PartsCube Global

Back to Home
Semiconductor Engineering

Are AI Agents The New Semiconductor IP?

Semiconductor Engineering

Key Takeaways:

  • AI agents could make customized design workflows faster and less expensive to develop.
  • Multi-agent systems introduce new challenges for orchestration, verification, and tool integration.
  • Consumption-based pricing could reshape EDA business models, but trust and proprietary data remain obstacles.

AI agents are emerging as a new category of reusable design capability, but with a twist. Rather than providing tightly defined design blocks, agents can be used to explore different options, integrate them, and use those changes to improve designs.

At the leading edge, soft IP developers have been customizing their code for large customers for more than a decade to boost performance and/or reduce power. But that customization is typically narrow and costly, and optimizations for one design may not directly transfer to derivative chips — especially multi-die assemblies at the leading edge.

Agents, in contrast, can be used to customize different functions with little or no human intervention, and they can self-optimize for different workloads and use cases, and some agents can even adjust for aging hardware. And they can do all of this much faster than a human, even though an experienced engineering team is needed to verify those results and recommendations.

“We have in-house specialized AI agents to do different things, to do spec analysis, RTL generation for design, for formal and functional verification, and then CDC linking, and place-and-route,” said William Wang, CEO of ChipAgents. “We provide our optimized agents to a customer, and sometimes a customer builds their own AI agent. But there are some similarities with the IP business. Sometimes we’re specifically building an agent for one customer for a particular flow, which is similar to what happens in the IP business. What’s different is that we can do it in a very short amount of time. With IP, you need 10 months. Our cycle is much shorter because AI agents are very easy to customize. In some cases we can give solutions to the customer the same day.”

Those solutions still need to be verified and debugged. But the ability to sort through many different options quickly is an attractive formula, especially with rapid evolution of AI.

“I draw parallels to the early 2000s gigahertz wars, where all of a sudden everything was multicore,” said Mo Faisal, CEO of Movellus. “Now we’ve gone from,’ Everything is going to be a GPU,’ to ‘Oh, no, it’s CPUs and GPUs, and this is a one-to-one ratio.’ So effectively we’re going through the same architectural churn that we go through every 10 to 15 years.”

The business of agents

Business models still need to solidify around agents. but they also need to be flexible enough to evolve at the same pace as the technology.

“If you look at how the EDA industry has evolved, we’re not selling perpetual licenses anymore,” said Rob Knoth, senior group director of Strategy & New Ventures at Cadence. “We moved to time-based licenses many years ago. Now we see similar things as agents roll around. They’re hitting up the tool usage even more because they want to find answers quickly, and they’re trying lots of experiments. So putting together more of a consumption-based model is really good because customers know how to use that.”

This isn’t a simple transition, though. While the promise of agents is significant, their adoption in chip design will take time. In addition, security and permissions aren’t always well-defined.

“There are some things where you’d really rather sell an agent than sell the software,” said Wally Rhines, CEO of Silvaco. “People are trying to figure out if anybody in EDA is actually pricing based upon tokens rather than traditional pricing models. It doesn’t seem like anyone is, but their customers are all talking about it. I’m very much open to selling an agent that will automate [manufacturing equipment] calibration, then jump back to my store of synthetic simulated data against the physical data. We don’t get to see the physical data, but we do know how many tokens our agents consumed. It’s a way that EDA can sell a product, get the benefit, and no one ever sees the underlying data that companies want to keep proprietary.”

Divide and conquer, but with caveats

One of the big changes with the shift from generative AI to agentic AI is the need for orchestration, and this has broad business implications. Orchestration is basically a blueprint that defines how tasks are assigned, how agents access tools and data, and how that work is coordinated. But it’s especially challenging if multiple agents execute at the same time and their actions overlap.

“What’s different about traditional AI and agentic AI is the coordination of multiple agents that are working in parallel, but doing different types of work,” said Robbie Williamson, vice president of customer engineering for cloud AI at Arm. “It’s not like there’s a unique workload that we haven’t seen before. It’s more about how do you coordinate all of these at the CPU level, and have enough bandwidth, memory, etc., to do all of this at scale. Instead of an individual making a request and driving AI, we have computers that can do it much faster. The transition that we see when we talk to hyperscalers is now about what they need from the CPU. With training, that wasn’t as important because you’re just batching the job and waiting for it to come back.”

