NVIDIA expanded its NVIDIA Agent Toolkit for engineering on July 26, adding PhysicsNeMo and CUDA-X libraries designed to help developers build autonomous AI engineers. The company said the tools will support chip design, verification, simulation, packaging and industrial engineering by connecting AI agents with physics models, accelerated solvers and quantum chemistry capabilities.
The expansion comes as semiconductor and engineering companies seek to manage increasingly complex design cycles involving physical simulation, performance analysis and verification. NVIDIA presented the toolkit as a software layer through which specialized AI agents can use engineering models, data and computational tools within defined workflows.
NVIDIA Expands Its Engineering AI Toolkit
NVIDIA said it re-architected PhysicsNeMo as a collection of agent-friendly libraries and added new and updated CUDA-X components to NVIDIA Agent Toolkit. PhysicsNeMo provides tools for training and deploying customizable AI physics models, while the CUDA-X additions support mathematical solvers and electronic-structure calculations.
The company described these capabilities as callable software tools that autonomous agents can invoke while completing engineering tasks. In practice, the agents could prepare simulations, process results and generate high-fidelity data, although their performance would remain dependent on the models, software controls and engineering processes used by each organization.
Autonomous AI Engineers Address Design Complexity
NVIDIA said the toolkit connects specialized assistants with domain-specific tools, models and data. These systems are intended to augment engineers across chip architecture, verification, advanced packaging, printed circuit board design and system development rather than operate as a single general-purpose design application.
Additionally, NVIDIA Vice President Timothy Costa said AI can now work with tools used for physics, simulation and design. His statement framed agentic engineering as a way to increase the scale of computational work, while the disclosed deployments show that companies are applying the technology to bounded simulation, coding and verification tasks.
PhysicsNeMo and CUDA-X Add Engineering Skills
The expanded toolkit includes PhysicsNeMo for AI physics models, cuISS for iterative sparse calculations, cuDSS for direct sparse calculations and cuEST for quantum chemistry. NVIDIA said these libraries are designed for GPU execution and can be integrated into production engineering workflows.
| Indicator | New Capability | Context |
|---|---|---|
| AI physics | PhysicsNeMo re-architected as agent-ready libraries | NVIDIA said developers can turn customizable physics-model architectures into callable tools for design and simulation workflows. |
| Iterative solvers | New cuISS library added | NVIDIA said cuISS accelerates large sparse linear systems used in physics-based and engineering simulations. |
| Direct solvers | cuDSS integrated into agent workflows | NVIDIA said cuDSS supports electronic design automation and scientific simulations across multi-GPU and multi-node environments. |
| Quantum chemistry | cuEST added for production-scale calculations | NVIDIA said cuEST supports density functional theory and post-DFT methods for increasingly large ground-state and excited-state simulations. |
New Agent-Ready Libraries and Capabilities
The libraries address computational operations that frequently appear inside larger engineering processes. For example, sparse solvers are used when simulations generate extensive systems of equations, while quantum chemistry tools help model material and electronic behavior at scales relevant to device development.
However, NVIDIA did not present the toolkit as eliminating conventional validation. The announced components provide accelerated computational functions, while engineering organizations remain responsible for configuring workflows, testing outputs and determining whether results meet design and production requirements.
Nemotron Advances Agentic Chip Coding
NVIDIA also highlighted Nemotron 3 Ultra, an open model used with the company’s ACE-RTL research agent for register-transfer level coding. RTL describes the digital logic and data movement used in semiconductor design, making accurate code generation and verification important before a chip reaches fabrication.
They reported that ACE-RTL with Nemotron 3 Ultra achieved a 97.1% average pass rate across nine categories in the Comprehensive Verilog Design Problems benchmark. NVIDIA said the system used an iterative generate, test and reflect process and consumed up to 71% fewer tokens per iteration than selected comparison models.
Enterprise Deployment and Model Customization
NVIDIA said Nemotron 3 Ultra can be post-trained with proprietary information and deployed on premises or within private computing environments. That approach may allow semiconductor companies to customize agents while retaining greater control over confidential design data and internal workflows.
