Siemens Uses NVIDIA NeMo Gym to Boost Chip AI Accuracy

Siemens EDA has applied NVIDIA NeMo Gym to optimize its Fuse EDA AI Agent, significantly enhancing how the tool manages standard-cell characterization for microchips. By systematically evaluating performance across full tasks rather than isolated intermediate actions, the team raised the agent's benchmark composite score from roughly 55% to 95%. The optimization process relied on Claude Opus 4.8 executed within the Claude Code harness, driving the Solido Characterization Suite agent through complex electronic design workflows.

The evaluations focused on Liberty file generation using the open-source GF180MCU process design kit, which features 229 standard cells across 36 process, voltage, and temperature (PVT) corners. To streamline tuning, Siemens restricted each rollout to two cells across the 36 PVT corners, running about 25 rollouts that each lasted between 1.5 and 2 hours. By refining the agent's skills and clarifying its Model Context Protocol tool descriptions with explicit JSON schemas, Siemens reduced the average required agent turns from 155 to 53, marking a 66% drop. Output token consumption dropped similarly, declining 64% from approximately 81K to 29K tokens per rollout.

To enforce strict quality control, NeMo Gym utilized a domain-specific verifier that checks five gating criteria: coverage, validity, relative validation, accuracy, and integrity. Under this setup, if an agent misses a single gate, it receives a zero composite reward, preventing plausible-looking but flawed artifacts from advancing undetected. For engineering practitioners, this methodology demonstrates that evaluating complete end-to-end output trajectories is critical for preventing early setup errors from corrupting downstream chip simulations. The approach offers a repeatable framework for developers across electronic design automation and computer-aided engineering to build reliable, self-verifying autonomous agents.