TSMC Applies NVIDIA CUDA-X Libraries and Vision AI Across Lithography, Process Control and Defect Inspection

NVIDIA announced at GTC Taipei on May 31, 2026 that TSMC is using NVIDIA accelerated computing and AI to advance semiconductor design and manufacturing. According to NVIDIA, the technology delivers a 20-50% improvement in cost effectiveness or cycle time in computational lithography compared with CPU-based computational lithography, while maintaining the same cost of ownership. C.C. Wei, chairman and CEO of TSMC, said: "By using NVIDIA accelerated computing and AI across fab operations optimization, lithography, process control and inspection, TSMC is strengthening our technology leadership and manufacturing excellence to support our customers' future products and success." As chips move to more advanced nodes, bringing them from design to high-volume production requires massive-scale simulation and real-time optimization. TSMC is using NVIDIA CUDA-X libraries and AI models to accelerate these workloads on NVIDIA GPUs. For computational lithography, a printing method for chip mask design, it uses NVIDIA cuLitho, a GPU-accelerated library. For transistor, equipment and process simulation, it uses NVIDIA cuEST, a GPU-accelerated electronic structure simulation library for 50x faster chemistry simulations, on average, for semiconductor material design. For advanced process control, it uses the NVIDIA cuML machine learning library to distill hundreds of thousands of process parameters spanning thousands of steps as inputs for machine learning models. For fab operations optimization, GPU-accelerated scheduling computation using CUDA has led to notable improvements in fab productivity with NVIDIA H200 GPUs. TSMC is also using the NVIDIA Metropolis platform and NVIDIA TAO Toolkit for defect classification. Using vision AI, it has improved detection of defects at nanometer scale while reducing the need for repeated labeling and retraining as process conditions, inspection tools and defect types change. TSMC is exploring NVIDIA Omniverse libraries to build FabTwin, a virtual fab environment for evaluating process tool layouts and related simulation workflows, so that design scenarios can be tested digitally before physical implementation.