NVIDIA Accelerates Chip Design with In-House CPUs, Marking an AI Turning Point for EDA
0xBroomberg
Nvidia announced at DAC 2026 that it is using its in-house Vera CPU and AI agent tools to reshape chip design workflows, with early benchmarks showing up to 15× speedups. This means → EDA, the foundational infrastructure of the chip industry, is being pulled into Nvidia's hardware ecosystem.
What is the Vera CPU doing here?
Nvidia is running chip-design software on its in-house Vera CPU, targeting EDA — electronic design automation, the software stack that takes a chip from schematic to verified design.
Cadence's formal-verification tool Jasper and Synopsys's logic-simulation tool VCS both showed 1.5× performance gains on Vera.
This means → Nvidia is no longer just a chipmaker. It is using its own chips to design the next generation of chips — a self-accelerating loop.
How are AI agents entering chip design?
Nvidia plugged its physics-AI library PhysicsNeMo and its GPU math libraries CUDA-X into the Nvidia Agent Toolkit, letting AI agents — autonomous programs that call tools and execute multi-step tasks — invoke simulation solvers the way a human engineer would.
CUDA-X added three new libraries: cuISS (iterative solver), cuDSS (circuit-simulation solver), and cuEST (quantum-chemistry simulation) — the first time iterative sparse solvers run on GPUs.
In plain terms = engineers used to write code and run simulations manually. Now AI agents call the tools, run the math, and deliver results. Humans make the decisions.
What numbers did partners report?
Keysight Technologies used the new cuDSS library and saw electromagnetic simulation speed up by as much as 10×.
Silvaco ran a 3.2-billion-node photonic edge-coupler simulation on a 32-GPU cluster in under four hours.
Cadence's AI super-agent AuraStack, running cuDSS on its Millennium M2000 supercomputer, achieved a 15× speedup in design-verification workflows.
This means → the gains are not Nvidia's claim alone. Third-party benchmarks land in the 10–15× range, signaling that GPU-accelerated EDA has moved past proof-of-concept.
The software is free — so what's the catch?
PhysicsNeMo ships under the Apache 2.0 open-source license. The new CUDA-X libraries are free and drop in as replacements for code engineers typically write by hand.
But all of it runs only on Nvidia hardware.
In plain terms = the software costs nothing; the hardware bill is where Nvidia collects. Same logic as cheap printers, expensive ink. This reflects Nvidia's real play: locking EDA workloads into its own ecosystem.
Where does this lead next?
Nvidia is already using Vera to accelerate the design of Rosa, its next-generation CPU built on the new Rigel core.
That creates a chain: current chips speed up design tools → faster tools speed up next-gen chip design → next-gen chips speed up even newer tools.
This means → if the loop holds, Nvidia's chip-iteration pace becomes tied to its design-tool performance — each generation designed faster than the last. Whether the EDA toolchain can deliver similar gains across broader workflows is the key test for this strategy.
Content is for reference only, not financial advice.