NVIDIA adds physics and CUDA-X libraries to Agent Toolkit for engineering AI agents

NVIDIA has expanded its Agent Toolkit for engineering with PhysicsNeMo and updated CUDA-X libraries, positioning the package as a way for developers to build specialized AI agents that can work with simulation, physics models and accelerated computing tools. The company announced the update on July 26, framing it around chip design, verification, packaging, systems engineering and scientific simulation workflows.

The core change is that PhysicsNeMo has been reworked into agent-ready libraries, while CUDA-X additions give agents access to accelerated solvers and quantum chemistry capabilities. In practice, NVIDIA says developers can connect engineering assistants to domain tools and data so the agents can train or deploy AI physics models, call sparse solvers, and handle simulation tasks that would otherwise require multiple specialist systems.

Why it matters

The announcement is another sign that AI agent work is moving beyond general productivity assistants and into technical design environments where tool access, data control and repeatable workflows matter. NVIDIA is aiming the toolkit at engineering teams that already rely on high-performance computing and electronic design automation, not at consumer chatbot use cases.

The company listed several new or updated components. NVIDIA cuISS is described as an iterative sparse solver library for large linear systems in engineering simulations. cuDSS targets direct sparse solving for workloads such as device, circuit and system simulations. cuEST brings electronic structure theory tools for quantum chemistry simulations into GPU-accelerated production workflows. Those libraries are being presented as callable capabilities for agentic engineering systems rather than standalone utilities.

NVIDIA also tied the release to its Nemotron 3 Ultra open model and ACE-RTL, an NVIDIA Research agent for hardware design. The company says Nemotron 3 Ultra leads among open models on agentic register-transfer-level coding tasks using a Verilog design benchmark. That claim matters because chip design agents have to operate in a narrow technical domain where syntax errors and incorrect assumptions can become expensive quickly.

Industry use

NVIDIA named Cadence, Synopsys, Siemens, Samsung, Silvaco, Keysight and others as companies using or integrating parts of the stack. Reported examples include Cadence using Nemotron, accelerated computing and CUDA-X with its AuraStack AI Super Agent and Millennium M2000 platform; Synopsys using Agent Toolkit, NIM microservices and Nemotron models with AgentEngineer; and Siemens using NeMo Gym, Nemotron and CUDA-X libraries in EDA agent workflows.

Some of the performance claims are substantial but should be read as vendor-reported results. NVIDIA says Cadence is seeing up to 20x faster multiphysics performance, Siemens is reporting more than 10x faster library characterization with more than 10x lower token costs, and Keysight is using cuDSS to accelerate electromagnetic simulations by up to 10x. NVIDIA also says Samsung is applying PhysicsNeMo to chip-scale thermal-stress analysis across domains containing up to 10 billion cells.

The release does not make every feature generally available immediately. NVIDIA cautions that many products and features described remain in various stages and will be offered only when and if available. Even with that caveat, the update shows how the company is packaging its GPU software stack for a world where AI agents increasingly operate inside engineering pipelines rather than outside them.

Sources

• nvidianews.nvidia.com

Cover photo by Tima Miroshnichenko on Pexels, used under the Pexels License.