The South Korean chipmaker is building a seventh-generation high-bandwidth memory product to meet a new custom specification for Nvidia's next-generation AI systems. The design avoids the industry trend of increasing stack height, opting for a lower profile to reduce manufacturing complexity and increase supply volumes. Nvidia requested target speeds of 17 to 18 Gbps per pin, a 20% to 22% increase over the 14.4 Gbps found in Samsung's initial HBM4E samples.
The custom memory will feature in the Rubin Ultra AI GPU, which scales GPU interconnects from 72 to 576 units to offset lower memory capacity per chip. Samsung holds a competitive advantage over SK Hynix and Micron by designing and manufacturing both the DRAM and the required logic-based base dies in-house.