What are the details of TSMC's deployment of NVIDIA's accelerated computing and AI technologies across semiconductor design and manufacturing, including the specific NVIDIA tools and platforms used, the improvements achieved in lithography (20–50% cost/cycle time improvement via

Here are the confirmed details from the announcement made at the opening of GTC Taipei on June 1, 2026, where NVIDIA and TSMC disclosed a broad deployment of AI and accelerated computing across TSMC's design and manufacturing workflows .

---

Announcement Context

NVIDIA CEO Jensen Huang used the GTC Taipei keynote to announce that TSMC will run a widening share of its design and manufacturing workloads on NVIDIA accelerated computing, with goals of faster turnaround, higher energy efficiency, improved yield, and tighter process control . TSMC Chairman and CEO C.C. Wei framed the collaboration as extending TSMC's role "from supplying AI chips to using AI chips to make chips," deepening the technological tie between the two companies .

---

NVIDIA Technologies Deployed and Measured Improvements

1. Computational Lithography — NVIDIA cuLitho

• Tool: cuLitho, a GPU-accelerated computational lithography library .
• Improvement: 20–50% reduction in cost or cycle time for chip mask design compared to CPU-based methods, while maintaining the same total cost of ownership .
• cuLitho also accelerates inverse lithography by roughly 40x versus CPU systems .

2. Transistor, Equipment, and Process Simulation — NVIDIA cuEST

• Tool: cuEST, a GPU-accelerated electronic structure simulation library .
• Improvement: 50x faster chemistry simulations on average for semiconductor material design, enabling rapid evaluation of advanced transistor materials and chemical formulations .

3. Advanced Process Control & Variation Reduction — NVIDIA cuML

• Tool: cuML, a GPU-accelerated machine learning library (part of RAPIDS) .
• Improvement: TSMC uses cuML to run large-scale analytics on GPUs, distilling hundreds of thousands of process parameters across thousands of manufacturing steps into precision inputs for ML models — significantly reducing process variation and improving yield .

4. Fab Productivity — GPU-Accelerated Scheduling on NVIDIA H200 GPUs

TSMC runs its fab scheduling and operations optimization workloads on NVIDIA H200 Tensor Core GPUs, using CUDA-X libraries to accelerate discrete optimization algorithms, reducing scheduling turnaround time and improving overall fab throughput .

5. Defect Inspection — NVIDIA Metropolis + TAO Toolkit

• Platforms: NVIDIA Metropolis (vision AI platform) and NVIDIA TAO Toolkit (fine-tuning toolkit) .
• Improvement: Automated defect inspection with vision AI that improves detection of nanometer-scale defects while significantly reducing repeated manual labeling and retraining cycles .

6. Digital Twin Exploration — NVIDIA Omniverse (FabTwin)

• Platform: NVIDIA Omniverse, used to build and explore a FabTwin — a digital twin of a semiconductor fab .
• Purpose: Enables TSMC to simulate and optimize fab layouts, material flows, equipment configurations, and production logistics in a physically accurate virtual environment before making changes to the physical factory .

---

Summary Table

Workload Area	NVIDIA Technology	Key Improvement
Computational lithography	cuLitho	20–50% cost/cycle time improvement
Transistor/material simulation	cuEST	50x faster chemistry simulations
Process control / variation reduction	cuML	Faster analytics on hundreds of thousands of process parameters
Fab scheduling & optimization	H200 GPUs + CUDA-X	Accelerated discrete optimization algorithms
Defect inspection	Metropolis + TAO Toolkit	Improved nanometer-scale defect detection
Digital twin / fab simulation	Omniverse (FabTwin)	Virtual fab planning and optimization

The full stack runs on NVIDIA CUDA-X libraries and AI models deployed on NVIDIA GPUs inside TSMC's production fabs .