TSMC, the world chief in semiconductor manufacturing, is shifting to manufacturing with NVIDIA’s computational lithography platform, known as cuLitho, to speed up manufacturing and push the bounds of physics for the subsequent era of superior semiconductor chips.
A essential step within the manufacture of laptop chips, computational lithography is concerned within the switch of circuitry onto silicon. It requires complicated computation — involving electromagnetic physics, photochemistry, computational geometry, iterative optimization and distributed computing. A typical foundry dedicates huge information facilities for this computation, and but this step has historically been a bottleneck in bringing new know-how nodes and laptop architectures to market.
Computational lithography can be essentially the most compute-intensive workload in all the semiconductor design and manufacturing course of. It consumes tens of billions of hours per yr on CPUs within the modern foundries. A typical masks set for a chip can take 30 million or extra hours of CPU compute time, necessitating giant information facilities inside semiconductor foundries. With accelerated computing, 350 NVIDIA H100 Tensor Core GPU-based programs can now substitute 40,000 CPU programs, accelerating manufacturing time, whereas decreasing prices, house and energy.
NVIDIA cuLitho brings accelerated computing to the sector of computational lithography. Transferring cuLitho to manufacturing is enabling TSMC to speed up the event of next-generation chip know-how, simply as present manufacturing processes are nearing the bounds of what physics makes doable.
“Our work with NVIDIA to combine GPU-accelerated computing within the TSMC workflow has resulted in nice leaps in efficiency, dramatic throughput enchancment, shortened cycle time and diminished energy necessities,” stated Dr. C.C. Wei, CEO of TSMC, on the GTC convention earlier this yr.
NVIDIA has additionally developed algorithms to use generative AI to boost the worth of the cuLitho platform. A brand new generative AI workflow has been proven to ship an extra 2x speedup on high of the accelerated processes enabled by means of cuLitho.
The applying of generative AI allows creation of a near-perfect inverse masks or inverse resolution to account for diffraction of sunshine concerned in computational lithography. The ultimate masks is then derived by conventional and bodily rigorous strategies, rushing up the general optical proximity correction course of by 2x.
Using optical proximity correction in semiconductor lithography is now three many years outdated. Whereas the sector has benefited from quite a few contributions over this era, not often has it seen a metamorphosis fairly as fast because the one supplied by the dual applied sciences of accelerated computing and AI. These collectively permit for the extra correct simulation of physics and the conclusion of mathematical methods that have been as soon as prohibitively resource-intensive.
This monumental speedup of computational lithography accelerates the creation of each single masks within the fab, which speeds the whole cycle time for creating a brand new know-how node. Extra importantly, it makes doable new calculations that have been beforehand impractical.
For instance, whereas inverse lithography methods have been described within the scientific literature for 20 years, an correct realization at full chip scale has been largely precluded as a result of the computation takes too lengthy. With cuLitho, that’s not the case. Modern foundries will use it to ramp up inverse and curvilinear options that may assist create the subsequent era of highly effective semiconductors.
Picture courtesy of TSMC.