Samsung Electronics (005930.KS) is building a self-reinforcing ecosystem in which it supplies memory for Nvidia graphics processing units while deploying those same GPUs in its own chip manufacturing lines. The idea is to use GPUs to predict and correct light distortion — a chronic problem in lithography, a core chipmaking step — and shorten process development time.
Analysts say the relationship between the two companies is evolving beyond that of a parts supplier and a customer into a strategic partnership that lifts manufacturing competitiveness and productivity through technology exchange.
Nvidia singled out Samsung Electronics on the 26th as a leading example of GPUs spreading onto factory floors, speaking on its second-quarter earnings call.
Colette Kress, Nvidia's chief financial officer, said Samsung Electronics is using Nvidia's cuLitho to achieve up to 20 times higher performance in computational lithography.
cuLitho is software that uses GPUs to calculate in advance how light will spread and how much circuit shapes will deviate before circuits are physically etched onto a wafer.
Even when chipmakers try to transfer a designed circuit onto a wafer exactly as drawn, the actual pattern can shift slightly because of diffraction, the spreading of light. That requires technology to predict such shifts beforehand and make small corrections to the circuit shapes on the photomask, the master plate that carries the chip circuit pattern. The process involves running repeated simulations on high-performance computers to find the optimal circuit design that minimizes light distortion.
Samsung Electronics adopted Nvidia's cuLitho late last year, raising its lithography simulation speed by up to 20 times. Faster calculation shortens the time engineers need to predict and revise circuit patterns, pulling forward the development timeline for chip processes.
Samsung Electronics is also running proof-of-concept work on lithography simulation using quantum computing together with Samsung SDS (018260.KS), its information technology services affiliate. It is a two-track strategy: easing immediate computing bottlenecks with Nvidia GPUs while researching quantum computing to prepare for the far heavier computational loads expected in ultra-fine process nodes.
An industry official said Nvidia wants AI applications to expand beyond language-centered computing into manufacturing AI that handles visual information and shop-floor data, adding that the company will accelerate related demonstration projects with Korean firms that own large-scale manufacturing infrastructure, including Samsung, LG and Hyundai Motor (005380.KS).
