NVIDIA Expands Agent Toolkit With PhysicsNeMo and CUDA-X Libraries for Engineering

What NVIDIA Added to the Agent Toolkit

NVIDIA announced an expansion of the NVIDIA Agent Toolkit for engineering, adding NVIDIA PhysicsNeMo and CUDA-X libraries as agent-ready tools and skills. According to NVIDIA's newsroom, the toolkit is "built to transform how the world designs and develops products" (E1). A second NVIDIA newsroom posting of the same announcement confirms the identical wording and scope of the expansion (E12).

Why NVIDIA Says Engineering Has Reached an "Inflection Point"

Explaining the rationale behind the expansion, Timothy Costa, vice president and general manager of computational engineering at NVIDIA, said: "Engineering has reached an inflection point. AI can now work with tools of physics, simulation and design" (E2). NVIDIA frames the PhysicsNeMo and CUDA-X additions (E1) as the toolset that puts that stated inflection point into practice, giving agents access to physics, simulation and design capabilities rather than general-purpose functions alone.

Three New CUDA-X Libraries: cuISS, cuDSS and cuEST

NVIDIA detailed three new libraries added to the CUDA-X family:

• cuISS (CUDA Iterative Sparse Solvers): NVIDIA states this library "accelerates large sparse linear systems in physics-based and engineering simulations" (E3).
• cuDSS (CUDA Direct Sparse Solvers): according to NVIDIA, cuDSS "accelerates large, complex sparse linear systems central to electronic design automation (EDA) and scientific simulation" (E4).
• cuEST (CUDA Electronic Structure Theory): NVIDIA says cuEST "brings high-accuracy quantum chemistry simulations to device-relevant scales," enabling density functional theory (DFT) and post-DFT methods to run in production workflows at scale (E5).

Each library targets a distinct engineering bottleneck — physics/engineering simulation for cuISS, EDA and scientific simulation for cuDSS, and device-scale quantum chemistry for cuEST — per NVIDIA's own descriptions.

Nemotron 3 Ultra's Position in Agentic RTL Coding

NVIDIA also highlighted NVIDIA Nemotron 3 Ultra's role in chip-design workflows. Paired with ACE-RTL, described by NVIDIA as "an agent for designing hardware from NVIDIA Research," NVIDIA states that Nemotron 3 Ultra "leads among open models in agentic RTL coding on the comprehensive verilog design problems benchmark across RTL coding tasks" (E6). The same claim appears verbatim in the duplicate NVIDIA newsroom posting (E13), indicating NVIDIA is presenting this as a consistent, repeated claim about the model's benchmark standing among open models — not an independently verified third-party result.

Partner Results: Reported Speedups Across the Industry

NVIDIA's announcement cites specific performance figures reported by five industry partners using the new tools. According to NVIDIA, Cadence is using NVIDIA Nemotron, accelerated computing and CUDA-X libraries with its AuraStack AI Super Agent and Millennium M2000 platform to "autonomously drive advanced packaging and printed circuit board (PCB) design from exploration through signoff, delivering up to 20x faster multiphysics performance" (E7). Siemens reports that agentic AI workflows in its Solido Characterization Suite are "delivering more than 10x faster library characterization while reducing token costs by more than 10x" (E8).

Samsung reports two separate results: up to 20x greater performance for computational lithography using cuLitho and CUDA-X libraries, and chip-scale thermal-stress analysis using PhysicsNeMo across domains containing up to 10 billion mesh cells (E9). Silvaco, running on 32 NVIDIA GPUs interconnected via NVLink, "completed a 3.2-billion-mesh-node photonic edge coupler simulation in under four hours," which NVIDIA describes as "a workload beyond the practical limits of CPU-based simulation" (E10). Keysight reports accelerating electromagnetic simulations by up to 10x using cuDSS, while Samsung, Synopsys and TSMC (台積電) report integrating cuEST into GPU-accelerated pipelines to achieve up to a 50x speedup for quantum-chemistry workloads (E11). The Keysight/Samsung/Synopsys/TSMC figures are repeated identically in a duplicate NVIDIA newsroom posting (E15), as are the Cadence figures (E14).

What This Means

Across the figures NVIDIA cites, the reported speedups cluster around two bands: 10x-to-20x for packaging, PCB, lithography and library-characterization workloads (Cadence, Siemens, Samsung), and up to 50x for the cuEST quantum-chemistry pipeline used by Samsung, Synopsys and TSMC — the single largest multiple in the announcement. Two different sparse-solver libraries appear in separate contexts: cuDSS is positioned by NVIDIA for EDA and scientific simulation (E4) and is the library Keysight names for its 10x electromagnetic-simulation result (E11), while cuISS is positioned for physics-based and engineering simulation generally (E3). Notably, all figures in this announcement — from the Nemotron 3 Ultra benchmark claim (E6, E13) to the partner speedups (E7-E11, E14, E15) — originate from NVIDIA's own newsroom rather than from independent benchmarking published by Cadence, Siemens, Samsung, Silvaco, Keysight, Synopsys or TSMC directly.