NVIDIA has added PhysicsNeMo and CUDA-X libraries to its Agent Toolkit, giving developers components to build AI engineering assistants that run physics simulations, solve sparse linear systems, and handle quantum chemistry workloads for chip and system design.
“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.”
NVIDIA Agent Toolkit Adds PhysicsNeMo and CUDA-X
PhysicsNeMo was re-architected into agent-callable libraries that train and deploy AI physics models for design and simulation tasks. The CUDA-X update adds three solver libraries.
Key capabilities include:
- AI physics skills: NVIDIA PhysicsNeMo libraries let agents train and deploy customizable AI physics models, turning model architectures into callable tools for engineering workflows.
- Iterative sparse solvers: NVIDIA cuISS (CUDA Iterative Sparse Solvers) library accelerates large sparse linear systems in physics-based and engineering simulations. Its composable solvers and preconditioners are built for GPU performance and scale to production simulation workloads.
- Direct sparse solvers: NVIDIA cuDSS (CUDA Direct Sparse Solvers) accelerates large sparse linear systems for electronic design automation (EDA) and scientific simulation, covering device, circuit, and system workloads with multi-GPU and multi-node scaling.
- Quantum chemistry: NVIDIA cuEST (CUDA Electronic Structure Theory) brings density functional theory and post-DFT methods into production workflows, supporting a wide range of functionals for ground-state and excited-state simulations on NVIDIA GPUs.
NVIDIA Nemotron 3 Ultra Open Model
NVIDIA reported that Nemotron 3 Ultra leads among open models in agentic register-transfer level (RTL) coding on the Verilog design problems benchmark. The ACE-RTL agent for hardware design, built by NVIDIA Research, runs on the model. NVIDIA Nemotron 3 Ultra can be post-trained on proprietary data and deployed locally or on premises. Developers can access it through Cadence, Synopsys, Siemens, and Hugging Face.
Partner Deployments
Cadence is using Nemotron, accelerated computing, and CUDA-X with its AuraStack AI Super Agent and Millennium M2000 platform to drive advanced packaging and PCB design from exploration through signoff, delivering up to 20x faster multiphysics performance. Cadence Jasper is being optimized for the NVIDIA Vera CPU for chip design validation.
Synopsys is using the NVIDIA Agent Toolkit, NVIDIA NIM microservices, Nemotron models, NVIDIA NeMo Gym, and NVIDIA NemoClaw blueprints with its AgentEngineer platform. Using Ansys Icepak, Synopsys ran an agentic workflow that executed simulation setup and pre- and post-processing for GPU cooling design optimization. Synopsys is developing cuISS use cases to speed up simulation workloads, and Synopsys VCS is being optimized for the NVIDIA Vera CPU to improve verification throughput.
Siemens is using NVIDIA NeMo Gym, Nemotron models, and CUDA-X with its Fuse EDA AI Agent for multi-tool, multi-agent workflows across semiconductor, 3D-IC, PCB, and system design. In the Solido Characterization Suite, these workflows delivered more than 10x faster library characterization and reduced token costs by more than 10x.
Samsung achieved up to 20x greater performance for computational lithography using NVIDIA cuLitho and CUDA-X libraries, and applied PhysicsNeMo to chip-scale thermal-stress analysis with numerical solver accuracy across domains containing up to 10 billion cells.
ChipAgents is fine-tuning Nemotron models for chip design and verification workflows including debug, formal verification, and coverage.
Silvaco ran a 3.2-billion-mesh-node photonic edge coupler simulation in under four hours on 32 NVIDIA GPUs interconnected by NVIDIA NVLink. The workload exceeded the practical limits of CPU-based simulation.
Keysight is using NVIDIA cuDSS to speed up electromagnetic simulations by up to 10x. Samsung, Synopsys, and TSMC are integrating NVIDIA cuEST into GPU-accelerated pipelines for up to a 50x speedup on quantum chemistry workloads.
Source: NVIDIA
About NVIDIA
NVIDIA, founded in 1993 and headquartered in Santa Clara, CA, designs and manufactures graphics processing units, systems on chips, networking hardware, and AI intelligence software such as CUDA. Its products serve industries including gaming, data centers, autonomous vehicles, professional visualization, robotics, health care, and energy. The company introduced the GPU in 1999 and later expanded into accelerated computing and AI infrastructure. In gaming, its GPUs support high-performance rendering, while in AI and high-performance computing, its systems provide the infrastructure for training and deploying large-scale models. NVIDIA also develops tools for robotics and autonomous driving.