News
NVIDIA says NeMo Gym helped Siemens improve a chip-design agent’s Liberty-file benchmarking score from about 55% to 95%.
NVIDIA said its NeMo Gym environment was used to help improve the reliability of a Siemens Fuse EDA AI agent for standard-cell characterization, with the measured benchmark composite score rising from about 55% to roughly 95% after changes to the agent’s skills and tool descriptions. The company said the work mattered because agentic workflows in chip design can produce plausible but incorrect artifacts, allowing early errors to propagate into final outputs.
Evidence and context
In the Oct. 14, 2026 post, NVIDIA described NeMo Gym as an open library for evaluating and improving models and agents through environments, and said it was applied to a Solido Characterization Suite agent that orchestrates characterization tools through Siemens Fuse EDA AI Agent. The workflow covered tool setup, SPICE simulation, distributed compute, Liberty file generation, debugging and final validation, with the post emphasizing that end-to-end checks were necessary because intermediate tool calls could appear successful even when final artifacts were wrong.
NVIDIA said the Siemens team used a domain-specific verifier to inspect Liberty (.lib) files against known references and trace failures through rollout trajectories. The verifier assessed coverage, validity, relative validation, accuracy and integrity, and returned zero if any gate failed. The company said this approach exposed missing cells, missing operating conditions, incorrect slew thresholds and timing or power values outside tolerance that intermediate checks did not catch.
For benchmarking, NVIDIA said the team used the open source GF180MCU PDK and evaluated Liberty-file generation across 229 standard cells and 36 PVT corners, though each run was restricted to two cells across the same 36 corners for tuning. The company said it conducted about 25 rollouts using Claude Opus 4.8 in Claude Code as the fixed model and harness. After simplifying overlapping file and shell interfaces and adding explicit JSON schemas for tool inputs, NVIDIA said the mean composite score increased to about 95%, while agent turns fell from 155 to 53 and output tokens declined from about 81,000 to 29,000.
The post also placed the work in the context of recent NeMo Gym releases, including v0.7.0 on Oct. 6, 2026, which added reusable Environment Servers, larger-scale evaluation support, token-level training capture and OpenSandbox pause/resume support. NVIDIA said the same evaluation method could be applied to other engineering domains, but the article did not provide a next-date catalyst for the Siemens effort beyond the publication of the post itself.
Be the first to comment