Agent-based AI for Chip and PCB Design Siemens Lets EDA Agents Check their Own Work

Related Vendors

Taiwan Machine Tool & Accessory Builders' Association (TMBA)

SIMTOS (Seoul International Machine Tool Show)

Initiating a simulation is one thing. Monitoring thousands of calculations over hours, fixing errors, and correctly interpreting the results is another. Siemens and Nvidia are therefore expanding Fuse EDA AI Agent with features for long and "self-verifying" development processes.

Siemens presented the next development stage of the Fuse EDA AI Agent at the Design Automation Conference (DAC) 2026 in Long Beach in July 2026. The system is designed to not only plan development tasks, call EDA tools, and consolidate their results in the future. The agents are also intended to validate their decisions throughout the entire process with Siemens' deterministic and physics-based EDA engines.

To achieve this, Siemens combines the EDA tools with AI models, runtime environments, and GPU-accelerated computing from Nvidia. The expansion is aimed at developers of semiconductors and PCBs and is set to be gradually implemented across much of the Siemens portfolio.

Agent Takes Over Ongoing Error Handling

What this could look like in practice is demonstrated by Siemens using the example of standard cell characterization. Developers must determine the timing, power consumption, and other properties of the cells under various process, voltage, and temperature conditions. Thousands of simulations often run simultaneously on HPC systems for this purpose.

The effort is not only due to computing time, as engineers also have to set up and monitor jobs, respond to failed compute nodes or non-converging simulations, change configurations, and restart calculations. They then check whether the results are plausible. The entire process can take weeks.

Agentive AI in EDA

When AI No Longer Just Assists but Controls EDA Workflows

A long-running Fuse agent is intended to handle most of this work. After an initial instruction, it plans the workflow, launches the necessary tools, and monitors the simulations. If errors occur, it is designed to adjust configurations, restart calculations, and subsequently verify the results. For use with the Solido Characterization Suite, Siemens reports a reduction in runtime by more than a factor of ten. At the same time, token costs are expected to decrease by five to ten times. These figures are provided by Siemens and Nvidia; independent benchmarks are not yet available.

What "Self-Verifying" Means

The term "self-verifying" requires clarification. The AI agent does not independently decide whether a result is correct. Instead, it verifies its work using Siemens' established EDA tools. These tools remain the authoritative instance for simulation, verification, and sign-off. Developers are also meant to maintain control over the process. According to Siemens, it is possible to track the plan the agent is following, the specific task it is executing, and the tools it is using. If an engineer deems the approach unsuitable, they can intervene.

That is no trivial matter. An error in chip design can become costly at the latest after tape-out and delay the market launch by months. An agent that works independently but cannot explain its decisions or verify them with reliable tools would only be of limited use for such tasks.

Nvidia Provides Models and Secured Runtime

Nvidia contributes several components to this partnership. Nemo Gym is intended for training and evaluating the EDA agents. Openshell provides a secure runtime environment with access controls and audit trails. The Nemotron models handle conclusions and tool calls within the workflows. Meanwhile, Cuda-X and GPU-accelerated computing are designed to speed up the actual calculations. This divides the task across multiple layers: the agent plans and coordinates, the Nvidia infrastructure supports its execution, and the EDA engines deliver technically reliable results.

Fuse is not intended to be limited to Siemens tools. According to the company, EDA tools from other manufacturers can also be integrated into the processes via interfaces based on the Model Context Protocol. The agents could then select and combine different tools depending on the specific task.

Deployment Across the EDA Flow

Siemens does not intend to limit the functions to standard cell characterization. The planned applications include synthesis, digital verification, physical implementation, sign-off checks, 3D IC integration, and PCB design. For digital verification, Siemens connects the Questa One Agentic Toolkit with Nvidia's Nemotron 3 Ultra model. The agents are designed to coordinate test workflows, verify results against predefined test environments, and detect errors earlier.

Also new is the Solido Layout Analyzer, which is designed to analyze parasitic and layout-dependent effects in post-layout designs. Developers can query results in natural language and receive suggestions for possible corrections. According to Siemens, STMicroelectronics plans to test and validate the solution during ongoing design operations. However, this does not yet constitute a completed practical application.

Fuse will also be integrated into the Siemens Intelligence Center X. This is intended to allow the agents to eventually connect with processes in development, manufacturing, and the supply chain. When developers will actually be able to use the new functions remains unclear. Siemens has announced them for upcoming releases of its AI-native EDA portfolio but has not yet provided a specific release date.