Cadence and Nvidia Introduce AI Models for Automated EDA Workflows

Generative AI is increasingly being used across technology, including chip and system design. According to Paul Cunningham, senior vice president at Cadence Systems, between 20 and 30 percent of chips currently under development are being designed with assistance from AI agents, and that share is growing rapidly.

Cadence's recent products illustrate how generative AI and high-performance computing can work together in both industry and scientific research.

Cadence's AI Super Agents

Over the past year, Cadence has introduced a series of AI Super Agents:

• AuraStack, for advanced packaging and PCB layout
• ChipStack, for system-on-chip design and verification
• InnoStack, for PHY routing
• ViraStack, for analog circuit design

These systems work with large language models to automate design and testing workflows by coordinating Cadence's existing electronic design automation, testing, and verification tools. The approach is similar to AI coding assistants, which automate software development tasks and rely on tests to determine whether the resulting code works as intended. In EDA, Cadence uses comparable tests to validate chip designs.

However, the effectiveness of these systems depends partly on how closely their training data matches the inputs and outputs they encounter. Cadence's early agents relied largely on general-purpose LLMs, which provided limited exposure to specialized formats such as register-transfer level, or RTL, design data.

Paul Cunningham described general-purpose foundation models as a starting point rather than a complete solution. He said they can still be specialized for particular applications.

Maestro Models and Nvidia Partnership

To provide greater specialization, Cadence announced a partnership with Nvidia and introduced the Maestro AI model family. The models were fine-tuned from Nvidia's Nemotron family to interpret and work with the results produced by Cadence's tools.

Cadence reported that early testing of Maestro showed up to a 30 percent improvement in output quality and up to five times greater token efficiency compared with general-purpose foundation models. According to these results, models post-trained on Cadence's tools produced more capable outputs while requiring fewer tokens to run.

The Maestro platform also supports customization using customer data. Cunningham said Cadence expects this capability to be provided primarily as a managed service rather than as a system customers configure independently.

Like Cadence's existing agent framework, Maestro models and agents can be accessed through an API, operated inside a secure virtual private cloud, or deployed on premises when the required hardware is available.

AI Agents and Scientific Computing

Cadence is not the only company exploring AI agents for coordinating HPC workloads. Nvidia has been working for years to combine artificial intelligence with high-performance computing. Its GPUs are used for many workloads beyond AI training and inference, although those remain major reasons for their adoption.

AI-driven laboratories are also a central part of the US Department of Energy's Genesis Mission, which has been accompanied by the introduction of several new Nvidia-based supercomputers.

Researchers at Sandia National Laboratories have demonstrated how AI agents could help develop and test new hypotheses in what they called a self-driving laboratory. Those experiments, however, did not use modern agent frameworks or even LLMs.

For publicly funded researchers operating with limited resources, an AI laboratory assistant that can continue carrying out tasks after the workday ends could be useful. Cadence's Maestro models and agent systems represent one example of how similar automation is being applied to specialized engineering workflows.