Samsung cuts an 18-hour chip simulation to 17 seconds on Nvidia GPUs as SK hynix speeds up memory modeling - MLQ.ai

Samsung Electronics has reported a breakthrough in chip design simulation, reducing a thermal and mechanical analysis that previously took about 18 hours on a 32-core CPU to just 41 seconds on a single Nvidia H100 GPU. Using four GPUs, the same task was completed in approximately 17 seconds. This represents a speedup of roughly 1,600 times for a single GPU and about 3,800 times for the four-GPU configuration, according to Korean press accounts of the presentation at Nvidia AI Day Seoul.

The dramatic reduction in processing time is not simply a result of moving existing code to a GPU. Samsung engineer Lee Sung-yeop explained that the company split large chip designs into smaller regions. AI models were then used to predict representative material properties for each region, allowing for calculations on a coarser computational grid. The results were run on Nvidia GPUs, with an in-house algorithm used to reduce errors between neighboring regions.

Samsung also highlighted its collaboration with Nvidia, noting it is working on the company's cuISS and cuDSS libraries for TCAD (technology computer-aided design). Additionally, Nvidia's cuEST library sped up quantum chemistry calculations on Samsung's sMOL materials platform by up to five times. The company's next step is to develop engineering agents that can link simulation tools and AI models to compare design options more efficiently.

SK hynix presented similar results at the same event. Kim Jung-han, who leads the TCAD Intelligence team at SK hynix's Future Technology Research Institute, said full-chip thermal and deformation analysis ran more than 30 times faster than its conventional CPU method. An AI model for predicting etch profiles was also reported to be dozens of times faster than standard simulation.

Kim explained that multi-step etch processes involve more than 10 input variables per step, such as pressure, temperature, and gas flow. As the number of steps increases, the number of possible cases grows exponentially. SK hynix addressed this by training PhysicsNeMo models on its own simulation data and reusing results computed for small regions. The company also split full-chip runs across GPUs, with performance rising nearly in proportion to the number of GPUs used.

Both companies built their models using Nvidia's open-source PhysicsNeMo framework. The work is part of a multiyear partnership between Nvidia and SK hynix, announced in June, which focuses on CUDA-X libraries and PhysicsNeMo for TCAD workflows and in-house engineering codes. The figures reported are for individual calculations under test conditions and do not cover the entire chip development process.

Market context

The advancements come as both Samsung and SK hynix are racing to supply high-bandwidth memory (HBM) for AI accelerators. This market has become a critical revenue driver for the companies, with both reporting record margins. The ability to simulate chip designs more quickly could potentially accelerate the time to market for new products, though the companies provided no data on yield improvements.