Search Ctrl+K
Change language Switch ThemeSign In
[](https://www.bestblogs.dev/en/explore/brief "Daily Brief")[](https://www.bestblogs.dev/en/explore/newsletter "Weekly Picks")[](https://www.bestblogs.dev/en/explore/topics "Topics")
[](https://www.bestblogs.dev/en/settings "Settings")[](https://www.bestblogs.dev/en/docs "Help Center")
Narrow Mode
78
英伟达大举进军 EDA
NVIDIA expands its Agent Toolkit with PhysicsNeMo and CUDA-X libraries, driving AI assistants from generic helpers toward autonomous engineering systems to enable intelligent semiconductor EDA design. Follow·
Today·1980 words (about 8 min)
AI Summary & Key Points
Summary
The article describes how NVIDIA expands the NVIDIA Agent Toolkit using the open‑source PhysicsNeMo framework and the CUDA‑X acceleration libraries (including cuISS, cuDSS, cuEST, etc.), enabling AI assistants to connect to physical models, numerical solvers, and simulation environments. This creates an ‘intelligent engineer’ capable of autonomously devising technical plans, invoking specialized software, evaluating results, and iterating optimizations. The piece further details the workflow of this architecture in thermal‑design scenarios and its integration with Nemotron 3 Ultra and ACE‑RTL in semiconductor design, emphasizing that the technology’s value hinges on strict governance and auditability.
Main Points
* 1. NVIDIA extends the Agent Toolkit with PhysicsNeMo and CUDA‑X, enabling AI to evolve from a general‑purpose assistant into an autonomous engineering system.
PhysicsNeMo offers trainable physics‑informed neural network surrogate models, while CUDA‑X provides high‑performance numerical libraries such as sparse solvers, allowing the agent to invoke simulation and solvers for plan generation, evaluation, and iteration.
* 2. The architecture separates reasoning from computation: the language model interprets engineering goals, maintains workflow state, and selects tools; PhysicsNeMo supplies fast approximations; CUDA‑X delivers high‑fidelity numerical solutions; and domain software imposes constraints and sign‑off rules.
This separation creates a hypothesis‑simulation‑measurement‑optimization‑verification loop; taking a thermal‑design agent as an example, it shows how the system can autonomously generate cooling geometries, rank surrogate models, run high‑fidelity simulations, and automatically adjust the design.
* 3. In the semiconductor domain, combining Nemotron 3 Ultra with ACE‑RTL enables agent‑driven RTL coding, with Cadence, Synopsys, Siemens, ChipAgents and other vendors integrating the technology, promising shorter iteration cycles and better utilization of simulation resources.
This combination can generate, debug, and verify RTL code while protecting IP, and industry collaborations show significant performance gains in multiphysics acceleration, computational lithography, and related areas.
* 4. The technology’s value depends on governance; the agent must disclose assumptions, maintain version control, quantify uncertainty, enforce execution limits, and undergo independent verification before manufacturing.
Only within an auditable and verifiable scope is the agent’s speed beneficial; otherwise its fluent explanations may mask non‑convergent or erroneous outcomes, posing design risks.
Sign in to highlight text and take notes as you read. Sign in now
半导体行业观察 2026-08-11 08:45 安徽
半导体行业观察：NVIDIA 通过 PhysicsNeMo 和 CUDA-X 库扩展了 NVIDIA Agent Toolkit，这标志着人工智能助手的发展方向从通用型人工智能助手转向自主工程系统。
公众号记得加星标⭐️，第一时间看推送不会错过。
NVIDIA 通过 PhysicsNeMo 和 CUDA-X 库扩展了 NVIDIA Agent Toolkit，这标志着人工智能助手的发展方向从通用型人工智能助手转向自主工程系统。扩展后的工具包不再局限于文档检索、代码生成或工作流程自动化，而是允许开发者将推理模型直接连接到物理模型、数值求解器、仿真环境和电子设计工具。其目标是打造一个“智能工程师”，能够制定技术方案、调用专用软件、评估结果并迭代优化，最终形成经过验证的设计。
