NVIDIA has fundamentally expanded how AI agents can work on engineering problems by adding physics simulation and quantum chemistry capabilities to its Agent Toolkit, allowing autonomous AI engineers to design and verify chips up to 20 times faster than traditional methods. The company announced the addition of re-architected PhysicsNeMo libraries and updated CUDA-X libraries, giving developers new tools to build AI agents that can reason about physics, run complex simulations, and generate high-fidelity design data.

What Makes Physics-Aware AI Agents Different?

Until now, most AI agents have been confined to working with text, code, or simple data processing tasks. But chip design and engineering require something fundamentally different: the ability to understand how physical systems behave. NVIDIA's new toolkit bridges that gap by embedding AI physics models directly into agent workflows. This means an AI agent can now train custom physics models, run simulations, and generate design insights without requiring a human engineer to manually set up each step.

The toolkit includes three major new capabilities. First, PhysicsNeMo provides AI physics skills that agents can use to train and deploy customizable models for design and simulation tasks. Second, new sparse solver libraries called cuISS and cuDSS accelerate the complex mathematical calculations that underpin physics simulations, making them practical for production environments. Third, cuEST brings quantum chemistry simulations to the table, enabling density functional theory calculations at scales that were previously too computationally expensive.

"Engineering has reached an inflection point. AI can now work with tools of physics, simulation and design," said Timothy Costa, Vice President and General Manager of Computational Engineering at NVIDIA. "With NVIDIA Agent Toolkit, developers can build agentic engineers that reason using physics, run complex simulations and generate high-fidelity data to become a new engine for innovation in chip and system design."

Timothy Costa, Vice President and General Manager of Computational Engineering at NVIDIA

How Are Industry Leaders Already Using This?

The toolkit isn't theoretical. Major semiconductor and design companies are already deploying these physics-aware agents in production workflows. Cadence reports achieving up to 20 times faster performance for multiphysics simulations in advanced packaging and printed circuit board design. Siemens has deployed agentic workflows that deliver over 10 times faster library characterization, along with more than 10 times reduction in token costs, which translates to lower computational expenses.

The quantum chemistry capabilities are proving particularly valuable. Samsung, Synopsys, and TSMC are using the cuEST quantum chemistry library and achieving up to 50 times speedup in key quantum chemistry workloads. These aren't marginal improvements; they represent fundamental shifts in how long design cycles take and how much computational resources companies need to allocate.

How to Get Started Building Physics-Aware Engineering Agents

• Leverage Nemotron 3 Ultra: NVIDIA's open-source Nemotron 3 Ultra model leads among open models in agentic register-transfer level coding, the specialized language used for chip design. Developers can deploy it locally or on premises, giving enterprises control over their data and customization options without relying on cloud-based APIs.
• Integrate with Existing Design Tools: Developers can get started using Cadence's harness, Synopsys' autonomous agents for design verification, or Siemens' Questa One smart verification platform. These integrations mean you don't need to rebuild your entire design workflow from scratch.
• Start with Sparse Solver Libraries: If you're working on physics simulations or electronic design automation, the new cuISS and cuDSS libraries can be integrated into existing simulation engines to accelerate computations without requiring a complete architectural redesign.

NVIDIA's Nemotron 3 Ultra model is particularly noteworthy because it performs well on comprehensive Verilog design benchmarks, the standard language for describing hardware circuits. Unlike closed-source models, Nemotron 3 Ultra can be fine-tuned on proprietary company data, meaning organizations can customize it for their specific design methodologies and maintain data privacy.

Why Does This Matter Beyond Chip Design?

The implications extend far beyond semiconductor manufacturing. The toolkit demonstrates a broader shift in how AI agents are evolving from general-purpose assistants into specialized domain experts. By embedding physics, simulation, and quantum chemistry capabilities directly into agent frameworks, NVIDIA is showing that the next generation of AI agents won't just process information; they'll understand the underlying physical principles that govern complex systems.

This matters because many industries face similar challenges: aerospace engineers need agents that understand aerodynamics, materials scientists need agents that can predict material properties, and pharmaceutical researchers need agents that can model molecular interactions. NVIDIA's approach suggests a template for how other domains might build specialized AI agents equipped with domain-specific reasoning capabilities.

The expansion of NVIDIA's Agent Toolkit represents a significant step toward making AI agents practical for engineering work that demands precision, domain expertise, and the ability to reason about complex physical systems. With major industry players already reporting dramatic speedups in design cycles and computational costs, the toolkit is moving from research project to production reality.