In an effort to transition from pure CPU-based computational lithography to create the most advanced chips humanity has ever created, TSMC has announced it will be integrating NVIDIA cuLitho into its production process.
With each smaller nanometer increment in node sizes, the impact of optical diffraction increases exponentially and has to be offset with brand new or “unorthodox” ways while the need to proactively manipulate mask patterns with optical proximity correction (OPC) or inverse lithography technology (ILT) for accurate image wafers will only grow.
Yet computational lithography requires gargantuan levels of computing power with fabrication factories often having their own data center-class servers and hardware to support that cause (Just not at the scale of actual data centers). As NVIDIA CEO Jensen Huang said many times, the CPU is not efficient in doing complex levels of math involving long chain floating point numbers. And we definitely need as much precision as we need when we are talking about creating nanometer components.
Here’s where NVIDIA cuLitho comes in to accelerate it using GPU computing where Team Green claims 350 systems of H100 Tensor Core GPUs can replace more than 40,000 CPU systems, resulting in significantly faster production as well as reduced TCO for fabs (And also some profit for NVIDIA of course).
This is also a win-win situation for both of them because what TSMC produces for its clients often powers its business, resulting in a positive loop. For instance, TSMC’s current servers are running AMD chips which are manufactured by themselves and are used to create newer and better chips for AMD. The same thing is going to happen to NVIDIA as well and we shall see how many advancements and speedups TSMC can benefit from cuLitho.
Prior to TSMC, the last time cuLitho was mentioned was from US-based Synopsys where they received the same “care package” from Team Green to accelerate their chip fabrication process.
