NVIDIA NeMo Gym cuts Siemens EDA chip-design agent turns 66%, lifts score to 95%
NVIDIA published a technical blog describing how Siemens EDA used the open evaluation library NVIDIA NeMo Gym to improve the reliability of its chip-design agent. The agent handles Liberty (.lib) file generation for standard-cell characterization; a domain-specific verifier scores final artifacts on coverage, validity, relative validation, accuracy and integrity, awarding zero reward if any gate fails. On a benchmark based on the open GF180MCU PDK — 229 standard cells across 36 PVT corners, with each tuning run restricted to two cells — the composite verification score rose from roughly 55% to about 95%. Agent turns fell from 155 to 53 (about 66% fewer) and output tokens from roughly 81K to 29K (about 64% fewer), after the team simplified overlapping file and shell interfaces and defined explicit JSON schemas for Model Context Protocol tool inputs based on verifier diagnostics. The runs used Claude Opus 4.8 in Claude Code as the fixed model and agent harness.
It shows that grading final engineering artifacts rather than individual tool calls, then refining skills and tool descriptions, can improve both reliability and efficiency for engineering agents.
Optimizing Agent Reliability and Tool Calling Efficiency for Chip Design Using NVIDIA NeMo Gym
Siemens EDA used agentic evaluations to cut agent turns by 66% while raising benchmark scores from 55% to 95%.
Oct 14, 2026
By Joshua Mabry, Chris Wing, Niranjan Sitapure, Kenton DeLathouwer and Ming Ming Zhang
#NVIDIA#NVIDIA NeMo Gym#Siemens EDA#Siemens Fuse EDA AI Agent#Solido Characterization Suite#Claude Opus 4.8#Claude Code#Model Context Protocol#GF180MCU PDK