Chip Verification Consumes 70% of Design Time! Siemens and NVIDIA Launch Self-Validating Agentic AI, Reducing Characterization Cycles by Over 10x

As AI chips, chiplets, and 3D ICs grow increasingly complex, the real bottleneck to product time-to-market may no longer be how chips are designed—but how to prove those designs are correct. Chip verification can consume up to 70% of total design time. Today (3rd), Siemens AG (German stock code: SIE) announced an expanded collaboration with NVIDIA (NASDAQ: NVDA) to upgrade its Fuse EDA AI Agent system, enabling AI agents to evolve from automatically executing engineering tasks to autonomously checking, validating, and correcting their own decisions. This technology has already been applied to characterization workflows for custom IC component libraries, reducing cycle times by over 10x and cutting token costs by 5–10x.

Siemens stated that this collaboration integrates NVIDIA’s AI infrastructure, inference models, and accelerated computing technologies into an Agentic AI workflow with built-in 'self-validation' capability. AI agents can not only reason, plan, and invoke various EDA tools to execute tasks, but also continuously validate results using deterministic, physics-based EDA engines—creating a closed-loop process of 'reasoning, execution, validation, and correction'.

It should be noted that the expanded EDA AI features will be rolled out gradually with Siemens’ upcoming AI-native EDA product versions and are not all immediately available. However, certain Solido custom IC workflows have already been validated and deployed in production, becoming the first quantifiable success case of Agentic AI in EDA.

Merely enabling AI to operate tools is insufficient—it must also prove that design outcomes are correct. Unlike general generative AI applications, semiconductor and PCB design involve engineering constraints such as timing, power, electrical performance, layout, manufacturing rules, and reliability. Even if an AI-generated answer appears semantically reasonable, it may fail engineering validation or be unsuitable for mass production. This makes the ability of Agentic AI to validate decisions using trusted engineering tools a critical threshold for real-world adoption in EDA.

Amit Gupta, Senior Vice President and AI Strategy Chief at Siemens EDA, stated that the expanded collaboration with NVIDIA combines domain-specific industrial AI, physics-based EDA engines, and accelerated computing to establish self-validating AI workflows. Gupta emphasized that by enabling long-running EDA AI agents to continuously and accurately validate decisions using mature engineering tools, customers can accelerate product development, improve design quality, and increase engineering teams’ confidence in outputs.

Timothy Costa, Vice President and General Manager of Computing Engineering at NVIDIA, noted that semiconductor and PCB design represent some of the world’s most complex engineering challenges, requiring AI agents to have access to trusted tools for reasoning, execution, and validation. By integrating Siemens’ EDA software with NVIDIA’s accelerated computing, AI software, and Nemotron open models, the goal is to enable Agentic AI to cover broader segments of the design process.

In the overall technical architecture, Siemens will use NVIDIA’s NeMo Gym open library to build and optimize domain-specific EDA agents, allowing the system to accumulate experience across projects, refine execution strategies and workflows, and improve the efficiency of design knowledge and project context utilization.

NVIDIA’s Nemotron models and Switchyard provide the reasoning power needed for complex engineering trade-offs, while OpenShell delivers a secure execution environment, access control, and audit trails required by enterprises—ensuring AI agents remain within corporate governance when accessing design data and EDA tools. NVIDIA’s accelerated computing and CUDA-X libraries simultaneously support AI model inference and EDA engine computation.

Siemens stated that this architecture improves output quality, tool invocation reliability, and token efficiency for long-duration engineering tasks. In certain applicable workloads, design teams can achieve signoff-quality results in hours without sacrificing accuracy—tasks that previously took days.

The Fuse EDA AI Agent has also been integrated into Siemens’ Intelligence Center X, enabling enterprises to build, manage, and orchestrate multiple AI agents within an industrial-grade AI environment, extending applications from single design tools to cross-domain decision-making across design, manufacturing, and supply chain.

From chip synthesis to PCB design: Agentic AI connects the full EDA workflow. Siemens plans to deploy specialized AI agents across semiconductor and PCB lifecycles, covering synthesis, verification, physical implementation, signoff, and design-for-test (DFT), improving convergence efficiency across tools and domains.

Products involved include Catapult high-level synthesis, Questa One and Veloce verification platforms, Solido custom IC design and verification tools, Aprisa physical implementation, Calibre signoff verification, and Tessent DFT solutions. Applications extend to Innovator 3D IC for advanced 3D integration and Xpedition PCB design.

This signifies that Siemens’ Agentic AI strategy goes beyond enabling engineers to control single software via natural language—it aims to deploy multiple specialized AI agents that collaborate, exchange information, and continuously validate each other’s outputs using engineering tools across the entire EDA process.

Custom IC leads the way: token costs reduced by up to 10x. In custom IC design, Siemens has further enhanced the AI agent capabilities of the Solido Characterization Suite. This workflow can invoke Solido LibSPICE, Solido Generator, and Solido Analytics to automatically run Solido Characterizer, generating and validating Liberty files and automating component library characterization.

Liberty files record timing, power, and electrical characteristics of standard cells and are foundational for downstream digital chip design and verification. Siemens stated that this solution has been validated and deployed in production for advanced process standard cell libraries, memory, and custom IP libraries, reducing characterization cycle times by over 10x and lowering token costs by 5–10x.

However, the 'over 10x cycle reduction' refers specifically to certain component library characterization workflows and does not imply all chip design or verification processes can achieve the same improvement.

Siemens also launched Solido Layout Analyzer, extending AI agents into custom IC layout analysis. This tool helps engineers observe parasitic and layout-dependent effects in post-layout designs, interpret results via natural language prompts, suggest corrections, and automatically generate reports.

Gianbattista Lo Giudice, Design Manager for Non-Volatile Memory at STMicroelectronics (NASDAQ: STM), stated that for memory design teams, linking layout insights directly to electrical performance is crucial. Solido Layout Analyzer enables early-stage layout-dependent effect analysis, preventing late-stage discovery of issues and potentially reducing debugging time for complex design modules by several weeks.

STMicroelectronics plans to validate actual benefits in ongoing design projects, so the 'several weeks reduction in debugging time' remains an expected outcome pending further project validation.

Digital verification is another key focus. AI agents must detect issues across billions of test scenarios. Digital verification is another major focus of Siemens’ Agentic AI strategy. Siemens noted that verification can consume up to 70% of total chip design time. As AI chips, chiplets, and 3D ICs increase in scale and complexity, traditional verification methods face growing volume and schedule pressures.

Building on its previously launched Questa One Agentic Toolkit, Siemens has further integrated NVIDIA’s Nemotron 3 Ultra inference model, designed for complex, long-duration AI agent tasks. In RTL benchmarking with ACE-RTL agents, Nemotron 3 Ultra demonstrated superior performance among open models, enabling AI agents to evaluate different engineering trade-offs and continuously validate results against standard test environments—accelerating issue detection and verification convergence.

Abhi Kolpekwar, Senior Vice President and General Manager of Digital Verification Technologies at Siemens EDA, stated that the industry is at a critical inflection point: the complexity of AI chips, chiplets, and 3D ICs is gradually exceeding the limits of traditional verification methods, making Agentic AI a natural evolution. He noted that in the future, AI agents will collaborate…