According to Paul Cunningham, senior VP at Cadence Systems, between 20 and 30 % of chips currently under development are designed with the help of AI agents, and this proportion is rising rapidly. Cadence has recently rolled out a series of tools called “AI Super Agents” and announced a partnership with Nvidia around a new family of models named Maestro.
- Cadence et Nvidia lancent Maestro, une famille de modèles spécialisée pour leurs outils.
- Les agents IA pilotent déjà les outils EDA existants de Cadence.
- Les modèles Maestro améliorent la qualité des résultats jusqu’à 30 % et l’efficacité en tokens jusqu’à 5 fois.
- L’IA agentique gagne aussi le calcul haute performance et les laboratoires autonomes.
Agents that drive existing tools
Specialized in computer‑aided design software for electronics (EDA), Cadence offers, among others, AuraStack for advanced packaging and printed‑circuit‑board (PCB) design, ChipStack for system‑on‑chip (SoC) design and verification, InnoStack for PHY routing, and ViraStack for analog circuit design.
Based on Anthropic’s Claude Code or OpenAI’s Codex, these agents rely on software layers called “harnesses” that pair language models (LLMs) with automation of design and test flows. They orchestrate Cadence’s EDA, test and verification tools. The logic mirrors that of programming assistants: code either compiles or it doesn’t, and Cadence has tests intended to verify that its designs will work as expected.
The effectiveness of these tests, however, depends heavily on how close the models’ training data are to the inputs and outputs they process. Cadence’s early agents relied heavily on off‑the‑shelf LLMs, which had little exposure to register‑level description (RTL), the language that describes a circuit’s operation. “It is not a question that these ready‑to‑use foundation models, despite their remarkable power, can do everything on their own. They are the starting point, and they can still be specialized,” said Paul Cunningham.
Maestro models trained on Cadence tools
Cadence therefore announced a partnership with Nvidia and launched the Maestro model family. These models were fine‑tuned (i.e., specialized from an existing model) using Nvidia’s Nemotron family, to work with Cadence’s tools and interpret their results.
According to Cadence, preliminary tests showed an improvement “up to” 30 % in result quality, as well as token efficiency “up to” 5× that of general‑purpose foundation models. The token, a unit of text processed by a model, is used here to measure compute cost. The company says that models post‑trained on its tools are therefore both more performant and cheaper to run.
The Maestro platform also allows models to be customized from customer data. Paul Cunningham clarified that this option would be offered as a managed service, not as an approach that customers would undertake on their own. The models and agents can be accessed via an API, in a secure VPC (isolated virtual private cloud) or, for enterprises with the necessary hardware, on‑premises.
An interest that goes beyond chip design
Cadence is not the only one exploring the use of AI agents to drive high‑performance computing (HPC) workloads. Nvidia has been bringing HPC and AI together for several years, its GPUs (graphics processing units) serving many purposes beyond training and running models, even though most buyers use them for that.
This idea of AI‑driven labs is also at the heart of the U.S. Department of Energy’s Genesis Mission, under which numerous new large Nvidia‑based supercomputers have been announced. Researchers at Sandia National Labs have also presented a so‑called “self‑driving lab,” where AI agents develop and test new hypotheses. These experiments, however, did not benefit from modern harnesses nor even from LLMs. In a context of limited public‑research funding, the notion of a lab assistant capable of continuing tasks after researchers leave can seem attractive.
