News Categories

NVIDIA cuLitho to accelerate next-gen chip manufacturing, will be adopted by TSMC, ASML, Synopsys (Updated)

By Zachary Chan & Vijay Anand - on 19 Mar 2024, 4:23pm

NVIDIA cuLitho

Note: This article was first published at GTC 2023 and is re-published with a few momentum updates made known at GTC 2024.

NVIDIA cuLitho brings GPU acceleration to computational lithography. Image source: NVIDIA

Everyone knows that as microprocessor circuitry gets smaller and smaller, it gets harder and harder to manufacture. Moore's Law (the doubling of transistors every two years) has been decreed as dead. And one of the main reasons for this is photolithography, the process of printing circuitry onto silicon wafers using lasers. In fact, the semiconductor industry is fast approaching the hard limits of physics as we reach 2nm nodes, where the size of said circuitry itself has become smaller than the wavelength of light that can be used to etch them. The most advanced EUV (Extreme Ultraviolet) lithography machines by ASML can cost 100s of millions of dollars, have a wavelength of about 13.5nm.

To put that into context, 2nm is smaller than a strand of human DNA. I'll let you chew on that nugget for a second or two. 

To workaround this problem, chipset manufacturers have to create increasingly complex light- or photomasks just to ensure that the right spot on a chip is 'lasered' correctly. This software portion of chip manufacturing is incredibly compute-heavy and a single photomask can require days or even weeks to process. 

Just like almost every other industry in the world that used to rely solely on CPUs for compute, there is now an accelerated GPU to take over. This piece of news is actually very simple, but we had to give context to it. 

One of NVIDIA's announcements last night at GTC 2023 is the release of the cuLitho software library (cu for Cuda), which is essentially a Cuda accelerator specifically designed to accelerate computational lithography workloads. NVIDIA claims cuLitho can provide up to 42x performance improvements over current CPU processing, with 9x lower power consumption. Just 500 NVIDIA DGX H100 systems with cuLitho can achieve the work of 40,000 CPU systems before, and a photomask that used to take two weeks to process can now be done overnight. This provides manufacturers the ability to compute even more complex and accurate photomasks, reducing errors and improving chip yield, speeding up time to market of new chips, while reducing costs.

There are only a few key players in the chip-making business and almost all of them - ASML, TSMC, Synopsys - are already on the cuLitho bandwagon.

GTC 2024 Update: At NVIDIA's premier AI conference of the year this year, NVIDIA shared that TSMC and Synopsys are now both going into production with NVIDIA's computational lithography platform to accelerate manufacturing, as they've integrated NVIDIA cuLitho with their software. In the future, they'll speed up fabrication of NVIDIA's newest Blackwell GPUs, which were just announced.

Additionally, NVIDIA has developed algorithms to apply generative AI to improve the value of cuLitho platform and speed up more accurate photomask generation. It looks like Gen AI has no bounds and Jensen Huang, NVIDIA's CEO, is right to say their next-gen Blackwell GPUs will bring about an industrial revolution. Read more about that here.

Source: NVIDIA

Join HWZ's Telegram channel here and catch all the latest tech news!
Our articles may contain affiliate links. If you buy through these links, we may earn a small commission.