NVIDIA Partners with Industrial Software Giants to Accelerate Simulation and Chip Design with AI

Title

NVIDIA Partners with Industrial Software Giants to Accelerate Simulation and Chip Design with AI

Meta Description

NVIDIA partners with industrial software vendors to focus on accelerated simulation, open benchmarks and datasets for Physics AI, and agentic AI evaluation in CAE/EDA, driving AI adoption in industrial engineering.

Article Content

NVIDIA has made another move in the industrial engineering space. Recently, an industrial software vendor (hereinafter referred to as "the company") announced via social media a new partnership with NVIDIA aimed at deeply integrating AI into industrial engineering scenarios, with a focus on faster simulation speeds, open benchmarks and datasets for Physics AI, and an evaluation framework for agentic AI in CAE/EDA.

This collaboration directly addresses two core pain points in industrial R&D: the lengthy computation time of simulations and the exponentially rising complexity of chip design. By combining NVIDIA's accelerated computing and AI capabilities, the two parties aim to provide engineers with more efficient iteration tools.

Collaboration Focuses on Three Directions

According to public information, the partnership has identified three clear focus areas.

1. Accelerated Simulation: From "Waiting for Results" to "Real-Time Feedback"

Simulation runs through the entire lifecycle of industrial products, but traditional solvers often take hours or even days. The primary goal of the collaboration is to significantly shorten simulation cycles through AI surrogate models and GPU acceleration. This means engineers can quickly validate more design options early in the design phase, rather than being held back by computational resources.

AI surrogate models are the core technology behind this "accelerated simulation" approach. Rather than directly solving physical equations, they first train a machine learning model on large amounts of high-fidelity simulation results to learn the mapping from "input parameters → output field quantities." At inference time, the model can produce approximate results in milliseconds to seconds, replacing some expensive numerical solves. Surrogate models are typically used in conjunction with GPU-accelerated solvers: surrogate models for rapid screening during conceptual design, and high-fidelity solvers for verification during detailed design. Their accuracy depends on the coverage of the training data; extrapolation beyond the training distribution can significantly increase error, so practical engineering often requires uncertainty quantification and online correction mechanisms.

2. Open Benchmarks and Datasets: Establishing a "Yardstick" for Physics AI

Physics AI refers to combining physical laws with machine learning to predict behaviors such as fluid, structural, and electromagnetic. Currently, the field lacks unified evaluation standards, making it difficult to compare different models horizontally. The two parties plan to release open benchmarks and supporting datasets to provide researchers and developers with a reproducible testing environment. This move is expected to lower the barrier to entry for Physics AI and promote community collaboration.

Physics AI, unlike purely data-driven models, emphasizes embedding physical constraints such as mass conservation, momentum equations, and boundary conditions into machine learning models. Common approaches include physics-informed neural networks (PINNs), operator learning (such as Fourier Neural Operators), and treating numerical solvers as differentiable layers. The advantage of these methods is lower data dependence and stronger extrapolation capability, but training stability and computational cost remain challenges. One of the main obstacles in the field today is the lack of recognized benchmarks: different papers use self-built datasets and evaluation metrics, making results difficult to reproduce and compare. A high-quality open benchmark needs to cover typical geometries, boundary conditions, and operating condition distributions, and define unified error metrics (such as relative L2 error and conservation quantity deviation). This is exactly the value of the datasets the two parties plan to release.

3. Agentic AI Evaluation: Enabling Reliable Deployment of Agents in CAE/EDA

The potential of agentic AI in industrial software lies in automatically executing multi-step tasks, such as automatically setting simulation boundary conditions or optimizing chip layouts. But industrial scenarios demand extremely high reliability, and agent decisions must be evaluable and traceable. The collaboration will establish an agent evaluation framework for CAE (Computer-Aided Engineering) and EDA (Electronic Design Automation), clarifying their capability boundaries in real-world workflows.

Agentic AI refers to systems capable of autonomous planning, tool invocation, and execution of multi-step tasks, typically powered by large language models as the reasoning core, combined with function calling, memory, and reflection mechanisms. In CAE/EDA scenarios, an agent might be asked to "read the geometry model → generate mesh → set boundary conditions → submit for solving → check convergence → adjust parameters and re-run." Unlike general office scenarios, industrial agent evaluation cannot just look at whether a task is completed; it must also measure the correctness of numerical results, compliance with software API calls, recovery capability after failure, and auditability of the decision process. Therefore, evaluation frameworks typically need to build task sets with standard answers, define hierarchical scoring dimensions, and record every tool call in a sandbox environment—this is also the technical challenge of this collaboration direction.

