IN Brief:
- Fuse EDA AI Agent will orchestrate engineering tasks across semiconductor, package, and PCB workflows.
- AI-generated decisions are checked against deterministic, physics-based Siemens EDA tools before progressing.
- Initial characterisation work has recorded faster turnaround and lower token consumption than conventional agentic approaches.
Siemens is introducing an artificial-intelligence layer that connects generative engineering workflows with deterministic electronic design automation tools, allowing autonomous tasks to be tested against established simulation, verification, implementation, and sign-off engines.
Built around the Fuse EDA AI Agent system and NVIDIA infrastructure, the architecture is intended to coordinate work across semiconductor, advanced-package, and PCB development. Rather than passing an AI-generated decision directly into the design flow, Fuse can submit it to physics-based tools that assess electrical behaviour, timing, manufacturability, test coverage, and other engineering constraints.
Across the full development chain, the planned coverage extends from high-level synthesis and functional verification to physical implementation, design-for-test, sign-off, 3D IC integration, and board development. Siemens tools within the wider environment include Catapult, Questa One, Veloce, Solido, Aprisa, Calibre, Tessent, Innovator3D IC, and Xpedition.
NVIDIA NeMo Gym, OpenShell, Nemotron models, Switchyard, and CUDA-X libraries provide the supporting AI and accelerated-computing infrastructure. Their inclusion allows specialist agents to run for extended periods, invoke design tools, review intermediate results, and revise a task without requiring an engineer to issue each command individually.
Within the Solido Characterization Suite, agentic workflows are being applied to standard-cell and IP library characterisation, including the generation and verification of Liberty files. Selected production tasks have recorded turnaround improvements exceeding tenfold, while token expenditure has fallen by between five and ten times.
Library characterisation suits this form of automation because it combines a large number of repeatable simulations with data checks, exception handling, corner analysis, and model generation. The workload remains computationally substantial, yet much of its structure can be defined precisely enough for an agent to execute and verify.
Verification remains the limiting discipline
Although chip design contains many opportunities for automation, verification can consume as much as 70% of the development schedule as functional modes, power states, safety requirements, software interactions, and physical effects accumulate. Complexity rises further when several dies, memory stacks, and package structures must be evaluated as one system.
An agent’s ability to create scripts, constraints, or implementation alternatives is therefore only one part of the engineering task. Syntactically correct output may still introduce a timing violation, weaken testability, break a physical rule, or produce a layout whose behaviour changes materially after parasitic extraction.
By connecting each autonomous action to deterministic tools, the workflow can turn failed checks into structured feedback. A generated constraint can be tested through timing analysis, a physical implementation can be evaluated by sign-off tools, and a verification agent can measure whether newly produced tests have improved coverage rather than merely increased test volume.
Such autonomy also increases the importance of traceability, since every decision must retain a record of the model, process-design kit, constraint set, tool version, and verification result that informed it. Automotive, medical, aerospace, and industrial programmes will require those records to survive formal review, configuration control, and later fault investigation.
Cadence has pursued a related direction through agentic automation spanning PCB and package development, indicating that EDA suppliers are moving beyond isolated copilots towards coordinated engineering systems. Differentiation will depend increasingly on the depth of tool integration, the quality of engineering feedback, and the extent to which automated decisions remain auditable.
Engineers are unlikely to delegate final sign-off wholesale, particularly where the consequences of a missed defect are severe. Repetitive investigation, regression analysis, constraint checking, result triage, and library work nevertheless offer substantial scope for automation when the supporting evidence remains accessible.
Forthcoming releases will extend the number of Fuse-supported tools and workflows. Production adoption will rest on repeatability across large designs, mixed tool environments, changing process technologies, and organisations whose existing automation has developed over several generations of products.
