AIApplications on AIdataNewsGlobal: How an Agentic EDA Workflow Can Coordinate 3D IC Electrothermal Design

For teams that evaluate AI applications in chip and system design, the practical question is no longer whether a language model can summarize tool documentation. The more useful question is whether an agent can keep a multi-tool engineering cycle moving. The Siemens and NVIDIA update described here does exactly that in a narrow, high-stakes setting: it uses seven EDA tools, coordinates electrothermal analysis, and evaluates hundreds of design points per cycle. It connects agentic AI capabilities introduced at DAC 2026 with NVIDIA accelerated computing so that electrical, power, thermal, and co-design analysis for 3D integrated circuits can be handled as one coordinated application workflow.

The application problem this workflow addresses

Modern AI hardware changes the packaging conversation. Systems that serve large models need larger compute capacity, greater memory bandwidth, and more thermal management inside compact packages. The NVIDIA Rubin GPU example illustrates the scale of the demand: it supports up to 50 petaflops of NVFP4 inference and up to 288 GB of HBM4 per GPU. Those GPUs can be combined into larger systems such as NVIDIA Vera Rubin NVL72. When compute and memory are integrated more closely, power density and heat concentration rise at the same time. For an engineering team, that means electrothermal co-design is better treated as an early design activity, not as a last verification task after the architecture is already locked.

Manual coordination is the starting point for improvement

The workflow becomes meaningful because of a familiar bottleneck. Existing workflows often require specialists to coordinate multiple tools, transfer data among electrical, power, thermal, and extraction analyses, and repeat evaluations across design iterations. As complexity increases, time and computing constraints may lead teams to reduce the number of scenarios they evaluate. That can delay identification of power and thermal issues. In application terms, the agent is introduced where the work is repetitive, cross-domain, and difficult for a human team to scale by hand.

How the agentic application is used

An engineer defines the design objectives, and the Fuse EDA AI Agent, integrated with Intelligence Center X, coordinates multiple Siemens tools. The agent runs coupled electrothermal analyses, finds unmet targets, and evaluates whether causes involve power delivery, placement, materials, or a combination of factors. Using analysis results, the agent guides design-space exploration, assesses trade-offs, and reruns selected configurations against the objectives. That helps engineers evaluate additional alternatives and identify power and thermal issues earlier in development.

The coordination layer depends on Siemens Design Space Exploration and Optimization, abbreviated as DSE/DSO. According to the reported benchmark studies, Siemens reported up to a 95% reduction in engineering effort, up to a 99% reduction in workflow setup time, and up to 50 times greater design exploration coverage compared with manual workflows. For application readers, the important point is not only that more simulations run, but that the agent helps decide which alternatives deserve attention.

Operation path an application team can follow

1. Define the design objectives in concrete terms, including the electrical, power, thermal, and co-design targets that should be checked.
2. Let the Fuse EDA AI Agent coordinate the work across the relevant Siemens tool chain rather than asking specialists to move data manually between every analysis stage.
3. Run coupled electrothermal analyses, then let the agent identify unmet targets and possible causes, such as power delivery, placement, materials, or a combination of factors.
4. Use the DSE/DSO layer to guide design-space exploration, assess trade-offs, and rerun selected configurations against the original objectives.
5. Validate alternatives through physics-based signoff solvers before treating a candidate design as acceptable.

Tool chain embedded in the workflow

The workflow coordinates nine steps across seven Siemens EDA tools. It performs 50 to 100 reasoning steps during an eight- to 10-hour runtime and can evaluate heat transfer coefficient (HTC) boundaries, power budgets and material stack-ups across 200 to more than 500 design points per cycle. This level of repeated analysis is what allows the application to move beyond a few manual spot checks and evaluate hundreds of design points per cycle during one coordinated cycle.