Why NVIDIA and an Industrial Software Vendor?

NVIDIA's advantages in GPU acceleration and AI infrastructure need no elaboration, while industrial software vendors possess deep domain knowledge and user scenarios. The combination is not simply "AI + software" but an attempt to restructure the underlying workflow of industrial R&D.

Looking at industry trends, giants such as Siemens, Dassault, and Ansys are all advancing the integration of AI and simulation. NVIDIA has previously launched platforms for physics simulation such as Omniverse and Modulus. This collaboration further indicates that NVIDIA is penetrating the core toolchain of CAE/EDA with its CUDA ecosystem and AI frameworks by partnering with vertical players.

For industrial software vendors, leveraging NVIDIA's compute power and AI toolchain allows them to inject intelligent features into existing products faster while avoiding the high cost of building their own AI infrastructure.

The CUDA ecosystem is key to understanding NVIDIA's penetration into industrial software toolchains. Many core computations in CAE/EDA solvers (sparse matrix solving, FFT, particle/ray tracing) already have mature CUDA library support, such as cuSOLVER, cuSPARSE, and cuFFT; NVIDIA also provides Modulus (a physics machine learning framework) and Omniverse (a simulation and digital twin platform) as upper-layer tools. Industrial software vendors can migrate solvers to GPUs or call these libraries to achieve multiple-fold acceleration without changing the user interface, but this also deepens dependence on specific hardware and software stacks. This binding of "domain knowledge + accelerated computing stack" is a typical strategy for NVIDIA to expand ecosystem coverage by partnering with vertical players.

Practical Impact on Engineers and Enterprises

In the short term, engineers may first feel the improvement in simulation speed. AI surrogate models can replace some high-fidelity computations while maintaining accuracy, especially suitable for rapid exploration during conceptual design.

In the medium term, open benchmarks and datasets will lower the development barrier for Physics AI models. Small and medium-sized enterprises and research teams can validate their algorithms against unified standards without building test environments from scratch.

In the long term, once the evaluation system for agentic AI in CAE/EDA matures, it is expected to drive "automation of automation"—AI assisting in the setup and optimization of simulation workflows, allowing engineers to focus on higher-value decisions.

Challenges and Uncertainties

The collaboration is still in its early stages, and the timing of specific product launches, the coverage of open-source benchmarks, and industry acceptance of agent evaluation all remain to be seen. The stringent demands for accuracy and reliability in industrial scenarios mean AI models must undergo rigorous validation before entering production workflows. In addition, maintaining open benchmarks requires sustained investment, and ecosystem building is not achieved overnight.

However, the direction is clear: the combination of AI and industrial engineering is moving from proof of concept to the infrastructure layer. This collaboration between NVIDIA and an industrial software vendor provides a noteworthy example of scaled application of Physics AI and agents in CAE/EDA.

One-line Summary

NVIDIA partners with an industrial software vendor to accelerate simulation with AI and establish evaluation standards for Physics AI and agents.

Paragraph Summary

NVIDIA and an industrial software vendor announced a collaboration focused on AI adoption in industrial engineering scenarios, defining three directions: first, shortening simulation cycles through AI surrogate models and GPU acceleration, enabling engineers to iterate quickly in early design; second, releasing open benchmarks and supporting datasets for Physics AI to provide a reproducible unified evaluation environment for prediction models in fluid, structural, and electromagnetic domains; third, establishing an evaluation framework for agentic AI in CAE/EDA to clarify capability boundaries in high-reliability industrial workflows. The article argues that this collaboration combines NVIDIA's advantages in accelerated computing and AI infrastructure with the domain knowledge of industrial software vendors, aligning with the trend of Siemens, Dassault, and Ansys advancing AI-simulation integration. In the short term it improves simulation speed, in the medium term it lowers the development barrier for Physics AI, and in the long term it drives automation of simulation workflows. However, product launch timing, benchmark coverage, and industry acceptance remain to be seen.