Specialization and orchestration are becoming more important as the role of AI agents in chip design widens, but coordination, control, and trust remain challenging
Key Takeaways:
AI agents are becoming smarter, more capable, and increasingly optimized for specific tasks, creating new opportunities to use smaller language models instead of relying on large language models for every task.
This shift has broad implications for the semiconductor industry, from the actual design and manufacturing of chips, to how those chips function under different workloads. While LLMs can deploy swarms of agents for continuous reasoning and independent action, that’s overkill for most applications.
But reducing the size of the model, and potentially the number or capability of the agents, adds its own challenges. One approach is to divide complex workflows among smaller language models and agents, while still retaining the compute capability needed to design, optimize, floor-plan, verify, and debug chips, and to identify anomalies and variation in manufacturing processes that might limit yield. This approach combines both human expertise with AI’s ability to crunch massive amounts of data.
“In our early work on agents, which we created for each of our tools, we have questions and answers,” said Prith Banerjee, senior vice president of innovation at Synopsys. “We had an agent for this, and an agent for that. But in this world of multi-agent workflows, our customers are trying to design a chip. Maybe they’re having a problem with the PPA, and some engineers say they know exactly what changes are needed to lower the power. So they’ll do clock gating, but when you do clock gating it increases the area. Then, someone who is the area expert will come in. This is exactly what happens in the real world where multiple people come together and they design a system with coordination. The human manager is coordinating the activities of the human agents in a multi-agent framework”
The next step is how to automate and combine more of those processes, and that requires more guardrails to limit what AI can do.
“The model is drawing a boundary,” said Mo Faisal, CEO of Movellus. “It’s like, ‘Hey, you’re getting in the domain that I work in. Stay in your lane.’ We will specialize. A couple of [huge] companies will take the LLMs, but everybody else has to specialize. And AI, by the way, is not just LLMs. That’s the biggest use, but it’s not the definition of AI. AI intelligence can be used in many areas. People are building AI for chips, but there’s also AI for everything — accounting, lawyers, weather prediction, and all of those are different and unique workloads. But you don’t need to use LLMs for weather prediction. That’s inefficient.”
The good and bad of specialization is that agents need to be configured and integrated into existing tools and workflows. Making all of that work seamlessly is hard work. It requires orchestration by other agents, or teams of agents, that have been trained to fit together these potentially overlapping steps. And it requires a whole different way of approaching the problem, where CPU-based orchestration may be as important, if not more important, than GPU-based multiply-accumulate operations.
“We’ve gone to the aggregation level with ‘super agents,’ and we think their complexity and the amount of value they can deliver is as big or bigger than EDA tools,” said Matt Graham, senior group director of verification software product management at Cadence. “These are a new category of tools, and we’re starting to talk less about agents than skills that are dynamically assembled into agents. But the idea is that it’s all handled by a kind of orchestrating super-agent. And just like our customers have always arranged our tools into flows with their own custom scripting, they’re going to arrange our super agents and agentic workflows and solutions the way they need by writing their own scripts and putting them into a flow.”
Those super agents (or lead agents, managing agents, or just agents — this is so new that naming conventions are constantly changing), as well as sub-agents that are managed by those super-agents, and they can be used in tandem or managed separately.
“You can have sub-agents do different things, and then for other things we have a multi-agent solution,” said William Wang, CEO of ChipAgents. “We released a multi-agent solution to the industry in Q4 of last year, and it’s multi-agent coordination for things like debugging. If your SoC is so big, how are you going to optimize the performance? To do that, you need multiple agents with parallel access to different contexts to do the job. Similarly, with functional/formal, you can have multiple agents looking at different color groups to do the debugging and be able to optimize. Those are extremely critical.”
The integration problem
Chip design and manufacturing require repeatability and predictability, but many of the leading-edge chips are unique. This is particularly true for multi-die assemblies, which are often optimized for specific data types or applications, such as data centers, smartphones, and robotics. At the same time, whatever actions agents take need to be compatible with existing tools or flows or processes, which must adapt to the rapid evolution of agentic AI.
“Typically customers will work within a platform, and that could be Cadence Virtuoso, Synopsys Custom Compiler, or Keysight Advanced Design system, and within that environment we guide them through an application-specific workflow,” said Niels Faché, senior vice president at Keysight EDA. “We have in-product documentation to explain all the different steps and setups and possibilities of the tools. That’s all designed for human engineers to collaborate and work on complex products. With AI, that’s becoming a very different model where human engineers work with digital engineers or agents. 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, ‘How can we now modularize it, provide access to different capabilities with the proper APIs. All of that coding is typically done in Python, and 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.”
Making that approach work requires better-defined goals for agents to work both independently and in concert with other agents, and better training of the agents to achieve those goals. “If you want an agent to perform a certain task — it could be synthesizing a filter — you would say, ‘I’d like to design a filter with these characteristics, a bandpass, and this frequency range,'” Faché said. “Then this agent needs to ask, ‘Okay, what is a filter? How do I set up a simulation?’ It will need examples. But then you also need an MCP server that allows you to break down that ask into a sequence of steps and call the right functionality to execute that step. What used to be guided in a graphical user interface is now in the form of example skills that you give to an agent and say, ‘This is how you execute this task.'”
Those agents also need to be able to break through silos where necessary, but obey limitations at other times. Some of the current concern about rogue AI agents is due to increasingly complex queries that require AI agents to look beyond silos, which to AI are just virtual constructs, as well as too much agent autonomy. Using small language models and agents that are incapable of deconstructing those sandboxes or scraping the entire internet will help, so barriers are only crossed when agents are allowed to do so.
