NVIDIA has extended the NVIDIA Agent Toolkit with the PhysicsNeMo and CUDA-X libraries, marking a shift in the development direction of AI assistants from general-purpose AI helpers to autonomous engineering systems. The expanded toolkit is no longer limited to document retrieval, code generation, or workflow automation, but allows developers to directly connect inference models to physical models, numerical solvers, simulation environments, and electronic design tools. 

Its goal is to create an "intelligent engineer" capable of developing technical solutions, calling specialized software, evaluating results, and iteratively optimizing to ultimately create a validated design.

Physical Artificial Intelligence Layer

PhysicsNeMo provides a physical AI layer. This open-source framework supports building, training, fine-tuning, and deploying scientific machine learning models that combine simulation data with physical constraints. Its architecture includes neural network operators, graph neural networks, point cloud models, physically-informed neural networks, and generative models. These methods can serve as alternative models for computationally intensive simulations in fields such as computational fluid dynamics, structural mechanics, thermal analysis, electromagnetics, and semiconductor manufacturing. In the agent workflow, the trained alternative model becomes a callable tool: the agent can quickly estimate the behavior of the design, identify promising candidates, and reserve high-fidelity solvers for final validation.

CUDA-X provides the infrastructure for accelerating numerical computation. NVIDIA's new cuISS library provides composable iterative sparse solvers and preprocessors for large systems generated from discretized partial differential equations. cuDSS provides direct sparse solvers designed for numerically robust devices, circuits, systems, and scientific simulations, and is scalable across multiple GPUs and nodes. cuEST extends the computation stack to electronic structure computation, including density functional theory and post-DFT methods. These libraries collectively enable agents to do much more than just recommend parameters: they can initiate GPU-accelerated computations whose output is based on well-established numerical methods.

The resulting architecture separates inference from computation. A language or inference model is responsible for interpreting engineering goals, maintaining workflow state, and selecting tools. The PhysicsNeMo model provides rapidly learned approximations. 

The CUDA-X solver generates higher-fidelity numerical results, while domain software applies constraints and signature rules. Thus, the agent can go through a cycle of hypothesis, simulation, measurement, optimization, and validation. For example, a thermal design agent can generate cooling geometries, rank them using an agent model, run detailed simulations of the final selected solutions, check for hot spots and pressure losses, and then automatically modify the geometry.

Semiconductor design

NVIDIA has also set its sights on the semiconductor design field. The Nemotron 3 Ultra, combined with NVIDIA Research's ACE-RTL agent, aims to achieve agent-based register-transfer level coding. NVIDIA states that this model outperforms other open models in comprehensive Verilog design benchmarks and can be post-trained on proprietary data for local or enterprise deployment. This combination is crucial because chip development agents must handle sensitive intellectual property while generating RTL code, debugging, running verification tools, and maintaining traceability.

Major engineering software vendors are integrating various parts of their technology stacks. NVIDIA reports that Cadence is applying accelerated computing and agent systems to packaging and PCB design; Synopsys is developing autonomous thermal optimization and verification workflows; and Siemens is coordinating multiple tools and agents across semiconductor, 3D-IC, PCB, and system design. Siemens has reportedly achieved more than ten times faster library characterization, Cadence twenty times faster multiphysics acceleration, and significant advancements in computational lithography, electromagnetic simulation, and quantum chemistry. ChipAgents is using the NVIDIA Agent Toolkit to build domain-specific AI agents for chip design and verification. The team is fine-tuning NVIDIA Nemotron models to accommodate complex end-to-end semiconductor design and verification workflows, including debugging, formal verification, coverage testing, and more.

Practical results could include shorter iteration cycles, broader exploration of alternatives, and better utilization of expensive simulation infrastructure in globally distributed engineering organizations.

However, the technological value hinges on governance. Engineering agents must expose assumptions, store input and solver versions, quantify uncertainties, enforce runtime constraints, and require independent verification before manufacturing decisions are made. An agent's speed is only useful within its validated scope, and a fluent interpretation by the agent cannot replace convergence testing, physical sign-off, or expert review.

In short, NVIDIA's strategy goes far beyond adding another AI interface. It is transforming accelerated computation libraries and physical models into standardized agent skills. If these skills remain interoperable, auditable, and tightly coupled with validated solvers, agents can compress design space exploration from a manually coordinated sequence into a continuous computational process. This will change the way teams design products: humans will be responsible for defining goals, constraints, and acceptance criteria, while autonomous systems will be responsible for executing and optimizing most of the simulation and optimization loops.

Source: Compiled from semiwiki

View more at EASELINK