Samsung, Nvidia Co-Develop AI to Accelerate Ultra-Low-Power NAND Research

Samsung Electronics and Nvidia have jointly developed artificial intelligence technology that can dramatically accelerate research and development of ferroelectric materials, a next-generation NAND flash memory technology.

According to industry sources on the 13th, researchers from Samsung Electronics' Semiconductor Research Institute, Nvidia, and Georgia Institute of Technology developed a "Physics-Informed Neural Operator (PINO)" model capable of analyzing ferroelectric-based NAND device performance more than 10,000 times faster than conventional methods. The research findings were published to the academic community on the 6th.

Ferroelectric materials can maintain a polarized state with separated positive and negative poles without applying an external electric field or high voltage. This separation of poles enables electrons to move between them, creating digital information of 0 or 1. Silicon, the primary material in current semiconductors including NAND, requires relatively high voltage to achieve pole separation and data storage. Replacing silicon with ferroelectric materials could significantly reduce power consumption.

Ferroelectric NAND, made with ferroelectric materials instead of conventional silicon, is attracting attention as a breakthrough technology that could address both supply shortages and power constraints facing big tech companies including Nvidia. The technology enables high-density stacking up to 1,000 layers while reducing power consumption by up to 96%.

Samsung Electronics published research in Nature late last year detailing the mechanism of "ferroelectric transistors for low-power NAND flash memory" with these capabilities. The company announced it would pursue follow-up research targeting commercial product development in response to growing AI storage demand. Samsung currently possesses NAND stacking technology at 200-300 layers and is reportedly focusing on ferroelectric materials as a core technology to achieve 1,000-layer stacking. According to the Korean Intellectual Property Office, Samsung Electronics leads global ferroelectric patent share at 27.8%, ahead of Intel, TSMC, and SK hynix.

However, commercializing ferroelectric materials requires analyzing the complex characteristics of the material and identifying optimal device structures. The joint AI technology addresses this by enabling analysis speeds 10,000 times faster than before. TCAD, the analysis tool widely used in the semiconductor industry, typically requires 60 hours per operation, limiting research speed. The research team succeeded in reducing operation time to under 10 seconds using AI trained on physical laws.

This research draws particular attention as a collaboration between Samsung Electronics and Nvidia, its largest memory partner, to strengthen competitive advantage in ferroelectric technology. Industry observers note Nvidia's direct participation in next-generation memory R&D as an unusual move. Cases of Nvidia, a GPU manufacturer, participating in memory research—particularly ferroelectric technology that has not yet been commercialized—are rare.

This move is interpreted as Nvidia's strategy to proactively strengthen technology partnerships as the core of AI computing shifts from GPUs to memory including high-bandwidth memory (HBM) and NAND, going beyond simple supply management. For Nvidia, investing in new technology could address memory supply shortages that may disrupt its AI accelerator supply chain and power constraints that could burden its data center customers.

According to market research firm Omdia, global NAND supply peaked at 21.387 million wafers in 2022 and is projected to decline to 15.408 million wafers this year. Even by 2028, supply is expected to reach only 17.61 million wafers, insufficient to meet surging demand. Additionally, the International Energy Agency (IEA) projects global AI data center power consumption will surge from approximately 450 TWh in 2024 to 550 TWh this year and double to 950 TWh by 2030.