Agentic EDA for AI Hardware: How Siemens and NVIDIA Coordinate Seven EDA Tools in One 3D IC Workflow

The value of this update is easier to see when it is treated as an applied AI system rather than a new model claim. For readers evaluating where artificial intelligence can move enterprise workflows forward, the Siemens and NVIDIA EDA package shows agents operating inside a high-consequence design loop: they connect task orchestration, accelerated simulation, and signoff checks so electrical, power, thermal and co-design questions for 3D ICs can be handled as one coordinated process.

Agentic AI capabilities introduced at DAC 2026 are combined with NVIDIA accelerated computing. The target is 3D integrated circuits, where compute density, packaging, power delivery, and thermal behavior interact too tightly to be reviewed only after each tool finishes. A new coordination layer aligns Siemens electrothermal analysis tools through one workflow. Design Space Exploration and Optimization, also called DSE/DSO, gives engineers a way to test alternatives against physics-based signoff solvers. At the application layer, the workflow can span seven EDA tools while evaluating hundreds of design points per cycle.

The benchmark reporting adds a quantitative frame. It describes 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. The pattern is not agent answers design questions in isolation; it is agent keeps the analysis loop moving while the evidence is still reviewed through solvers.

Why the packaging shift matters for AI application teams

Compact AI packages require more compute capacity, more memory bandwidth, and more thermal control. The NVIDIA Rubin GPU is cited with up to 50 petaflops of NVFP4 inference and up to 288 GB of HBM4 per GPU. GPUs of this class can be assembled into systems such as NVIDIA Vera Rubin NVL72.

When compute and memory are placed closer together, power density and heat concentration increase. The workflow lesson is that electrothermal co-design should be considered during design rather than only during final verification. For application owners, this is where an agent can provide value: it makes cross-domain review repeatable and helps teams avoid narrowing scenario coverage because time or compute constraints feel unavoidable.

Case pattern: an agent-managed 3D IC electrothermal workflow

The workflow begins with a person setting design objectives. The Fuse EDA AI Agent, integrated with Intelligence Center X, coordinates multiple Siemens tools. It runs coupled electrothermal analyses, finds unmet targets, and evaluates whether causes involve power delivery, placement, materials or a combination of factors.

From those results, the agent guides design-space exploration, assesses trade-offs, and reruns selected configurations against the objectives. The result is that engineers can examine more alternatives and surface power and thermal problems earlier in development. In runtime terms, 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 boundaries, power budgets and material stack-ups across 200 to more than 500 design points per cycle.

In plain terms, the agent acts less like a chatbot and more like a persistent test coordinator. It keeps moving between objective, evidence, hypothesis, rerun, and selection, while the human engineer retains the role of defining goals and reviewing signoff-relevant outcomes.

Tool stack: one agent workflow coordinating seven EDA tools

The chain combines tools that cover layout, library characterization, digital power, thermal signoff, system thermal simulation, temperature-aware extraction, and SPICE simulation. The named tools are:

• Innovator3D IC Integrator software: used for floor planning, 3D IC assembly, and package design.
• Solido Characterizer software: used to characterize libraries and generate Liberty files.
• mPower Digital software: used for power analysis, including temperature-dependent instance power.
• Calibre 3DThermal software: used for 3D IC thermal signoff and power mapping.
• Simcenter Flotherm software: used for system-level thermal simulation and visualization.
• Calibre xACT software: used for temperature-dependent parasitic extraction.
• Solido SPICE software: used for electrical SPICE simulation that accounts for power and temperature.

Platform layer: NVIDIA elements that support the agentic workflow

Siemens combines the NVIDIA Agent Toolkit with NVIDIA accelerated computing and CUDA-X libraries to support the agent-driven EDA workflow. Accelerated simulation tools provide analysis results that the agent evaluates during design exploration. The NVIDIA CUDA-X libraries accelerate supported engineering solvers used in the workflow, reducing simulation time during iterative design.

• Nemotron open models: support 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).

