NVIDIA Expands AI Agent Toolkit to Accelerate Product Innovation

NVIDIA (NVDA-US) today announced the expansion of its engineering-specific NVIDIA Agent Toolkit, introducing NVIDIA PhysicsNeMo and CUDA-X libraries as agent-ready tools and skills to revolutionize global product design and development.

Designing next-generation chips and systems requires NVIDIA’s teams to integrate physics, simulation, and performance analysis across increasingly complex design cycles. A new generation of autonomous AI engineers is emerging to handle these intricate tasks. These AI engineers can use specialized tools, run simulations, and generate high-fidelity data to scale chip design, verification, packaging, and system development.

NVIDIA has re-architected PhysicsNeMo into a series of agent-ready libraries and added and updated multiple CUDA-X libraries, both now integrated into the NVIDIA Agent Toolkit to support complex engineering work. PhysicsNeMo provides AI physics skills for training and deploying models, while CUDA-X libraries bring accelerated solvers and quantum chemistry capabilities into agent-based engineering workflows.

Timothy Costa, Vice President and General Manager of NVIDIA’s Computing Engineering organization, stated that engineering has reached a critical inflection point. AI can now leverage physics, simulation, and design tools. With the NVIDIA Agent Toolkit, developers can build agent-based AI engineers that reason using physical principles, run complex simulations, and generate high-fidelity data—becoming a new engine driving innovation in chip and system design.

The NVIDIA Agent Toolkit enables developers to create dedicated engineering AI assistants and connect them to domain-specific tools, models, and data. With the addition of NVIDIA PhysicsNeMo and CUDA-X libraries, these agents can now apply AI physics skills, accelerated solvers, and quantum chemistry capabilities to chip, system, and industrial engineering domains.

Key features introduced include AI physics skills, iterative sparse solvers, direct sparse solvers, and quantum chemistry.

Leveraging NVIDIA Research’s hardware design agent ACE-RTL, NVIDIA Nemotron 3 Ultra outperformed other open-source models on the comprehensive Verilog Design Problems benchmark, covering various RTL coding tasks.

This achievement demonstrates how Nemotron 3 Ultra delivers industry-leading accuracy and efficiency, and can be post-trained on proprietary data whether deployed on-premises or within enterprise environments, giving companies greater control, customization, and data privacy when building chip design AI agents.

Developers can begin using Nemotron 3 Ultra through Cadence’s testing framework, Synopsys’ fully autonomous long-running agents for design verification and analog/mixed-signal workflows, Siemens’ Questa One intelligent verification agent toolkit, and Hugging Face.

Industrial engineering leaders are already using the new and expanded components of the NVIDIA Agent Toolkit to develop autonomous AI engineers.

Cadence is integrating NVIDIA Nemotron, accelerated computing, and CUDA-X libraries with its recently launched Cadence AuraStack AI Super Agent and Cadence Millennium M2000 platform to autonomously advance advanced packaging and printed circuit board (PCB) design from exploration to signoff, achieving up to 20x faster multi-physics simulation speeds. These technologies will be added to Cadence’s comprehensive chip design super agent portfolio, covering end-to-end chip design workflows from architecture to manufacturing signoff.

Additionally, the collaboration extends from agent-based design to the underlying computing platform. Cadence’s portfolio of electronic design automation (EDA) and system design automation (SDA) tools, including the formal verification platform Cadence Jasper, is being optimized for the NVIDIA Vera CPU to help engineering teams verify advanced chip designs more quickly.

Synopsys is combining the NVIDIA Agent Toolkit, NVIDIA NIM microservices, Nemotron open-source models, NVIDIA NeMo Gym library, and NVIDIA NemoClaw blueprints with Synopsys AgentEngineer to build secure and accelerated agent-based workflows in chip and system design. Using Ansys Icepak, Synopsys’ agent-based workflows can autonomously perform setup, preprocessing, and postprocessing tasks required for complex GPU thermal design optimization. Synopsys is also developing use cases for NVIDIA cuISS to accelerate simulation workloads.

The collaboration scope extends from agent-based workflows to the underlying computing platform. Synopsys’ VCS, a high-performance functional verification solution used to simulate and verify complex chip designs before manufacturing, is being optimized for the NVIDIA Vera CPU to help increase verification data throughput.

Siemens is combining NVIDIA NeMo Gym, Nemotron open-source models, and CUDA-X libraries with Siemens Fuse EDA AI Agent to orchestrate multi-tool and multi-agent workflows across semiconductor, 3D-IC, PCB, and system design, spanning from concept to signoff. In the Siemens Solido Characterization Suite, these agent-based AI workflows have accelerated component library characterization speeds by over 10x while reducing cell costs by over 10x.

Samsung is leveraging NVIDIA cuLitho and CUDA-X libraries to achieve up to 20x faster computational lithography performance and adopting NVIDIA PhysicsNeMo for chip-scale thermal stress analysis, achieving solver-level accuracy on computational domains with up to 10 billion mesh cells.

ChipAgents is using the NVIDIA Agent Toolkit to build AI agents specialized in chip design and verification. The team is fine-tuning NVIDIA Nemotron models for complex end-to-end semiconductor design and verification workflows, covering debugging, formal verification, and coverage analysis.

Silvaco is expanding high-accuracy 3D optical simulation capabilities in Silvaco Victory Device using NVIDIA accelerated computing. Using 32 NVIDIA GPUs interconnected with NVIDIA NVLink technology, the platform completed a photonics edge coupler simulation with 3.2 billion mesh nodes in under four hours—a workload beyond the practical limits of CPU-based simulation.

Keysight is using NVIDIA cuDSS to accelerate electromagnetic simulation speeds by up to 10x, while Samsung, Synopsys, and TSMC are integrating NVIDIA cuEST into their GPU-accelerated pipelines to achieve up to 50x faster execution of critical quantum chemistry workloads.