NVIDIA joins forces with TSMC and other partners to integrate AI technology into the 2nm process

NVIDIA announced at GTC 2023 a collaboration with TSMC, ASML, and Synopsys, three semiconductor industry giants, to accelerate the integration of computational lithography technology into the realm of chip design and fabrication, and introduce a computational lithography library dubbed “cuLitho.”

Computational lithography primarily employs software to model and simulate the entire lithography process, utilizing mathematical preprocessing of photomask files to adjust aberrations and effects in optical lithography, optimize light source and mask shapes, minimize discrepancies between lithographic imaging and chip design, and ultimately enhance yield rates. However, as chip manufacturing progresses towards 3nm and below, each photomask’s burden grows exponentially, increasing fabrication difficulty.

Currently, computational lithography constitutes the most significant computational burden in chip design and manufacturing. Large data centers must operate 24/7, consuming billions of CPU hours annually to create photomasks for lithography systems, with astonishing capital expenditures and energy consumption. NVIDIA, in partnership with TSMC, ASML, and Synopsys, has achieved a major breakthrough in computational lithography technology after four years, launching the cuLitho library and laying the foundation for the next-generation 2nm process.

NVIDIA asserts that using GPUs rather than CPUs for computation can boost computational lithography efficiency by 40 times. The cuLitho library enables workload conversion to GPU parallel processing, allowing 500 NVIDIA DGX H100s to accomplish the tasks of a 40,000-CPU system. This significantly alleviates the burden on wafer fabs, requiring only one-ninth of the power consumption to produce three to five times more photomasks daily, with photomasks previously taking two weeks to produce now processed overnight.

In the long run, with the assistance of AI technology, the cuLitho computational lithography library can achieve superior design rules, higher densities, and increased yields.