Comparison table: manual coordination versus agentic coordination

Dimension	Manual coordination	Agentic coordination in the described workflow
Starting point	Specialists coordinate multiple tools and decide the next hand-off.	Engineer defines design objectives; Fuse EDA AI Agent coordinates the tool chain.
Domain transfer	Data is transferred among electrical, power, thermal and extraction analyses.	Coupled electrothermal analyses are run through one coordinated workflow.
Scenario coverage	Time and computing constraints may lead teams to evaluate fewer scenarios.	The workflow can evaluate hundreds of design points per cycle and 200 to more than 500 design points per cycle in the described configuration.
Execution profile	Coordination depends on specialist bandwidth and sequential tool use.	Nine steps run across seven Siemens EDA tools, with 50 to 100 reasoning steps during an eight- to 10-hour runtime.
Problem discovery	Power and thermal issues may be identified later than desired.	Additional alternatives can be examined, helping power and thermal issues surface earlier in development.

Implementation checklist for AI application reviewers

• Confirm the objective is expressed as design targets, such as power budgets, material stack-ups, and heat transfer coefficient boundaries.
• Confirm which tool chain will be coordinated, including whether seven EDA tools are involved in the workflow being piloted.
• Require evidence to pass through physics-based signoff solvers before design choices are accepted.
• Use Design Space Exploration and Optimization as the exploration mechanism, and preserve DSE/DSO naming in workflow documentation.
• Define how agent trajectories and Model Context Protocol (MCP) tool descriptions will be reviewed and refined.
• Require observability and audit records across agent harnesses before scaling long-running agent workflows.
• Use sandboxed agent execution when intellectual property and foundry process design kits (PDKs) are involved.

Step-by-step deployment path

1. Begin with a bounded 3D IC electrothermal use case rather than a broad AI initiative.
2. State the design objectives that will guide the workflow.
3. Connect the Fuse EDA AI Agent through Intelligence Center X to the relevant Siemens EDA tool chain.
4. Run coupled analysis across electrical, power, thermal, and extraction domains.
5. Let the agent identify unmet targets and classify whether the cause relates to power delivery, placement, materials, or a combination of factors.
6. Use DSE/DSO to explore alternatives, then rerun selected configurations against the objectives.
7. Validate the choices with physics-based signoff solvers.
8. Track runtime behavior, reasoning step count, and design-point coverage to decide whether the coordination layer is expanding exploration while staying trustworthy.

Common pitfalls and boundaries

• Waiting until final verification to address electrothermal co-design. The described workflow argues for considering these interactions during design.
• Leaving manual hand-offs as the only coordination method. As complexity increases, time and computing constraints may reduce scenario coverage.
• Treating agent output as self-validating. The value depends on alternatives being checked against physics-based signoff solvers.
• Omitting agent observability, audit records, and trajectory review in long-running workflows.
• Running agent tools without protection for intellectual property and foundry process design kits (PDKs).

FAQ: applying this 3D IC agentic EDA pattern

What exactly was updated?

Siemens and NVIDIA updated their electronic design automation, or EDA, agentic AI portfolio. The update joins agentic AI capabilities introduced at DAC 2026 with NVIDIA accelerated computing for electrical, power, thermal and co-design analysis of 3D ICs.

Why is the number of tools relevant?

The workflow spans seven EDA tools and coordinates nine steps across seven Siemens EDA tools. For AI application teams, that detail matters because the value is not a single tool feature; it is the coordination of analysis across multiple domains inside one process.

What runtime profile does the agent manage?

The described workflow performs 50 to 100 reasoning steps during an eight- to 10-hour runtime. It evaluates heat transfer coefficient boundaries, power budgets and material stack-ups across 200 to more than 500 design points per cycle, which is also expressed as hundreds of design points per cycle.

What efficiency claims are attached?

Benchmark reporting describes 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.

Does the agent replace physics-based verification?

No. In the described model, Design Space Exploration and Optimization, or DSE/DSO, is used with physics-based signoff solvers. The agent guides exploration and reruns; the evidence is still checked through signoff.

What safeguards are mentioned?

The workflow description includes NVIDIA OpenShell for sandboxed agent execution, NeMo Relay for observability and audit records, and NeMo Gym for tracking agent trajectories and refining Model Context Protocol (MCP) tool descriptions.

Action advice for digital leaders and product teams

• Look for agentic AI pilots where the process is long-running, cross-tool, and outcome measurable, as this EDA case illustrates.
• Start with a narrow 3D IC-style problem definition: objectives, tools, data domains, solver checkpoints, and review responsibilities.
• Use the reported efficiency figures as a comparison frame for pilot design, while validating against your own baseline workflows.
• Make agent governance part of the application review: trajectory tracking, MCP tool description quality, auditability, task routing, and sandboxing.
• Do not separate AI workflow design from engineering validation; the application value here comes from coupling agent coordination with physics-based signoff.