Skip to the stories
Electronics & Semiconductors

AI Agents Could Move Chip Design Earlier, but Tool Coordination and Data Remain Obstacles

  • Ed Sperling
  • 906 reads

Layered AI agents could connect more of the chip design flow, but engineers, proprietary data and operating costs still set limits.

AI Agents Could Move Chip Design Earlier, but Tool Coordination and Data Remain Obstacles

AI agents may help move checks and design decisions earlier in the chip development process, but the harder question is how to make them work together across the full flow. The engineers discussing the shift see a layered system taking shape: agents handle local tasks, other agents coordinate them, and higher levels connect design stages. Building that system will require new methods, access to engineering knowledge and a way to keep the cost manageable.

From individual agents to a coordinated flow

Ashish Darbari, founder and CEO of Axiomise, expects the industry to develop a new methodology for combining agents, much as it adopted UVM to bring different verification activities into a shared approach. He says the right way to connect agents is still unclear. Some companies developing agent-based tools lack the field experience to anticipate what chip designers need, while large chip companies can draw on years of internal data and direct experience—and may build their own systems as a result.

Frank Schirrmeister, executive director of strategy programs and system solutions at Synopsys, describes the challenge as a stack of increasingly broad responsibilities. Chip development spans specification engineering, functional verification, digital implementation, manufacturing and lifecycle monitoring. Within verification alone, teams may use virtual prototyping, formal methods, simulation, emulation and prototyping. An agent could handle one activity; another could coordinate several activities; higher-level agents could connect those groups to the rest of the flow.

Jack Berg, vice president of business development at Silvaco, calls this a hierarchy. He expects the industry to build and improve agents at different levels, eventually including a system-level agent that orchestrates the broader process. Vendors describe some offerings as Level 5 systems, Schirrmeister says, though those capabilities apply within a defined operating domain. The label alone doesn’t establish how well agents will coordinate across tools and stages.

Earlier feedback must still be checked

Moving work earlier in the flow means feeding information from downstream activities back into the specification and functional design. Berg says that feedback can add constraints intended to improve the chance that a design can be manufactured. Schirrmeister asks whether an agent can predict the design path; Berg says it can, but both emphasize the need to check the result. In particular, teams need to assess multi-physics effects early enough to avoid a design that fails when it reaches manufacturing.

That requirement puts a limit on how much of the process can be abstracted. An agent may propose a path, but its output still needs validation against the physical effects and manufacturing constraints that shape the eventual chip. The practical challenge is therefore broader than automating individual tasks: a coordinated flow must carry useful information between stages while keeping the checks that establish whether its decisions hold up.

Ronen Laviv, vice president of business development and strategic partnerships at Hiveware, raises another possible change: the languages and methods used to run verification. Co-simulation can slow emulation, he says. One possibility would be an AI-specific language that produces synthesizable SystemVerilog RTL code for execution in the machine, reducing the software involved and potentially speeding emulation. Laviv presents this as a possibility rather than a settled direction; he also questions whether UVM will remain in use indefinitely.

Darbari points to claims that formal verification could replace simulation, but says the key issue is whether adoption across these technologies produces a coherent methodology. New agents or faster execution in one block won’t, on their own, resolve how the blocks work together.

Engineering experience and private data

Laviv estimates that AI handles about 80% of the challenges under discussion, while experienced engineers retain skills that current systems lack. He cites a discussion of 250 skills spanning power, emulation, formal methods and verification. Capturing the knowledge accumulated over decades—and making it available to agents—is a major outstanding task. Some of that expertise resides within companies such as Apple, Qualcomm and Intel.

Berg compares the long development of AI methods for predicting protein structures with the possibility of progress in semiconductor design. He notes that the 2024 Nobel Prize in Chemistry was awarded in part for work involving AlphaFold, a software system for predicting how proteins fold. But Laviv draws a distinction: data about proteins is open, while chip design information is often proprietary.

Schirrmeister says access to suitable data is a central hurdle for companies building AI systems for electronic design automation. Chipmakers can use their own design data internally, but may not share it externally. He describes Samsung using company data and its own large language model to improve quality of service for a network-on-chip, alongside a fast conventional engine, an SoC simulator and an emulator. The data stays within the company, which helps explain why chipmakers may develop their own systems.

Schirrmeister also describes a balance between language models and agents: models need training grounded in real chip designs and the knowledge of people who design chips, while agents perform tasks within a flow. Companies that know AI may still need to learn the details of EDA; established EDA vendors, in turn, bring knowledge of chip design tools and workflows.

Cost will shape who can adopt the approach

The expense of running agents is another concern. Darbari asks whether smaller companies can afford the computing and token costs of operating them at scale, and suggests that large companies may have an advantage. Schirrmeister sees potential benefits for smaller designs that have been difficult to develop because of the work involved. He adds that competing at the scale of Nvidia remains beyond what a small team can accomplish from a garage.

The likely path is incremental: build agents for specific tasks, improve them through use, and add coordination across the hierarchy. That could reduce some engineering grunt work and make parts of chip design more accessible. Its reach will depend on whether teams can connect the tools, verify predictions, preserve essential engineering judgment and bring operating costs within reach.

Source: Ed Sperling · semiengineering.com

From the same set