Last year, NVIDIA introduced its cuLitho software library, which promises to speed up photomask development by up to 40 times. Today, NVIDIA announced a partnership with TSMC and Synopsys to implement its computational lithography platform for production use, and use the company's next-generation Blackwell GPUs for AI and HPC applications.

The development of photomasks is a crucial step for every chip ever made, and NVIDIA's cuLitho platform, enhanced with new generative AI algorithms, significantly speeds up this process. NVIDIA says computational lithography consumes tens of billions of hours per year on CPUs. By leveraging GPU-accelerated computational lithography, cuLitho substantially improves over traditional CPU-based methods. For example, 350 NVIDIA H100 systems can now replace 40,000 CPU systems, resulting in faster production times, lower costs, and reduced space and power requirements.

NVIDIA claims its new generative AI algorithms provide an additional 2x speedup on the already accelerated processes enabled through cuLitho. This enhancement is particularly beneficial for the optical proximity correction (OPC) process, allowing the creation of near-perfect inverse masks to account for light diffraction.

TSMC says that integrating cuLitho into its workflow has resulted in a 45x speedup of curvilinear flows and an almost 60x improvement in Manhattan-style flows. Curvilinear flows involve mask shapes represented by curves, while Manhattan mask shapes are restricted to horizontal or vertical orientations.

Synopsys, a leading developer of electronic design automation (EDA), says that its Proteus mask synthesis software running on the NVIDIA cuLitho software library has accelerated computational workloads compared to current CPU-based methods. This acceleration is crucial for enabling angstrom-level scaling and reducing turnaround time in chip manufacturing.

The collaboration between NVIDIA, TSMC, and Synopsys represents a significant advancement in semiconductor manufacturing in general and cuLitho adoption in particular. By leveraging accelerated computing and generative AI, the partners are pushing semiconductor scaling possibilities and opening new innovation opportunities in chip designs.

Source: NVIDIA

Comments Locked


View All Comments

  • ballsystemlord - Monday, March 18, 2024 - link

    I'm kinda surprised that TSMC didn't use OpenCL or CUDA to accelerate these computations before now.
  • GeoffreyA - Tuesday, March 19, 2024 - link

    Semiconductor manufacturing becoming dependent on Nvidia is bad. This cunning company is entwining itself in every corner of the earth.
  • Samus - Tuesday, March 19, 2024 - link

    The irony is if anything happens to TSMC, nVidia is a junk company because all of their products are dependent on TSMC manufacturing them. nVidia hasn't even designed their architectures around alternative manufacturing technologies from competitors, meaning it would takes years for them to adapt to TSMC being inaccessible.

    This is an important consideration because as China's economy and government grow more and more desperate, Taiwan is less and less secure. Obviously this is all terrible and nVidia isn't the only company affected but they are the most valuable company affected and are among the only companies that are absolutely 100% dependent on TSMC as it manufactures the vast majority of their products outside of some SoC's.
  • André - Tuesday, March 19, 2024 - link

    Apple is TSMCs biggest customer hands down as well as being the "most valuable".
  • kn00tcn - Tuesday, March 19, 2024 - link

    um literally the previous gen ampere was on samsung so what do you mean hasnt designed architectures around alternate manufacturing technologies from competitors

    wasnt there an iphone gen that used both tsmc and samsung? how can you be so sure it takes years to switch fabs, you're not a chip maker/designer
  • Dante Verizon - Tuesday, March 19, 2024 - link

    Because the chip development cycle is really long...
  • hazmond - Tuesday, March 19, 2024 - link

    Nvidiia 30 series graphics card uses Samsung 8nm
  • kn00tcn - Tuesday, March 19, 2024 - link

    still not as much as MS
  • GeoffreyA - Tuesday, March 19, 2024 - link

    Agreed. It's not good.
  • ballsystemlord - Tuesday, March 19, 2024 - link

    Good point. I had thought that TSMC used AI as an optional part of it's processes, but perhaps not.

Log in

Don't have an account? Sign up now