AI Shrinks Chip Design Time
Nvidia described a workflow where AI reduced a GPU design task that once took ten months and eight engineers into an overnight job, while still keeping humans in the loop for oversight. The example frames AI as a concrete cycle-time and labour-compression tool in engineering work. (tomshardware.com)
Chip design still starts with tiny logic blocks, the silicon equivalent of Lego bricks. Nvidia said artificial intelligence now rebuilds one of those block libraries for a new process overnight instead of taking eight engineers 10 months. (tomshardware.com) Those libraries are called standard-cell libraries: pre-designed, pre-verified building blocks that chip tools use to turn a hardware design into real transistors and wires. Synopsys describes them as a core foundation for system-on-chip design, with speed, power and area data built in for downstream tools. (synopsys.com) When a chip company moves to a new manufacturing process, it has to port that library so each cell still works within the new process rules. Bill Dally, Nvidia’s chief scientist and senior vice president of research, said that means handling roughly 2,500 to 3,000 cells each time. (videocardz.com) Dally said Nvidia’s tool, called NB-Cell, uses reinforcement learning, a training method where software improves through repeated trial and feedback. He said the latest versions run that porting job overnight on one graphics processing unit instead of consuming about 80 person-months. (videocardz.com) Nvidia said the machine-generated cells can match or beat human versions on three measures that dominate chip tradeoffs: size, power dissipation and delay, which is the time a signal takes to travel. Dally described the work during a March 2026 conversation with Google DeepMind and Google Research chief scientist Jeff Dean at Nvidia’s GTC conference in San Jose. (videocardz.com) (nvidia.com) The company did not say humans are out of the loop. Tom’s Hardware reported that Nvidia still described fully autonomous chip design as distant, and Dally said the company is applying artificial intelligence to specific stages rather than handing over the full process. (tomshardware.com) That fits the rest of Nvidia’s chip-design research. In 2022, Nvidia researchers said a reinforcement-learning system called PrefixRL produced arithmetic circuits that were smaller and faster than designs from existing electronic design automation flows. (developer.nvidia.com) Nvidia has also been building internal language models for engineering help rather than full replacement. A Hot Chips 2024 tutorial from Nvidia listed ChipNeMo for design assistance and NVCell-RL for cell work among a broader set of tools aimed at automating manual design tasks. (hotchips.org) Electronic design automation vendors are moving the same direction. Cadence said in March 2026 that it was expanding work with Nvidia on “agentic” chip-design tools, and Synopsys said in March 2025 that it was optimizing more than 15 design tools for Nvidia Grace Blackwell systems to speed electronic design automation workloads. (cadence.com) (news.synopsys.com) So the immediate change is not an artificial intelligence system sketching an entire graphics processing unit alone. It is a narrower shift: one repetitive, process-by-process piece of chip engineering getting compressed from months of specialist labor into a single night of machine-guided work that engineers still have to check. (tomshardware.com)
