NVIDIA has folded PhysicsNeMo physics-AI libraries and GPU math libraries into the NVIDIA Agent Toolkit, announced July 26 from the Design Automation Conference in Long Beach. The move makes accelerated sparse linear-algebra solvers callable directly by autonomous design agents—no separate toolchain, no context switch.

Three New Solver Libraries Target Engineering Workloads

Sparse linear algebra underlies nearly every physical simulation engineers run: fluid flow, structural stress, electromagnetics, thermal analysis. NVIDIA now ships three libraries:

  • cuISS: iterative sparse solvers
  • cuDSS: direct sparse solvers for circuit and device simulation
  • cuEST: quantum-chemistry methods predicting material behavior at atomic scale

All are free, open-source under Apache 2.0 (PhysicsNeMo is built on PyTorch, CUDA-X libraries are drop-in replacements for hand-tuned code), and available in containers at no cost. The catch: they run nowhere except NVIDIA GPUs.

What's Already Running

Partner deployments are live or in advanced stages:

  • Samsung reports up to 20× faster computational lithography (cuLitho); applying PhysicsNeMo to thermal-stress problems spanning 10 billion cells per simulation.
  • Keysight is using cuDSS to accelerate electromagnetic simulation by up to 10×.
  • Silvaco ran a 3.2-billion-mesh-node photonic edge coupler simulation in under four hours across 32 GPUs with NVLink—work NVIDIA says CPU-based simulation could not practically handle.
  • TSMC, Samsung, Synopsys are integrating cuEST for what NVIDIA claims is up to 50× speedup on quantum-chemistry workloads.

Cadence supplies the clearest picture of the go-to-market. Its AuraStack AI Super Agent runs cuDSS on the Millennium M2000, a Blackwell-based multiphysics supercomputer Cadence sells. Cadence claims 15× faster design workflows and 2× faster time to market on top of the 20× multiphysics figure. The addressable market is vast: verification alone consumes billions of compute hours annually across the industry.

The Agent Distribution Play

The libraries sit at the bottom of agent stacks built by Cadence, Siemens, and Synopsys—each adding their own agents on top of NVIDIA's Agent Toolkit runtime. This is a lock-in strategy: whichever design software vendor wins a customer's flow, the solver underneath is the same, running on NVIDIA silicon. Combined with NVIDIA's push to place its Vera CPU in verification workloads that refuse to run on GPUs, the strategy covers both halves of a chip company's compute bill—simulation on accelerators, single-threaded verification on CPU.

What's Missing: Baseline and Timeline

Critical caveats:

  • cuISS has no ship date. No landing page exists; Synopsys is described as still developing use cases rather than running it in production.
  • No disclosed baselines. Every speedup claim is a vendor number with no comparison point (core count, machine specs, software version withheld). MLCommons publishes audited results for AI training and inference; nothing comparable exists for CAE, so only customers running their own designs through their own flows can validate cost/benefit.
  • Safe-harbor language. Many described products "remain in various stages and will be offered on a when-and-if-available basis."

Why It Matters for You

If you build agents that call external tools, this is the pattern: NVIDIA is embedding domain solvers into the agent platform so tool calls become first-class. If you architect EDA or multiphysics workflows, the bet is that your simulation farm CPU capacity gets retired and replaced with GPU clusters running these libraries. The signal to watch is not the next speedup claim at a trade show, but whether engineering organizations actually stop buying CPU capacity for simulation.

The libraries are free and open-source; the economics that matter—GPU acquisition cost, NVLink switching latency, power consumption at scale—remain vendor lock. PhysicsNeMo and CUDA-X lower the switching cost for CAE vendors to integrate with NVIDIA silicon, which is the entire point.

Sources