Global semiconductor manufacturing leader TSMC (Taiwan Semiconductor Manufacturing Company) is adopting NVIDIA's cuLitho computational lithography platform to accelerate production and overcome the physical limitations of next-generation advanced semiconductor chips.
Computational lithography, the process of transferring circuits onto silicon wafers, is a crucial step in chip manufacturing. This process requires complex computational tasks involving electromagnetic physics, photochemistry, computational geometry, iterative optimization, and distributed computing. Typically, semiconductor foundries establish large data centers to handle this computation, but this step has long been a bottleneck for the release of new technology nodes and computing architectures.
Computational lithography is the most computationally intensive workload in the entire semiconductor design and manufacturing process. In cutting-edge semiconductor foundries, billions of CPU hours are consumed annually. A typical chip mask might require 30 million or more CPU hours, necessitating large-scale data centers within the foundry. Through accelerated computing, 350 systems powered by NVIDIA H100 Tensor Core GPUs can now replace 40,000 CPU-based systems, speeding up production while reducing costs, space, and power consumption.
NVIDIA's cuLitho applies accelerated computing to the field of computational lithography. TSMC is implementing cuLitho in production to accelerate the development of next-generation chip technologies as manufacturing approaches the limits of physical laws.
At the GTC conference earlier this year, TSMC President Dr. C.C. Wei stated, "Our collaboration with NVIDIA integrates GPU-accelerated computing into TSMC's workflow, achieving significant performance boosts, notably increased productivity, shortened cycle times, and reduced power requirements."
NVIDIA is also developing algorithms that enhance the value of the cuLitho platform through generative AI. A new generative AI workflow has been shown to double the speed of processes already accelerated by cuLitho.
Generative AI can account for light diffraction in computational lithography to create near-perfect inverse masks or inverse solutions. The final mask is derived using traditional, physically rigorous methods, doubling the speed of the optical proximity correction process.
Optical proximity correction (OPC) has been used in semiconductor lithography for over 30 years. This field has seen many benefits over time, but few transformations have been as rapid as those brought by accelerated computing and AI. With these technologies, computationally intensive mathematical techniques that adhere to the laws of physics can now be applied with much greater precision.
Computational lithography significantly accelerates the time it takes for a semiconductor foundry to create each mask, as well as the overall cycle time for developing new technology nodes. More importantly, it enables previously unfeasible new calculations.
For example, inverse lithography techniques have been described in scientific literature for over 20 years, but their lengthy computation time made them largely impractical for full-chip scale applications. cuLitho changes that. Leading semiconductor foundries will use cuLitho to improve inverse and curvilinear solutions, contributing to the creation of the next generation of powerful semiconductors.
Comments