Autonomous design agents are becoming the newest workload for GPU data centers, and NVIDIA is giving them physics-grade tools.
The company expanded its Agent Toolkit with a reworked PhysicsNeMo and fresh CUDA-X libraries aimed squarely at engineering work. The additions include sparse solvers for massive simulations and cuEST, a quantum chemistry library, letting agents train physics models and run them as part of normal design flows.
The pitch is straightforward: instead of a human expert driving each simulation step, an agent with the right libraries can compose and validate the work. NVIDIA claims the approach can deliver validated RTL designs 50 times faster than alternative stacks.
Adoption is the real story. Cadence, Synopsys and Siemens are wiring the toolkit into their flagship agent products, while Samsung, TSMC and Keysight are testing the new solver libraries in production pipelines. Nemotron 3 Ultra, NVIDIA’s open model, tops open-model leaderboards for agentic RTL coding.
Announced around the Design Automation Conference, the move signals that agentic AI has crossed from code generation into physical engineering. EDA vendors standardizing on one GPU-centric stack also tightens NVIDIA’s grip on the design ecosystem, making its hardware the default compute layer for tomorrow’s chip tooling.
