Samsung Develops HBM5 Memory Using 2nm Process

According to reports, Samsung Electronics has confirmed that it is developing its eighth-generation High Bandwidth Memory, HBM5, which will utilize a 2nm process for the base die. For the next iteration, HBM5E, the company plans to adopt 1d (seventh-generation 10nm-class) DRAM technology for the core die while continuing to use the 2nm process for the base chip.

Meanwhile, a new AI inference–focused processor, the Groq 3 LPU (Language Processing Unit) developed for Nvidia, has been assigned to Samsung Electronics for foundry manufacturing using its 4nm process technology. According to Han Jin-man, president of Samsung's Foundry Business, the company is confident in maintaining stable supply. He revealed that the Groq 3 LPU is currently being produced at Samsung's Pyeongtaek fabrication facility, and order volumes for this year have exceeded expectations. Samsung also noted that the HBM4 base die is manufactured using the same 4nm process, and demand for this node is expected to rise significantly in the future.

Han further disclosed that Samsung's collaboration with Groq dates back to 2023, when Samsung engineers were already involved in the project and supported the chip’s design phase—even before Nvidia acquired Groq. During the early stages of the partnership, Nvidia and Groq evaluated Samsung's chip performance and ultimately selected Samsung as the manufacturing partner.

Hwang Sang-joon, Executive Vice President of Memory Development at Samsung's Device Solutions (DS) division, explained that the HBM5 core die will be produced using 1c (sixth-generation 10nm-class) DRAM technology, while the base die will be developed on the 2nm process. For HBM5E, the core die will transition to 1d-class DRAM, with the base die continuing to rely on 2nm technology. Samsung previously applied 1c DRAM to the HBM4 core die and paired it with a 4nm base die, achieving industry-leading performance.

At the Nvidia GTC 2026 held on March 16, 2026, Hwang and Han presented these technologies at Samsung's exhibition booth. Their appearance marked a rare occasion in which Samsung executives publicly introduced their technologies at Nvidia's flagship developer conference. On the same day, Jensen Huang posed for photos with the two Samsung executives and signed wafers for both the Groq 3 LPU and HBM4, writing the messages "GROQ SUPER FAST" and "AMAZING HBM4!"

Han Jin-man also noted that mass production of the Groq 3 LPU is expected to begin between the end of the third quarter and early fourth quarter, with market demand projected to rise significantly next year. These developments highlight Samsung's continued push in both high-bandwidth memory innovation and advanced semiconductor foundry manufacturing, while further strengthening its collaboration with Nvidia in the AI computing ecosystem.