That creates a new hierarchy in the design flow. “What has happened with this multi-agent workflow is that, in an organization, you’re trying to design a chip,” said Prith Banerjee, senior vice president of innovation at Synopsys. “When they’re having a problem with the PPA, some engineers say they know exactly what changes to make to lower the power. They’ll do clock gating. But when you clock gate power, then it changes something else. It increases the area. So some area will come in — and this is exactly what happens in the real world with coordination when multiple people come together to design a system. The human manager is coordinating the activities of the human agents in the world of a multi-agent framework. We have agents for EDA tools and agents for the CAE (computer-aided engineering) tools. We have simulation with three core solvers — CFD (computational fluid dynamics), HFSS (high-frequency structure simulation) for the electromagnetics, and mechanical for the structural part. In the EDA world, we have things for the implementation, and things for sign-off. Each of these tools has sub-areas, and we have agents to do these things. But what was happening by a team of engineers coordinated by a human manager is now being done by a hierarchy of agents with that framework.”

That changes some of the dynamics in chip design in not-so-subtle ways. Soft IP is often portable across process technologies, but it typically requires optimization and validation. AI agents can automate those tasks by coordinating EDA tools and weighing alternatives, and they can work with traditional IP to optimize specific functions in the context of a larger design.

“One of the key implications of agentic workflows is disaggregation,” said Niels Faché, senior vice president for Keysight‘s Design Engineering Software business. “We believe in platforms. Now we have to see how everybody’s going to deal with that because a platform is a very sticky concept. Many of our customers invest in customizing our platform for their needs. It’s very hard to move away from it. It’s the same for Cadence and Synopsys, but we believe that ultimately customers want that flexibility. We take an open approach to this, where we accept that disaggregation is going to happen. We need to play to the strength of our simulation capability and our understanding of the application. Customers still wanting to work with us on creating their workflows will need help with orchestration, so we need to be experts in that domain, while the most sophisticated customers will do it themselves.”

That requires some fundamental changes for EDA companies. “With AI, that’s a different model where human engineers work with digital engineers or agents,” Faché said. “So now we have to redesign our products. You’ve heard about MCP (Model Context Protocol) servers where we look at a product and ask, ‘Okay, how can we modularize it to provide access to different capabilities with the proper APIs?’ All of that coding is typically done in Python. That would then allow an agent to stitch together a code-based workflow. Our job is to make sure that our customers get this MCP functionality. We also provide examples so that an agent can learn from them and say, ‘Okay, based on these examples, I understand how this works. And now I’ve got a request for a new design, and I can organize that work based on what I’ve learned.’ That’s the world we’re moving toward, and we already see that with our front-runner customers. They are already building these agentic workflows. That means we need MCP servers, and we need all the proper documentation so we can train these agents.”

What else can agents do?

How much freedom to give agents isn’t entirely clear. AI is opaque, and unlike traditional IP, it needs to be observed from the outside. This becomes more complicated as more agents are assigned to different but connected jobs.

“Orchestration is going to be a big challenge, and it will become an even bigger challenge,” said Stelios Diamantidis, chief product officer at Cognichip. “As agents become smarter, they can deliver some outcomes, and it’s been proven to work in many other spaces, as well. But putting all these agents together to drive them as a cohort toward an answer to a problem can also have some interesting challenges. We’re still working with models that generally do not understand the deep physics of the chip design problem, and that means a lot of the logic being baked into agents. But what happens when those agents begin to disagree? Are we going to have ping-pong effects and convergence issues in EDA? We’re no strangers to those effects.”

This is particularly true with multi-die assemblies. “The stack is very heterogeneous,” Diamantidis said. “You’re looking at a model that was trained with sources that generally we don’t know very much about. You have agents that have been given enough logic about a problem by enough different people that the combination of agents and models is now a fairly chaotic model. You put them all together and you get chaos times chaos.”

These challenges help explain why IP and EDA companies are working to tightly integrate agents into existing tools.

“When you look at how we’re applying agents into our data center design and operation tooling, agentic AI coupled with physics-based simulation is great at being able to integrate large amounts of disparate data, ask questions, run experiments, evaluate the output of those experiments, and then do the loop again,” said Cadence’s Knoth. “And what’s very unique about the space we operate in is that, as opposed to other industries, we don’t need a person to look at the results and say that’s right or that’s wrong. EDA has the tools that we’ve used for decades to grade the decisions. ‘That floor plan worked better. It lowered the power.’ Or, ‘That floor plan was worse. The power went up, the heating was worse.'”

