Nvidia says AI sped GPU design
Nvidia reported that AI reduced a GPU design task that previously took eight engineers ten months down to an overnight run. (tomshardware.com) The company also cautioned that fully autonomous chip design remains distant and that human oversight is still required. (tomshardware.com)
Modern chip design still starts with tiny building blocks, and Nvidia says artificial intelligence now lays out some of them in hours instead of months. (tomshardware.com) Those building blocks are called standard cells: pre-made logic pieces that engineers reuse by the thousands when they build a graphics processing unit, or GPU. Every time a chipmaker moves to a new manufacturing process, that library has to be adapted to the new rules. (nvidia.com) Nvidia chief scientist Bill Dally said at the company’s March 2026 Graphics Technology Conference that porting a standard-cell library of about 2,500 to 3,000 cells used to take eight engineers about 10 months, or 80 person-months. He said Nvidia now runs that job overnight on a single GPU with a reinforcement-learning system called NVCell. (nvidia.com; tech.yahoo.com) Reinforcement learning is a trial-and-error method that scores many design choices and keeps the ones that work better, like software playing millions of rounds of a game. Nvidia’s 2021 NVCell paper said the system uses that approach to place devices and route connections while fixing rule violations from chip-manufacturing constraints. (nvidia.com; nvidia.com) Dally said the overnight results were better than human designs on measures including cell size, power dissipation and delay, which is the time a signal takes to move through a circuit. Nvidia’s 2023 NVCell 2 paper reported cleaner and more routable layouts on hard cells and said 98.9% of tested cells were design-rule clean. (tech.yahoo.com; nvidia.com) Nvidia said the software is one piece of a broader push to use artificial intelligence across chip planning, bug handling and verification, which is the phase where teams check whether a design actually matches its intended behavior. Dally said the company is “trying to use AI wherever we can” in the design flow. (tomshardware.com; videocardz.com) The timing matters because Nvidia is trying to ship new GPU generations on a fast cadence while chip manufacturing keeps getting more complex at advanced nodes. Faster cell-library work can shorten one of the repeated chores that comes with each move to a new process technology. (nvidia.com; tomshardware.com) Nvidia is not saying software can design an entire processor by itself. Dally said the industry is still “a long way” from fully autonomous chip design, and Tom’s Hardware reported that human oversight remains part of the process. (tomshardware.com) That leaves Nvidia describing artificial intelligence less as a replacement for chip engineers than as a tool for specific, repeated design jobs. For now, the company’s claim is narrower: one overnight run can now handle work that once tied up a team for most of a year. (tomshardware.com; nvidia.com)
