Chipmakers are handing more of the grinding engineering work to AI agents, and NVIDIA wants its software underneath them. The company announced that semiconductor leaders are adopting its CUDA-X libraries, including a new one called cuPotential, to build what it calls AI-defined fabs: factories where AI helps engineers tune processes, chase down production problems and raise efficiency.
The pressure behind this is plain. Demand for advanced chips keeps rising, and fabs must expand output while managing ever more complicated manufacturing steps. Engineers need to model materials, analyse inspection data and test process changes in digital twins before anything touches a production line.
“The semiconductor industry made the AI revolution possible. Now, AI is transforming how chips are built,” said Tim Costa, NVIDIA’s vice president and general manager of industrial and computational engineering.
Simulating atoms faster
Modeling how atoms and molecules interact is central to understanding semiconductor materials, but the calculations can eat up hours even in software already tuned for GPUs. cuPotential adds specialised GPU functions that developers can drop into existing software, and it lets AI agents run and refine molecular simulations themselves. In early-access testing it sped up already GPU-optimised workflows by a further 2-3x. Synopsys is among the early users.
Synopsys is also using cuEST and cuDSS. Early QuantumATK results showed a potential 30x speedup with cuEST, and S-Device benchmarks showed 12x with cuDSS. Cadence reported that its AuraStack AI Super Agent optimises advanced packages for heat and stress 8x faster with cuDSS. COMSOL cited simulations up to 10x faster. Voltai said it has cut complex electro-thermal-mechanical workflows for 3D-IC packaging from months to a day.
Preparing and checking production
Lam Research plans to build NanoVDB, an open source library for 3D data, into its SEMulator3D platform. The company expects select wafer-etch modeling to run up to 30x faster, and NVIDIA’s PhysicsNeMo framework has been shown to shrink plasma-prediction simulations from days to as little as milliseconds. Siemens is using cuLitho to speed up its Calibre software for mask preparation. Luminary says its physics AI models let engineers explore 100x more package designs in the same schedule.
On the inspection side, imec used cuLitho to improve a lithography process window by 30%. KLA built a new NPP kernel that runs 50x faster than earlier CPU versions. Teradyne reported roughly 85x faster probe-maintenance inference, which helps spot degrading probes before they hurt yield. Applied Materials cited chamber simulations up to 35x faster.
cuPotential is available in early access, NanoVDB is open source within OpenVDB, and NPP ships in the CUDA Toolkit. Lam Research targets release of its NanoVDB-accelerated SEMulator3D capability by the end of the year.
Sources
- NVIDIA and Semiconductor Leaders Advance AI Across Chip Design and Manufacturing for AI-Defined Fabs
- NVIDIA and semiconductor leaders advance AI across chip design and manufacturing – Silicon Semiconductor News
