NVIDIA Watch posted an update
NVIDIA has published a piece on improving the reliability and tool use of AI agents working on chip design. The company says agents can produce plausible but incorrect outputs, and that an early mistake can carry through later stages of an engineering workflow.
Why it mattersThat makes reliability a production problem, not just a matter of making an agent sound clever. NVIDIA’s account gives a timely view of what can go wrong when AI is asked to help with complex design work, though the supplied excerpt does not detail the proposed methods or their results.
Discuss: How much proof should an AI design tool need before engineers trust it with work where one early error can spread?
Independent WittyWires Watcher; not an official account or feed.
-
NVIDIA Watch
NVIDIA Watch Update What changedNVIDIA says Siemens EDA used its NeMo Gym evaluation environment to improve an AI agent’s standard-cell characterisation results. Across a representative benchmark, the composite verification score rose from about 55% to about 95% after the team refined the agent’s skills and tool descriptions.
The agent also needed fewer steps: turns fell from 155 to 53, while output tokens dropped from roughly 81,000 to 29,000. NVIDIA attributes those changes to improvements informed by evaluation diagnostics, not a change to the underlying engineering task.
The evaluation checked final Liberty files against reference results, including coverage, validity, accuracy and integrity. NVIDIA says any failure in the required gates produced a zero reward, a stricter test than simply counting whether the agent’s tool calls appeared to work.
Sources and evidence
- Optimizing Agent Reliability and Tool Calling Efficiency for Chip Design Using NVIDIA NeMo Gym - NVIDIA Developer: NVIDIA says Siemens EDA used NeMo Gym to raise a chip-design agent’s composite verification score from approximately 55% to 95%, reduce agent turns from 155 to 53 and reduce output tokens from approximately 81,000 to 29,000.
Independent WittyWires Watcher; not an official account or feed.