Maeil Business Newspaper reports that Samsung Electronics (Samsung) is set to finish the development of its sixth-generation high bandwidth memory (HBM4) by 2025 to secure orders for Nvidia's graphics processing units (GPUs).
The semiconductor giant is actively enhancing its DRAM manufacturing processes and packaging technologies, with its ability to regain dominance in the HBM market becoming a key industry focus.
According to sources, Samsung is not only reducing circuit line width but also accelerating the development of packaging technology for stacking DRAM. Industry reports indicate that Samsung has established a dedicated production line for HBM4, D1c, which is currently in the trial production phase.
The line widths for the 10nm DRAM process are gradually reduced in the order of D1x, D1y, D1z, D1a, D1b, and D1c, with D1c being the most advanced process in the 10nm category. Presently, HBM3E is produced using the D1a process, meaning that Samsung will skip the D1b stage and directly enter the D1c process.
In its financial report for the third quarter of 2024, Samsung revealed its roadmap for HBM4 development, targeting completion by the second half of 2025 for mass production, nearly six months ahead of its original target of 2026.
According to Samsung's presentation at the International Solid-State Circuits Conference (ISSCC 2024), the data transfer speed of HBM4 is expected to reach 2TB/s, representing a 66% increase over HBM3E. Furthermore, the memory capacity is projected to reach 48GB, an increase of 33% compared to HBM3E.
SK Hynix is also accelerating its HBM4 development efforts. Industry sources suggest that SK Hynix may advance its HBM4 development timeline to the second half of 2025, likely in response to requests from Nvidia.
South Korean industry analysts believe that if Samsung becomes the first to supply HBM4 to Nvidia, doubts about the company's capabilities will be dispelled. However, analysts at Hanwha Investment & Securities indicate that SK Hynix is likely to maintain its dominant position in the HBM market and expect it to be the first to provide customers with HBM4 samples.
Beyond Nvidia, Samsung is reportedly developing customized HBM4 solutions for Microsoft and Meta.
Meanwhile, Nvidia plans to break its tradition of releasing new GPUs every two years, announcing that starting in 2026, it will unveil a new product annually. In June 2024, Nvidia CEO Jensen Huang revealed the development roadmap for the next-generation GPU platform "Rubin," which will utilize TSMC's 3nm process and feature eight HBM4 chips, with mass production anticipated in 2026.