“There are still these silos,” said Ankur Gupta, executive vice president and head of EDA IC software at Siemens EDA. “There’s EDA for simulation, all the way into PLM, which is real-world deployment and operation and facilities are all kicking in. It’s not just EDA for design verification. This is where we started talking about digital twins and having some intelligence captured. Now, that’s reality, especially when you put the LLM and SLM in the picture. You’re capturing enough information for each of these silos — the chip, the system, the simulation, the PLM system — for data ontology. LLMs and SLMs will have to work off data, and we are connecting all of that as a layer. On top of that layer is each domain, and it’s interacting with the data in this layer. But it can just as easily interact across these silos.”
There are business cases for this kind of behavior, as well, so there is a lot riding on solving any problems. Pulling together fab data that incorporates multiple processes can save enormous amounts of money and time in chip manufacturing. “The cost of a mistake is $500 million at 3nm and 2nm,” said Gupta. “That’s where these engines have to remain high-fidelity. We use reinforcement learning, unsupervised learning — 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. That doesn’t change.”
Small language models are well-suited for many specialized tasks. “We are seeing that happening for specific applications that are RAG-based (retrieval-augmented generation),” said Shriniwas Kulkarni, engineering leader at SiMa. “Small language models are very efficient for specific use cases, rather than using large language models and quantizing them, and they can train them with their own data to make sure they’re very accurate.
Others agree. “We see people aligning to ideas about agents whose context is limited in scope, and then you have an orchestrator that is looking across this, and so you want that top-down and bottom-up together as you’re building your agent architecture,” said Saurabh Gayen, chief solutions architect at Baya Systems. “What’s often missed when we’re talking about these things is a human in the loop. If you have a black-box view of the entire problem, do you build some black box that is hierarchical to get the best result, or do you want to pause as you go through this process and have a human understand so they can bake into it, ‘Hey, don’t think about it this way. Think about it that way.’ AI isn’t so much a replacement for engineers as a way to enhance what they’re able to do and their understanding.”
Human engineers effectively are the glue that makes AI work. They are essential to figuring out how agents at different layers of a design, or a stack of manufacturing processes, are best optimized together. “We’re all learning this together,” Gayen said. “How you engineer those layers of agents is an art. And if somebody says they’ve found the perfect solution, that’s really just one among many.”
But human engineers also are in short supply. “We went through the wonderful dot-com period where nobody went into semiconductors because they all wanted to be software programmers,” said Michal Siwinski, chief product and marketing officer at Arteris. “There was a lull in the talent pool, which has been recovering slowly, but there’s still a gap in terms of the number of experienced people who actually impact designs. Agentic AI helps to fill some of those voids. In this case, it’s not actually an accelerant. It’s backfilling. How many weekends can someone work on these projects?”
The data crush
At the root of all these challenges and advances are changes in the way data is utilized. Concerns about sandbox break-throughs, reverse engineering, and even the performance and power draw of AI chips, are all dependent on the creation, movement, storage, and security of data. Real-world data is better, but synthetic data can approximate it with better training by expert engineers.
“There are new classes of jobs where, if you are a top 1% expert in any area, you can get a job talking to LLMs,” noted Movellus’ Faisal.
Models will follow, and agents will deploy them. “The fabs aren’t generating these AI models yet,” said Wally Rhines, CEO of Silvaco. “Everyone wants to use Claude for everything. People are simulating all sorts of stuff. The difference here is that in physical processes, you have basic device physics that controls the effects of changing a process. We have years of SPICE modeling and BSIM models. We have TCAD engineers who talk in a language called Jump, and they are a culture unto themselves and tend to continue doing things the way they’ve done them. What you need are innovative people who can generate real models from synthetic data that can be calibrated with real data to build a virtual model.”
“In some cases, you’d really rather sell an agent than software,” Rhines said. “It’s not clear if anybody in EDA is actually pricing based upon tokens rather than traditional pricing models. I don’t think anyone is, but their customers are all talking about it. Selling an agent that will automate the calibration and 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 consume, so it’s a way that EDA can sell a product, get the benefit, and no one sees the underlying data that companies want to keep proprietary.”
The future
Some of the changes underway with AI look very familiar.
“You can draw parallels to the early 2000s, when the gigahertz wars all of a sudden became multicore,” said Movellus’ Faisal. “Everything was going to be a GPU, and then, ‘No, it’s going to be CPUs and GPUs.’ We go through the same architectural churn every 10 or 15 years. The big difference is that back then, a chip was whatever you could do at reticle size. Now the chip is the whole data center. You may have many GPUs working together like one chip. So the bottleneck will shift to data movement, and then it will move to something else once we solve that.”
What’s different, though, are how the pieces are being assembled and how they will be used in the future. “We can talk about how great agentic flows are in terms of how much they improve productivity, and there’s a whole orchestration conversation,” said Arteris’ Siwinski. “You have multi-agent solutions where you orchestrate across domains, which is very interesting because you’re trying to get the analog designer and the verification person and the formal person talking the same language. Hopefully with agentic AI that becomes easier because there’s a common understanding of something. The smart edge is coming, and you will have really good reasons to optimize and to create more unique software.”
The big unknown is where and how exactly engineers will fit into this AI agent-filled world. All indications are that their expertise will still be needed, but how their jobs will change isn’t clear.
“With multi-agent coordination, in the future we’re going to have virtual chip companies that are also going to be virtual AI engineers working together and collaborating to form this agent organization,” said ChipAgents’ Wang. “Then, maybe part of the chip companies will be virtual in the future. If you look at physical AI and robotics, one of the reasons why robotics has not taken off yet is that they don’t have specialized chips. If I open the refrigerator door, it takes a second. A robot takes 20 seconds. So people gave up. But in the future, robotic companies will build their own ASICs. The chip engine, this kind of AI platform, will be very important to enable these physical AI robot companies. They can all build the chips themselves.”
Leave a Reply