Agentic AI Drives Verification Closure and Thermal Analysis

The Synopsys team recently returned from DAC, the Chips to Systems Conference. We’ve been exhibiting at DAC for nearly 40 years now, and it’s always a highlight event for us. This was the first time that it was held in Long Beach, California, but this new venue drew quite a crowd. There is no doubt that the use of AI for chip development was the key theme of this year’s show. In this post, we’d like to discuss some of the exciting applications for AI developed in partnership with Nvidia that we demonstrated at DAC.

It’s only been a few years since the electronic design automation (EDA) industry started including AI-based features in our products. For many vendors, initial support came in the form of chatbots that provided tool support. Modern EDA solutions are complex due to the number of features and options available. When an AI assistant or copilot has access to a knowledge based of domain-specific expertise, it can answer detailed questions and provide guidance on how to get the most benefit from EDA tools.

Of course, Synopsys has also made this type of technology available to our users. However, we have gone well beyond chatbots to offer generative AI, in which the tool results are improved or wholly produced automatically by AI. If an assistant advises users to do something a certain way, it is even more helpful if it just does it by itself. Of course, any generation is performed under user control and is subject to user constraints. It does no one any good if AI generates a result that doesn’t satisfy the requirements of the chip specification.

Today, we have embraced agentic AI, which takes automation to a whole new level. A typical definition of agentic AI is an autonomous system capable of detailed planning, advanced reasoning, using and coordinating external tools, and executing multi-step workflows. It is perhaps a stretch to say that agentic AI acts like a human, but it does perform independently with minimal user guidance. In the EDA domain, it can help produce a chip that meets project goals in less time and with fewer human resources.

AI agents can be classified into three different levels:

• Task-level agents call the EDA tools and execute iterative, repetitive tasks that generate data. The are similar to copilots in that they are prompted by the user and provide a response, but with the addition of generative capabilities.
• Multi-agent workflows are built from multiple task-level agents that both consume and generate design data. They operate autonomously, although with a human in the loop to provide guidance.
• Long-running agent workflows are fully autonomous, capable of executing multiple steps in the EDA development process and generating complete results.

Every chip verification team knows all too well the challenge of coverage closure. Achieving the targets for code coverage and functional coverage is slow and painful. Just running more constrained-random tests isn’t enough once progress stalls. Verification engineers spend weeks tweaking constraints and sometimes even hand-writing tests to hit the last few percentage points in the coverage metrics. This takes a lot of resources, leads to some sleepless nights, and delays tape-out.

The Synopsys Design Verification (DV) Agent bends the coverage curve upward, improving time to closure from weeks to hours. The verification team provides DV Agent with the design specification, the RTL code, the existing test repository, and the coverage targets. The agent then builds a verification plan, generates testbenches, runs regression, identifies failing regression and uncovered scenarios, closes coverage, and resolves test failures in one autonomous flow. Team benefits include faster time to results, higher quality coverage results, and automated root cause analysis (RCA) to fix any failures.

We developed our solution in cooperation with Nvidia to take advantage of GPU acceleration. It is built on the NVIDIA Agentic Stack and secured by NVIDIA OpenShell.

Each new generation of silicon brings higher power density, tighter thermal margins, and greater risk when thermal design falls short. However, finding an optimal thermal design manually remains time consuming, entailing a lot of trial and error. To address this challenge, Synopsys has built end to end fully autonomous EDA and system design agents for thermal analysis and simulation. It uses industry-proven solvers to ensure that accuracy is preserved in the faster automated flow.

This solution is built with Ansys Icepak electronics cooling simulation software, which is now part of the Synopsys portfolio, and open-source PyAEDT libraries. The fully autonomous agentic workflow was also developed with Nvidia. It uses NVIDIA CUDA-X libraries to extend the value of GPU acceleration to thermal analysis. Our agents integrate the NVIDIA NeMo agent toolkit, NVIDIA Nemotron open models, and NVIDIA OpenShell to accelerate the development of secure and governance-based workflows.

Our demo shows Synopsys Icepak Agent, an autonomous engineering assistant purpose built for thermal simulation. The agentic workflow autonomously executes simulation setup, pre-processing, and post-processing in a fraction of the time required for traditional approaches. The engineer provides the model, the simulation inputs, and the desired outputs through a single natural language prompt, and agentic AI does the rest.

As this year’s DAC clearly demonstrated, the EDA industry is investing heavily in agentic AI—and user interest is hitting a tipping point. Synopsys stood out at the show by moving past the hype to demonstrate proven, working agentic AI solutions. Crucially, our vision extends far beyond silicon alone; we are uniquely positioned to serve the full silicon-to-systems needs of the industry.

For chip developers, the benefits are already transformative. By deploying our autonomous, closed-loop Design Verification Agent for EDA, weeks of manual verification are dramatically reduced to hours. Meanwhile, our specialized Icepak Agent automates complex CAE thermal analysis, seamlessly aligning with the way modern system engineers actually work. This is just the beginning of our Silicon-to-Systems agentic AI journey—watch this space for future posts on how our solutions are expanding.