TSMC, the world chief in semiconductor manufacturing, is transferring to manufacturing with NVIDIA’s computational lithography platform, referred to as cuLitho, to speed up manufacturing and push the bounds of physics for the subsequent era of superior semiconductor chips.
A crucial step within the manufacture of pc 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 large information facilities for this computation, and but this step has historically been a bottleneck in bringing new expertise nodes and pc architectures to market.
Computational lithography can also be essentially the most compute-intensive workload in your complete 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 lowering prices, house and energy.
NVIDIA cuLitho brings accelerated computing to the sphere of computational lithography. Shifting cuLitho to manufacturing is enabling TSMC to speed up the event of next-generation chip expertise, simply as present manufacturing processes are nearing the bounds of what physics makes attainable.
“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,” mentioned Dr. C.C. Wei, CEO of TSMC, on the GTC convention earlier this yr.
NVIDIA has additionally developed algorithms to use generative AI to reinforce the worth of the cuLitho platform. A brand new generative AI workflow has been proven to ship an extra 2x speedup on prime of the accelerated processes enabled via cuLitho.
The appliance of generative AI permits creation of a near-perfect inverse masks or inverse answer to account for diffraction of sunshine concerned in computational lithography. The ultimate masks is then derived by conventional and bodily rigorous strategies, dashing up the general optical proximity correction course of by 2x.
Using optical proximity correction in semiconductor lithography is now three a long time previous. Whereas the sphere has benefited from quite a few contributions over this era, not often has it seen a metamorphosis fairly as fast because the one offered by the dual applied sciences of accelerated computing and AI. These collectively enable for the extra correct simulation of physics and the conclusion of mathematical strategies 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 expertise node. Extra importantly, it makes attainable new calculations that have been beforehand impractical.
For instance, whereas inverse lithography strategies have been described within the scientific literature for twenty 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 now not the case. Modern foundries will use it to ramp up inverse and curvilinear options that can assist create the subsequent era of highly effective semiconductors.
Picture courtesy of TSMC.
