Samsung Develops 8-Layer HBM4E for NVIDIA's NVHBM, Aiming for Core Partnership

ChainCatcher report, according to the Seoul Economic Daily, Samsung Electronics is developing an 8-layer HBM4E (seventh-generation) product that meets NVIDIA’s specifications, aiming to secure a core partnership position in supplying NVIDIA’s custom high-bandwidth memory, NVHBM. This product reduces stacking height compared to the originally planned 12-layer and 16-layer designs. NVIDIA has set a speed specification of 17–18 Gbps—approximately 20% faster than Samsung’s initial HBM4E sample speed of 14.4 Gbps—and it will be used in NVIDIA’s publicly disclosed NVLink-optimized NVHBM. Unlike traditional HBM approaches that increase capacity by adding more layers, the 8-layer design contradicts conventional logic by reducing stacking complexity, lowering post-processing difficulty and yield pressure, thereby facilitating larger-scale supply and serving as NVIDIA’s strategy to alleviate memory shortages. NVHBM is expected to be deployed starting with NVIDIA’s next-generation AI GPU, “Rubin Ultra,” launching next year. This GPU will expand inter-GPU connectivity from a maximum of 72 to up to 576 connections, enabling hundreds of GPUs equipped with faster HBM to work in tandem and significantly boost overall computational power. In the custom HBM market, Samsung holds a stronger competitive edge over SK Hynix and Micron, as NVHBM requires integrated capabilities in both DRAM and logic-chip-based interposer design—a capability Samsung can deliver end-to-end. Industry insiders note that Samsung has already validated the industry’s highest speed performance in HBM4, giving it a distinct advantage in speed competition.