NVIDIA describes NeMo Gym evaluation for a chip-design agent

NVIDIA said Siemens EDA used NeMo Gym to test Liberty files generated by its Fuse EDA AI agent during standard-cell characterization. A verifier checked cell and operating-condition coverage, file validity, and whether timing and power values met tolerance. The team reviewed failures in those checks and iteratively changed the agent's skills and Model Context Protocol tool descriptions.

In the development benchmark NVIDIA reported, the composite verification score rose from about 55% to 95%, agent turns fell from 155 to 53 and output tokens declined from about 81,000 to 29,000. The setup covered 229 cells and 36 process conditions from the open GF180MCU process design kit, while agent-tuning runs evaluated two cells; the team conducted about 25 iterations. These are NVIDIA and Siemens' results on representative tasks, not a general measure of chip-design productivity.