Computing lithography speed increases by 40 times, TSMC expects to introduce Nvidia AI acceleration technology into 2nm trial production in June

NVIDIA worked hand in hand with TSMC, ASML and Synopsys. After four years of development, NVIDIA finally completed the new AI acceleration technology cuLitho. According to reports, CuLitho can increase the computational optical scale of next-generation chips by more than 40 times, making it possible to manufacture 2nm and more advanced chips.

cuLitho is a computational lithography function library that will shorten the photomask process of advanced process chips, increase yield, and significantly reduce the energy consumption required for wafer manufacturing.

According to the Taiwan United Daily News, TSMC will conduct production qualification certification for cuLitho in June this year and complete 2nm trial production to improve the 2nm process yield and shorten the mass production schedule.

According to reports, the new NVIDIA cuLitho software library for computational lithography has been integrated by TSMC and Synopsys into the software, manufacturing processes and systems of their latest generation NVIDIA Hopper architecture GPUs, while equipment manufacturer ASML Works closely with NVIDIA and plans to integrate GPU support into all of its computational lithography software products.

NVIDIA's Jensen Huang said, "The chip industry is the foundation for almost every other industry in the world. As lithography technology reaches its physical limits, NVIDIA is launching cuLitho and working with our partners TSMC, ASML and Synopsys to enable fabs to increase production, Reduce carbon footprint and lay the foundation for 2nm and beyond processes.”

Nvidia said that compared to current photolithography technology - the process of etching patterns on silicon wafers - cuLitho can bring about a 40-fold performance leap, greatly accelerating the kind of large-scale applications that "consume tens of billions of CPU hours every year." Sizing work.

It is reported that it can use 500 NVIDIA DGX H100 systems to achieve the work that could originally be completed by 40,000 CPU systems, running all parts of the computational lithography process in parallel, thus helping to reduce power requirements and potential environmental impacts.

"Computational lithography is the largest computing workload in chip design and manufacturing, consuming tens of billions of CPU hours every year. Large data centers operate 24x7 to create masks for lithography systems. These data centers are the core of chip manufacturing "Part of the nearly $200 billion in capital expenditures that companies invest every year," Huang said that cuLitho can increase the speed of computational lithography by 40 times. For example, the manufacturing of the NVIDIA H100 GPU requires 89 masks. When running on the CPU, it takes two weeks to process a single mask, while running cuLitho on the GPU only takes 8 hours.

According to reports, TSMC can reduce the power from 35MW to 5MW by using cuLitho acceleration on 500 DGX H100 systems, replacing the 40,000 CPU servers previously used for computational lithography. Wafer fabs using cuLitho can produce 3-5 times more masks per day, but only require 1/9 of the current configuration of power.

Of course, although the cuLitho acceleration library is also compatible with Ampere and Volta GPUs, Hopper is currently the fastest solution.

The Fab 20 ultra-large wafer factory built by TSMC in the second phase of Zhuke Baoshan will be used as a 2nm production base and become a 2nm center. The first phase of the Fab20 factory is expected to start risk trial production in 2024 and start mass production in 2025. At the same time, the second phase is currently under construction, and it is expected to gradually begin risky trial production and mass production after the mass production of the first phase.

According to the cooperation plan of four major semiconductor manufacturers, the advancement of this technology will allow the use of finer circuits on chips than today, while speeding up time to market and improving the energy efficiency of large-scale data centers that operate around the clock to promote manufacturing processes. .

TSMC President Wei Zhejia said that the cuLitho team has made significant progress in accelerating computational lithography technology by transferring time-consuming tasks to the GPU. This development brings the possibility for TSMC to deploy lithography solutions such as reverse lithography and deep learning on a larger scale in chip manufacturing, making an important contribution to the continued semiconductor shrinkage.