Developers can access the model through systems associated with Cadence, Synopsys and Siemens, as well as through Hugging Face. The benchmark measures defined RTL tasks, however, and does not by itself establish performance across every commercial chip-design environment.
Engineering Companies Deploy Autonomous Workflows
NVIDIA named Cadence, Synopsys, Siemens, Samsung, ChipAgents, Silvaco, Keysight and TSMC among organizations using its accelerated computing or agentic engineering technologies. The reported applications cover packaging, circuit boards, verification, cooling analysis, lithography, optical simulation and quantum chemistry.
Cadence said its AuraStack AI Super Agent coordinates specialized agents across PCB and advanced-packaging workflows. NVIDIA reported that the combination of Cadence software, its Millennium M2000 platform and CUDA-X technology delivers up to 20 times faster multiphysics performance, while Cadence separately reported gains including up to 15 times higher productivity in selected workflows.
Reported Engineering Performance Results
- Library characterization: NVIDIA said Siemens’ Solido Characterization Suite is delivering more than 10 times faster characterization while reducing token costs by more than 10 times.
- Computational lithography: NVIDIA reported that Samsung is using cuLitho and CUDA-X libraries to achieve performance gains of up to 20 times.
- Thermal-stress analysis: NVIDIA said Samsung applied PhysicsNeMo to chip-scale domains containing up to 10 billion cells with numerical solver-level accuracy.
- Optical simulation: NVIDIA said Silvaco completed a 3.2-billion-mesh-node simulation in under four hours using 32 NVIDIA GPUs connected through NVLink.
- Electromagnetic simulation: NVIDIA reported that Keysight is using cuDSS to accelerate selected workloads by up to 10 times.
- Quantum chemistry: NVIDIA said Samsung, Synopsys and TSMC are integrating cuEST into GPU pipelines that deliver speedups of up to 50 times for selected workloads.
The figures describe workloads selected and reported by NVIDIA and its participating companies. They indicate substantial acceleration in defined environments, but results may vary according to model size, hardware configuration, simulation requirements and the baseline systems used for comparison.
NVIDIA Extends Its Industry Partnerships
NVIDIA said Synopsys is using Agent Toolkit components with AgentEngineer to automate simulation preparation and processing for GPU cooling optimization. Meanwhile, Synopsys VCS and Cadence Jasper are being optimized for NVIDIA’s Vera CPU to improve functional and formal verification workloads.
Siemens is combining Nemotron models, NeMo Gym and CUDA-X libraries with its Fuse EDA AI Agent to coordinate multiple tools and agents across semiconductor, 3D integrated-circuit, PCB and system design. These partnerships extend NVIDIA’s role from supplying GPUs into models, libraries, agent runtimes and engineering-software integrations.
Stakeholder Comments
Timothy Costa, Vice President and General Manager of Computational Engineering NVIDIA, said:
“Engineering has reached an inflection point. AI can now work with tools of physics, simulation and design. 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.”
The statement reflects NVIDIA’s commercial and technical position as the announcement’s originating company. The disclosed benchmarks and customer results provide measurable examples, while broader adoption will depend on whether organizations can integrate the tools with existing engineering controls, verification standards and computing infrastructure.
NVIDIA’s expanded Agent Toolkit combines AI physics models, GPU-accelerated solvers, quantum chemistry libraries and an open coding model within a common framework for specialized engineering agents. The initial deployments show applications across semiconductor design, simulation, verification and manufacturing preparation.
The reported performance gains are tied to specific partner workloads and configurations. The announcement illustrates how engineering-software companies are beginning to embed autonomous agents within defined engineering workflows rather than limiting AI to general coding or conversational assistance.
Sources: NVIDIA Newsroom, NVIDIA Technical Blog, Cadence.
Prepared by Ivan Alexander Golden, Founder of THX News, an independent news organization delivering timely insights from global official sources.
Research combines AI-assisted analysis with human-edited accuracy and context.
