Nvidia used AI to speed chip design

Nvidia says it now uses artificial intelligence to do one chip-design job overnight that previously took eight engineers about 10 months. (videocardz.com) The task is standard-cell library porting, a step where engineers rebuild the tiny logic blocks that get reused across a chip when moving to a new manufacturing process. Nvidia Chief Scientist Bill Dally described the result in a March 2026 conversation with Google Chief Scientist Jeff Dean at Nvidia’s GTC conference. (nvidia.com) (videocardz.com) Standard cells are the basic Lego bricks of a processor, and each new process node comes with new layout rules, routing limits, and manufacturing constraints. Nvidia’s research group says advanced nodes have made that work harder because routing tracks shrink and design rules multiply. (nvidia.com 1) (nvidia.com 2) Nvidia has been working on that problem for years through a tool called NVCell, which uses reinforcement learning to place and route those cells while fixing rule violations. In a 2021 paper, Nvidia researchers said NVCell produced equal or smaller layouts for more than 90% of single-row cells in an industry-standard library on an advanced node. (nvidia.com) Dally said Nvidia is using artificial intelligence in several parts of its internal chip-design flow, not just cell libraries. He listed design-space exploration, bug handling, verification, and standard-cell work, while saying the company is still far from a system that can design a chip without human input. (videocardz.com) (tomshardware.com) That fits Nvidia’s earlier push to build chip-specific language models from its own engineering data. In October 2023, Nvidia researchers introduced ChipNeMo, a domain-adapted large language model trained to act as an engineering assistant, generate electronic design automation scripts, and summarize bugs. (nvidia.com 1) (nvidia.com 2) Chip design is one of the most expensive engineering workflows in computing, and the software used for it is called electronic design automation, or EDA. Nvidia’s own design-automation lab says its work spans register-transfer level design, verification, logic synthesis, physical design, sign-off, and design-for-manufacturing. (nvidia.com) The company’s message is narrower than “artificial intelligence designs chips now.” Nvidia is saying it has automated specific, repetitive parts of the workflow with internal models trained on decades of its own graphics processing unit design data, while engineers still direct and check the work. (videocardz.com) (tomshardware.com) For Nvidia, the immediate payoff is not a self-driving chip team but faster internal iteration on the small building blocks every new processor depends on. That is a concrete example of the company using artificial intelligence on itself, inside the same semiconductor pipeline that feeds its graphics processing units and artificial intelligence accelerators. (nvidia.com) (tomshardware.com)