopsys’ Narayanaswamy. “There’s also an opportunity to infuse the model orchestration into individual phases of the tool. Many of our tools have a number of heuristic algorithms and we orchestrate sequences of optimizations. You could actually have that orchestration happen in ‘self-learning agents’. So long as there’s an objective metric, these things can learn.”

The EDA industry relies on selling tools, so the future health of the industry is directly related to the quantity of tools they sell and the value they provide. “If the EDA industry goes toward the development of agents, agents work like a virtual engineer for semiconductor companies,” says Cadence’s Shojaei. “Instead of paying that money to employees, and bringing more and more employees to do their chip design stuff, they can pay that to virtual engineers, which are coming from the EDA industry. The business model will change in the future, and the technology will bring new ideas, but definitely there is a huge opportunity for the EDA industry.”

There are several ways that this could work. “There is a broad spectrum of people who are applying analytical AI,” says Kolpekwar. “First are the people who have the money. They have deep pockets, and they have created a set of AI engineers or data scientists who are writing their own frameworks, writing their own agents. All they need is engine access. These are highly sophisticated users, and they can be divided into two categories. Category one is, ‘We know what we are doing. Just give us the tools and we will figure it out.’ Category two is, ‘We tried this. It didn’t work for us, and we think that EDA vendors can write better agents.’ Then there is a third category of customers. They have defined a strategy and they want EDA companies to come in, create the framework, drop in the tools, and create the agents. This can be a turnkey project.”

One of the big problems has always been access to the data required to train AI. “The semiconductor industry has been successful in the last 50 years in terms of IP protection,” says Keysight’s Petr. “No one is going to give away their IP just to enable a flow. That makes the discussion a little bit more difficult. If you can’t scrape knowledge off the Internet, that means you need to inject it somehow. The person who has it is the one who needs to inject it, and that makes it a distributed problem. There is knowledge that the EDA vendors have, there’s knowledge the foundries have, and there’s knowledge that the design houses have.”

That implies partnership and cooperation. “It’s very important for EDA vendors to recognize that they are partners, not the owners of a flow,” says Kolpekwar. “Partners means providing that contextual intelligence so that we are able, along with our end user, to manage the change better. This is one area that will evolve, and it will take some time to mature. But I really believe that we have the ability to create a context and contextual intelligence that not only understands the changes, but also makes more informed decisions, considering the history and a possible future in the given flow context.”

Engineer evolution
Based on early indications, the role of the engineer could increasingly become that of a supervisor and reviewer rather than a doer. “In the software space, there would be 1 project manager for every 10 developers,” says Narayanaswamy. “Today, engineers need to be thinking more about product, and organizing and architecting the product, while the details of the building, with guardrails, can be done by agents. They need to upscale to being this architect and product manager combo. That’s where it’s going to go. I suspect it will happen at a somewhat slower pace in EDA, because chip design has a higher bar of robustness, and is justifiably more risk-averse. That will make it a bit slower, but it’s likely to go there.”

It is clear that their role will change. “Verification teams will likely shift from manual processing of data toward managing intent, strategy, and risk,” says Nightingale. “Engineers may spend less time correlating logs and debugging repetitive failures, and more time defining meaningful scenarios, validating AI-generated results, and understanding system-level behavior. The most valuable engineers will increasingly be those who understand what needs to be verified and why, rather than simply generating verification content.”

Others agree. “Verification teams will evolve toward higher-level orchestration roles, and verification tools themselves become agentic platforms that coordinate across the entire design and validation stack rather than serving as isolated point solutions,” says William Wang, CEO at ChipAgents.

Over time, that could transform their job function. “If we are successful in implementing agentic verification to a reasonable level, then all our verification engineers will start to become verification scientists,” says Kolpeckwar. “A lot of their time will be spent analyzing what happens, providing higher levels of engineering judgment, doing some experimentation, implementing some hypotheses, and then driving a strategy on how they basically instill confidence that we can tape out this design.”

It is always easier to think about positive outcomes, but other side effects may appear. “Unlike previous advancements in verification, AI is not just a linear extension of existing techniques,” says Stefan Birman, partner at AMIQ Consulting. “There are many variables and unknown unknowns, and we all have a great deal to learn. For example, they are finding that humans need to check and validate all AI-generated output. This is generally less interesting than doing the work themselves, so engineers worry about their jobs becoming less creative, or even disappearing someday.”

Responsibilities may shift between team members. “In the past, designers have been encouraged to do more verification before handing it over to the DV team,” says Shojaei. “It was not very successful. But a unit testing agent will enable designers to do that verification. They don’t need to know how to write assertions anymore. An agent can do that for them. They don’t need to know how to run formal, or write a tcl file, or build a file to run the tool. AI is perfect at generating those EDA setup files. In minutes they have a very nice testbench in formal or SystemVerilog, and an agent that can run it. The agents even help them to debug it and root cause issues. They don’t need to close coverage 100%. They need to find the easy bugs so that the verification team does not spend a lot of time on them.”

It is unclear if agents encapsulate all of the knowledge that they need. “You have to build solutions that look over the shoulder of the engineers and learn from them,” says Petr. “Engineers never wrote down their knowledge. To train a young engineer, you normally pair them with a senior engineer. That concept can also be applied to AI. It’s a supervised learning approach and can be built into a flow. You can try to do this unsupervised, which is basically reinforcement learning, but the challenge becomes that you need to define the goals so the system can start running to try to hit the goal using a fully automated optimization loop. The difficulty here is, ‘Who is able to describe the goals perfectly so that the system doesn’t come up with some nonsense?'”

No matter what level of engineer, there is one key message: “Engineers will have to be able to adapt, to leverage AI to be very efficient, so that they’re not the one being replaced by the AI,” says Graykowski.

Related Articles
Verification Methodologies Struggle To Keep Up With AI
Engineers are flooded with new capabilities. The problem now is how best to deploy them.

Toward Agentic Verification
Using Agents for verifying designs holds huge potential, but can it deliver? And what comes next?

Creating Agentic EDA Methodologies
Current approaches involve multiple tools, vendors, designs, data formats, and abstractions. Can agents really use them all?