Establishing a hierarchy for agents is important, but it’s not always straightforward. “So you have one agent with different sub-agents to do different things, and then we have a multi-agent solution,” said Wang. “We released a multi-agent solution last year. So for debugging, your SoC is so big that it makes it hard to optimize performance. You need multiple agents to have parallel access to different contexts to do the job. Similarly, with formal, you have multiple agents looking at doing the debugging and being able to optimize the design. Those are extremely critical. So you need multi-agent coordination. In the future, we’re going to have virtual chip companies with virtual AI engineers working together, collaborating from this agent organization, and then maybe a part of the chip companies will be virtual in the future.”

That, in turn, could accelerate the adoption of agents.

“There could be a time where we start looking at agentic workloads and try to figure out how we can model those in pre-silicon to improve IP around that, and what that means in terms of the design of the CPU itself and our CSS offerings for those building their own,” said Arm’s Williamson.

Trust and verification

The bigger question is how much autonomy chip designers will be willing to give agents.

“The cost of a mistake is $500 million at 3 or 2 nanometers,” said Ankur Gupta, executive vice president of Siemens EDA IC products. “That’s why these engines have to remain deterministic. They have to remain high-fidelity. We, of course, use reinforcement learning, unsupervised learning, and we’ve been doing that for more than a decade. But the way we measure the result of that learning capability is whether we are still delivering a trusted outcome. So that doesn’t change.”

There’s more to the trust equation than just the outcome, though. It harkens back at least a few decades to the challenge of sharing process data that chipmakers and chip companies consider to be their IP.

“All of these agent technologies and models are pretty good at understanding the language components,” said Gupta. “But the physics component is still the big challenge. And if you think about SLMs, LLMs, where are they going to get the vast amounts of data from, because the whole universe of semiconductor and electronics designs is very siloed and customers are concerned about their IP. Everybody wants on-prem, but that argument persists because you’ve got the foundry, which is not giving up anything.”

Foundries do share PDKs with end customers and EDA companies, but they tightly restrict access to them, which makes it more difficult to train AI and to share models. And because AI is opaque, it becomes even more difficult to convince people to share data.

Nevertheless, the adoption of agents continues to grow. “The analogy is to what the human engineer is doing in a company,” said Synopsys’ Banerjee. “You do not have a single engineer doing the entire design of a chip or a system. You have teams of engineers, and teams of engineers are managed by managers, by directors, by vice presidents. And you have that hierarchy because individuals are doing different things, which they are skilled at. Then the manager coordinates their work. So the early work in AI applied to EDA was to make the tools go faster, better PPA. This is what we did with our DSO.ai (design space optimization). The early work on AI was GenAI. It was generating a script. So you ask a question. ‘I am struggling with this Fusion Compiler. What setting should I use?’ And based on hundreds of questions and answers, this thing essentially is a ChatGPT targeted to this particular line of questions. So it’s an agent for RTL assistance, or whatever.”

AI agents can even be used to fine-tune existing steps in the design flow. “In general, we want to have some very commonly used workflow agents,” said ChipAgents’ Wang. “For example, when you’re building an accelerator chip, it’s very different from a CPU. So how do you do verification? Do you do more formal, less formal? It’s a very different methodology. Having that flexibility is really critical. You don’t want to build a workflow that is very rigid so people cannot customize it. The platform needs to allow you to customize for your specific chip while also being general enough.”

Conclusion

“The model brings a lot of intelligence, so the agent doesn’t have to be terribly smart,” said Cognichip’s Diamantidis. “It can be there to introduce some guardrails and add some additional skills. But right now we’re flipping it. We’re saying, ‘You know, we don’t have billions of examples of hardware, and they’re spread out in all these different silos. So we can’t get them in one place. We can train a model that’s very focused on what it does and put everything on the agent, but that makes it top-heavy, and top-heavy things tend to lose their balance. That’s one of the big challenges.”

So even though AI agents are being integrated into traditional workflows, chip design is still cautious about when and where to deploy it.

“Should agents be managing the hardware that is running all kinds of agents? The answer is yes,” said Movellus’ Faisal. “Is that going to be easy? I don’t think so, because the silicon industry that is enabling the intelligence and adopting all of that for internal tooling and even management is probably the slowest and most conservative. A $500 million silicon program failure is a big deal. There are going to be all kinds of management agents that will be needed — managing the rack, managing thermals, managing workload orchestration.”

While agentic AI won’t replace traditional IP, it could impact the design process in which that IP is used. Change is coming, but there are still a lot of challenges and unknowns ahead.

The post Are AI Agents The New Semiconductor IP? appeared first on Semiconductor Engineering.