PhysicsNeMo 提供物理人工智能层。这个开源框架支持构建、训练、微调和部署结合仿真数据和物理约束的科学机器学习模型。其架构包括神经网络算子、图神经网络、点云模型、物理信息神经网络和生成模型。这些方法可以作为计算流体动力学、结构力学、热分析、电磁学和半导体制造等领域中计算量巨大的仿真的替代模型。在代理工作流程中，训练好的替代模型就成为一个可调用的工具：代理可以快速估计设计的行为，识别有希望的候选方案，并将高保真求解器预留用于最终验证。
CUDA-X 提供了加速数值计算的基础架构。NVIDIA 的全新 cuISS 库为由离散化偏微分方程产生的大型系统提供了可组合的迭代稀疏求解器和预处理器。cuDSS 提供直接稀疏求解器，专为数值鲁棒的器件、电路、系统和科学模拟而设计，并具备多 GPU 和多节点可扩展性。cuEST 将计算栈扩展到电子结构计算，包括密度泛函理论和后 DFT 方法。这些库共同使代理能够做到更多，而不仅仅是推荐参数：它可以启动 GPU 加速的计算，其输出基于成熟的数值方法。
由此产生的架构将推理与计算分离。语言或推理模型负责解释工程目标、维护工作流程状态并选择工具。PhysicsNeMo 模型提供快速学习的近似值。CUDA-X 求解器生成更高保真度的数值结果，而领域软件则应用约束和签核规则。因此，智能体可以经历假设、仿真、测量、优化和验证的循环。例如，热设计智能体可以生成冷却几何体，使用代理模型对其进行排序，对最终入选方案运行详细仿真，检查热点和压力损失，然后自动修改几何体。
NVIDIA 也将目光投向了半导体设计领域。Nemotron 3 Ultra 与 NVIDIA 研究院的 ACE-RTL 代理相结合，旨在实现代理式寄存器传输级编码。NVIDIA 表示，该模型在全面的 Verilog 设计基准测试中领先于其他开放模型，并且可以基于专有数据进行后训练，用于本地或企业内部部署。这种组合至关重要，因为芯片开发代理必须在生成 RTL 代码、调试故障、运行验证工具以及保持可追溯性的同时，处理敏感的知识产权。
各大工程软件供应商正在整合技术栈的各个部分。NVIDIA 报告称，Cadence正在将加速计算和智能体系统应用于封装和 PCB 设计；Synopsys正在开发自主热优化和验证工作流程；而Siemens 则正在协调半导体、3D-IC、PCB 和系统设计领域的多种工具和智能体。据报道，Siemens 的库表征速度提升了十倍以上，Cadence 的多物理场加速提升了二十倍，计算光刻、电磁仿真和量子化学方面也取得了显著进步。ChipAgents 正在 使用 NVIDIA Agent Toolkit 构建用于芯片设计和验证的特定领域 AI 智能体。该团队正在对 NVIDIA Nemotron 模型进行微调，以适应复杂的端到端半导体设计和验证工作流程，包括调试、形式验证、覆盖率测试等等。
实际结果可能是缩短迭代周期，更广泛地探索替代方案，以及在全球分布式工程组织中更好地利用昂贵的仿真基础设施。
然而，技术价值取决于治理。工程代理必须公开假设、保存输入和求解器版本、量化不确定性、强制执行运行限制，并在制造决策前要求进行独立验证。代理的速度仅在其已验证的范围内有用，而代理流畅的解释不能替代收敛性测试、物理签核或专家评审。
总而言之， NVIDIA 的战略远不止于添加另一个 AI 界面。它正在将加速计算库和物理模型转化为标准化的智能体技能。如果这些技能保持互操作性、可审计性，并与经过验证的求解器紧密耦合，智能体就能将设计空间探索从手动协调的序列压缩成一个连续的计算过程。这将改变团队的产品设计方式：人类负责定义目标、约束和验收标准，而自主系统则负责执行和优化大部分仿真和优化循环。
（来源：编译自semiwiki）
*免责声明：本文由作者原创。文章内容系作者个人观点，半导体行业观察转载仅为了传达一种不同的观点，不代表半导体行业观察对该观点赞同或支持，如果有任何异议，欢迎联系半导体行业观察。
_END_ 今天是《半导体行业观察》为您分享的第 4495 内容，欢迎关注。 推荐阅读
加星标⭐️第一时间看推送
求点赞
求分享
求推荐
Key Quotes
> NVIDIA has expanded the NVIDIA Agent Toolkit with the PhysicsNeMo and CUDA‑X libraries, marking a shift in the direction of AI assistants from general‑purpose helpers to autonomous engineering systems.
> The resulting architecture separates reasoning from computation. The language or reasoning model interprets engineering objectives, maintains workflow state, and selects tools. PhysicsNeMo models provide fast‑learned approximations. CUDA‑X solvers produce higher‑fidelity numerical results, while domain software applies constraints and sign‑off rules.
> However, the technology’s value hinges on governance. Engineering agents must disclose assumptions, retain input and solver versions, quantify uncertainty, enforce execution limits, and require independent verification before manufacturing decisions.
Tags
AI Agent
PhysicsNeMo
CUDA-X
Semiconductor Design
Simulation Acceleration
Related Articles
* Kimi K3: The New Frontier of Intelligence * One AI Is Not Enough * OpenClaw and Hermes: A Comprehensive Review of AI Agent Architecture from Source Code * Broadcom Told the Truth, But No One Noticed * Nvidia Buys 5 Million Optical Modules a Year, Then Dismantles Them Itself * Advanced Packaging: Entering a New Battlefield, exploring its advantages in addressing cost and size bottlenecks, TSMC and domestic vendors' layout, and the process challenges faced.") * The next bottleneck for AI may not be GPU and storage, but semiconductor equipment. * Rethinking Development Infrastructure: When Agents Become First-Class Citizens * The Missing Link in Co-Packaged Optics, emphasizing that merely improving photonic device performance cannot achieve scale; instead, package PDK must enable multi-domain system-level integration—covering optoelectrical, thermal, mechanical, and manufacturing aspects—and ensure manufacturability.") * Redefining Skill Development: A Step-by-Step Guide & One-Stop Development Assistant Launch
Make your daily reading actually fit you.A daily brief built from the sources you follow. Get started free HomeDiscoverSettings