NVIDIA, TSMC Integrate AI into Semiconductor Fabs
AI-generated image: synthetic visual, not an actual depiction of events, people, or locations.
Key Takeaways
- TSMC is using NVIDIA cuLitho to achieve a 20-50% improvement in cycle time or cost effectiveness for computational lithography.
- The cuEST library provides 50x faster chemistry simulations on average for semiconductor material design.
- TSMC is utilizing NVIDIA Metropolis and TAO Toolkit to improve detection of nanometer-scale defects using vision AI.
- The companies are developing FabTwin, a virtual fab environment built on NVIDIA Omniverse to simulate and optimize fab layouts.
NVIDIA said TSMC is using its accelerated computing and AI software across semiconductor design and manufacturing, extending the chipmaker’s role in artificial intelligence from supplying GPUs to helping optimize the fabs that make advanced chips.
The announcement, made Monday at NVIDIA GTC Taipei, adds another layer to the long-running partnership between NVIDIA and Taiwan Semiconductor Manufacturing Co., the world’s largest contract chipmaker and a critical manufacturing partner for NVIDIA’s AI processors. NVIDIA said TSMC is applying its CUDA-X libraries, AI models, Metropolis vision AI platform and Omniverse tools across areas including computational lithography, process simulation, defect inspection and fab operations.
The move comes as advanced-node manufacturing becomes increasingly dependent on large-scale simulation and real-time optimization. As chips shrink and designs grow more complex, manufacturers need to model physics, inspect wafers at nanometer scale, optimize production flows and manage thousands of process variables across increasingly automated fabs.
TSMC is using NVIDIA’s cuLitho, a GPU-accelerated computational lithography library, to improve the cost effectiveness or cycle time of chip-mask design by 20% to 50% compared with CPU-based workflows, according to NVIDIA. The foundry is also using cuEST, an electronic structure simulation library, for chemistry simulations used in semiconductor material design, with NVIDIA saying the tool can deliver average speedups of 50 times.