• Innovator3D IC Integrator software: supports floor planning and 3D IC assembly and package design.
• Solido Characterizer software: characterizes libraries to generate Liberty files.
• mPower Digital software: performs power analysis and calculates temperature-dependent instance power.
• Calibre 3DThermal software: supports 3D IC thermal signoff and power mapping.
• Simcenter Flotherm software: provides system-level thermal simulation and visualization.
• Calibre xACT software: performs temperature-dependent parasitic extraction.
• Solido SPICE software: supports power- and temperature-aware electrical SPICE simulation.

AI infrastructure supporting the application

The agent does not operate in isolation. Siemens combines the NVIDIA Agent Toolkit with NVIDIA accelerated computing and CUDA-X libraries to support the agent-driven EDA workflow. Accelerated simulation tools supply analysis results that agents evaluate during design exploration.

• Nemotron open models: support the orchestration and execution of agentic EDA subflows.
• NeMo Gym: tracks agent trajectories and identifies areas for refining agent skills and Model Context Protocol (MCP) tool descriptions.
• NeMo RL: supports post-training fine-tuning of language models with EDA knowledge and tool-calling capabilities.
• NeMo Relay: provides observability and audit records across agent harnesses.
• NeMo Switchyard: routes tasks based on compute requirements, accuracy and token costs.
• NVIDIA OpenShell: sandboxes agent execution to help protect intellectual property and foundry process design kits (PDKs).
• NVIDIA CUDA-X libraries: accelerate supported engineering solvers used in the workflow to reduce simulation time during iterative design.

Value judgment for application teams

From an application standpoint, this kind of agentic workflow is most useful when design decisions are tightly coupled and repeated simulation is the main cost. A 3D IC electrothermal program has exactly those conditions: electrical behavior, power delivery, thermal distribution, package choices, and material stack-ups interact, and a change in one area can require new evaluations in another. The value is therefore not only automation for its own sake, but broader coverage of alternatives under the same time and compute constraints.

The published metrics help frame that value. Up to a 95% reduction in engineering effort suggests less manual coordination. Up to a 99% reduction in workflow setup time suggests the application may reach productive runs faster. Up to 50 times greater design exploration coverage suggests that the workflow can consider more combinations than a manually operated process might reach. Still, these are benchmark reports, not universal guarantees; teams should compare them with their own design objectives, tool maturity, and hardware environment.

Common misconceptions

• The agent replaces design judgment. The described application still begins with an engineer defining design objectives and ends with validation against those objectives.
• Agentic coordination is only a model improvement. In this workflow, the outcome depends on the Fuse EDA AI Agent, Intelligence Center X, the Siemens tool chain, physics-based signoff solvers, and NVIDIA accelerated computing together.
• More explored points means less rigor. The workflow evaluates alternatives through DSE/DSO and uses physics-based signoff solvers, so speed is paired with formal analysis.
• Sensitive design assets can be freely exposed to any agent environment. The update mentions NVIDIA OpenShell to sandbox agent execution and help protect intellectual property and foundry process design kits (PDKs).

Action checklist for teams considering a similar AI application

1. Start with a clearly bounded workflow where objectives, constraints, and signoff criteria are written down before the agent begins exploration.
2. Map the current manual chain across electrical, power, thermal, extraction, package, and system-level analysis to understand where agent coordination would matter most.
3. Choose a pilot design set that can reveal whether the application improves coverage, setup time, and issue identification without reducing verification confidence.
4. Instrument the agent run through observability and audit records, so the team can see task routing, tool calls, and reasoning traces rather than only the final output.
5. Review tool interfaces and Model Context Protocol (MCP) descriptions before assuming the agent can operate every EDA step reliably.
6. Protect proprietary process data and PDK access through sandboxed execution where the platform supports it.

For AIApplications readers, the Siemens and NVIDIA update is useful as a pattern for where AI can be embedded in enterprise engineering: not as a chatbot over documents, but as a coordinating layer across a complex tool chain. If the application can reliably connect objectives, analysis, trade-off evaluation, and signoff, the practical gain is a workflow that examines more alternatives while giving engineers an earlier view of power and thermal risk.