HeadsUpAI

TSMC adopts NVIDIA AI stack to automate advanced chip manufacturing

TSMC is adopting the full NVIDIA AI and accelerated computing stack to automate the design and production of advanced semiconductors. The integration uses specialized libraries like cuLitho for computational lithography (the process of printing chip mask designs) and cuEST for electronic structure simulation. These tools move manufacturing workloads to GPUs to handle the massive scale required for nanometer-level nodes.
cuLitho efficiency
20-50% improvement
cuEST simulation speed
50x faster
Primary hardware
NVIDIA H200 GPUs
Virtual fab platform
FabTwin (Omniverse)
Defect detection scale
Nanometer-scale

As chip architecture shrinks, the computational burden of transistor modeling has outpaced general-purpose hardware. By using NVIDIA technologies, TSMC is achieving 50x faster chemistry simulations and up to 50% better cost effectiveness or cycle time in lithography. This follows the pattern of leaders like Google DeepMind using AlphaEvolve to design their own specialized hardware.

TSMC is also deploying NVIDIA Omniverse to build FabTwin, a virtual environment for testing tool layouts. This digital-first approach allows for real-time optimization of process parameters to reduce manufacturing variation. These internal optimizations establish the production foundation for the Vera Rubin launch and complement NVIDIA's agentic AI factory security launch.

View the full update on nvidianews.nvidia.com
NVIDIA Newsroom
NVIDIA Newsroom
@nvidianewsroom
X

News at #NVIDIAGTC: TSMC is using NVIDIA accelerated computing and AI to advance semiconductor design and manufacturing. The company is adopting NVIDIA CUDA-X, Metropolis, TAO Toolkit, Omniverse and Cosmos to accelerate chip design, optimize fab operations, and sharpen defect detection. Read the release ➡️ https://t.co/EchWt0YGeA

31retweets203likes
View on X

Still wondering? A few quick answers below.

FabTwin is a virtual fab environment built on the NVIDIA Omniverse platform. It allows TSMC to evaluate process tool layouts and simulate workflows digitally before physical implementation. This virtual-first approach helps identify constraints early and improves planning efficiency for complex semiconductor manufacturing facilities.

NVIDIA cuLitho is a GPU-accelerated library for computational lithography, the process used to design chip masks. By moving these workloads to GPUs, TSMC achieves a 20-50% improvement in cost effectiveness and cycle time compared to traditional CPU-based methods, while maintaining the same total cost of ownership.

TSMC is using the NVIDIA Metropolis vision AI platform and the NVIDIA TAO Toolkit to improve defect classification. These tools enable the detection of nanometer-scale defects and reduce the need for repeated manual labeling and retraining as manufacturing process conditions or inspection tools change.

Every HeadsUpAI update is written based on its original source and reviewed before it's published. Read our editorial standards →

Share this update