A new AI engineering toolkit combines physics-aware AI, GPU-accelerated simulation and domain-specific libraries to automate complex design, verification and simulation tasks across semiconductor, industrial and scientific applications.
NVIDIA has expanded its NVIDIA Agent Toolkit by integrating NVIDIA PhysicsNeMo and updated CUDA-X libraries, enabling developers to build AI engineering agents capable of performing physics-based reasoning, simulation and engineering analysis. Rather than acting only as coding assistants, these AI agents can execute complex engineering tasks across semiconductor, industrial and scientific workflows, helping reduce design time while improving engineering productivity.
The latest version transforms PhysicsNeMo into agent-friendly software libraries that allow AI models to train, deploy and apply physics-informed models within engineering workflows. Combined with CUDA-X libraries, engineering agents gain access to accelerated numerical solvers and computational capabilities required for simulation-intensive applications. This enables AI systems to work with engineering data, perform simulations and generate high-fidelity outputs instead of only producing software code.
The key features are:
- Supports agentic register-transfer level (RTL) coding with Nemotron 3 Ultra
- Includes GPU-optimized cuISS sparse iterative solver library
- Adds cuEST libraries for quantum chemistry workloads
- Modular, agent-ready APIs simplify integration into AI workflows
- Designed for scalable deployment on NVIDIA accelerated computing platforms
A key addition is support for AI-driven multiphysics simulation, sparse linear system solving and quantum chemistry computation. The toolkit incorporates GPU-accelerated libraries for solving large engineering problems encountered in semiconductor design, electronic packaging, PCB development, computational fluid dynamics, structural analysis and materials research. Developers can also build customised engineering agents that interact with domain-specific tools, models and datasets to automate repetitive engineering tasks while supporting human decision-making.
The platform targets semiconductor design and verification, advanced chip packaging, system-level engineering and industrial digital twin development. Engineering software providers are integrating the toolkit to automate workflows such as RTL coding, verification, library characterisation, simulation and packaging analysis. By combining AI reasoning with GPU-accelerated computation, the toolkit is intended to shorten engineering cycles while enabling more complex simulations that would otherwise require significant manual effort.
Beyond chip development, the expanded toolkit can be applied to aerospace, automotive, manufacturing, robotics, life sciences and scientific computing, where physics-aware AI models and accelerated simulations are increasingly required. As engineering projects grow in complexity, the combination of autonomous AI agents, physics-informed machine learning and GPU-accelerated computing aims to help organisations automate design exploration, improve simulation accuracy and accelerate product development from concept through verification.
Click here for the original announcement.
