Accelerated Computing Helps TSMC and NVIDIA Transform Semiconductor Manufacturing.
TSMC, a leading global semiconductor manufacturer, is using the NVIDIA Computational lithography platform to speed up production and push physics for next-generation semiconductor devices.
What Is Computational lithography?
Computational lithography, a critical computer chip manufacturing step, transfers circuitry onto silicon. Complex computations integrating photochemistry, computational geometry, iterative optimization, distributed computing, and electromagnetic physics are needed. Large data centers are often set aside by a typical foundry for this calculation, but up until now, this process has been a roadblock in the introduction of new computer architectures and technological nodes.
NVIDIA has created algorithms that use generative AI in order to increase the cuLitho platform’s value. It has been shown that a novel generative AI workflow, in addition to the faster processes made possible by cuLitho, may provide a 2x speedup.
The use of generative AI makes it possible to create an almost perfect inverse mask or inverse solution to take into consideration the light diffraction that occurs in computational lithography. After that, the final mask is created using conventional, physically demanding techniques, which results in a 2x acceleration of the optical proximity correction procedure overall.
The Challenge
It has been thirty years since optical proximity correction was used in semiconductor lithography. Throughout this time, the area has benefitted from many contributions, but it has seldom witnessed a revolution quite as quick as the one brought about by the twin breakthroughs of faster computing and artificial intelligence. Together, they enable the implementation of mathematical concepts that were previously impractically resource-intensive and the more realistic modeling of physics.
This massive computational lithography speedup shortens the entire cycle time for building a new technology node by accelerating the production of each and every mask in the fab. More significantly, it allows for previously unfeasible new computations to be made.
For instance, while inverse lithography methods have been documented in scientific literature for twenty years, the computational complexity has mainly prevented an accurate implementation at full chip size. That is no longer the case with cuLitho. It will be used by state-of-the-art foundries to ramp up inverse and curvilinear solutions that will contribute to the development of the next powerful semiconductor generation.
NVIDIA cuLitho
Accelerate computational lithography
NVIDIA cuLitho is a library that offers tools and algorithms that are designed to accelerate GPU-based computational lithography and semiconductor manufacturing processes by a significant margin compared to existing CPU-based techniques.
Computational lithography, a crucial process in the production of computer chips, is a sophisticated calculation combining distributed computing, photochemistry, computational geometry, electromagnetic physics, and iterative optimization.
Massive data centers are required for this computational lithography stage, which is currently one of the greatest compute workloads in semiconductor manufacturing. The calculation needs increase exponentially with the growth of silicon downsizing.
Advantages of NVIDIA cuLitho
The foundation of accelerated computing is acceleration libraries. A library called NVIDIA cuLitho aims to address the new difficulties in nanoscale computational lithography. By accelerating inverse lithography by 40X using GPUs, new approaches are being developed to make next semiconductor technologies more affordable and predictable. It makes use of decades’ worth of CUDA infrastructure investment while introducing fresh advancements.
Performance
A 40X performance speedup is powered by quicker inverse lithography technology (ILT), which produces correct photomasks more quickly.
Efficiency
Two weeks may be used to process photomasks in a single night. It is possible to create 3X to 5X more masks each day.
Cost Savings
The job of 40,000 CPU systems may be completed by 500 NVIDIA Hopper GPU computers running cuLitho. That is 1/8 of the space and 1/9 of the power.
Future Silicon Scaling
Faster OPC makes it possible to use novel lithography techniques, such as curvilinear OPC, high-numerical aperture extreme-ultraviolet (high-NA EUV) lithography, and subatomic modeling, to address the shrinking of next-generation semiconductors.
