NVIDIA says Nemotron 3 Ultra hit 97.1% on agentic RTL chip-design tasks

NVIDIAAI · x · 2026-07-27

NVIDIA says it tested Nemotron 3 Ultra on agentic chip design, where a model iteratively writes RTL code, runs a simulator, reads failures, and rewrites.

Across nine categories of real design work, the model averaged a 97.1% pass rate while using 6,629 tokens per iteration, which NVIDIA says is the best result among the open models it tested on both metrics. The image also shows a synthetic data pipeline built from public RTL snippets and in-context <Spec, RTL> examples, followed by syntax checks, benchmark decontamination, and rubric-based filtering.

Related event: NVIDIA Claims Nemotron 3 Ultra Hits 97.1% in Chip Design(2 posts)→

Original post →

More from coding & agent

coding & agent channel →