NVIDIA’s Next-Gen Vera Rubin Integrates Samsung’s HBM4, Claiming Title of “Fastest AI Memory Solution” in the Industry

Samsung’s resurgence in the High Bandwidth Memory (HBM) sector is noteworthy, highlighted by the recent announcement of its HBM4 technology, which boasts exceptional speed and is now available for commercial use.

Unveiling Samsung’s HBM4 Technology Enhancements

The competition surrounding HBM4 has intensified, with Samsung significantly enhancing its offerings. In a breakthrough announcement, the South Korean tech giant disclosed that its HBM4 memory units are among the first to be commercially shipped, constructed on a 6th-generation 1c DRAM alongside a 4nm logic die, utilizing components all sourced in-house.

A standout feature of Samsung’s HBM4 is its remarkable pin speed rating, with standard performance reaching up to 11.7 Gbps—an impressive 46% increase compared to the previous 8 Gbps standard. Remarkably, after overclocking, speeds can reach up to 13 Gbps, fulfilling a crucial requirement set by NVIDIA for their technology. Additionally, Samsung has introduced a current offering based on a 12-layer architecture, with plans for a 16-layer HBM4 module in development, potentially increasing per-module capacity to a substantial 48 GB.

Detailed view of a Samsung HBM4 chip highlighting its design

Although Samsung has not disclosed the specific customer behind this deployment, it is known that NVIDIA is the sole infrastructure provider leveraging this advanced memory technology, particularly within its Vera Rubin AI architecture. With the Vera Rubin system prioritizing low latency and optimal response times, memory performance is a critical focus. This alignment positions Samsung’s HBM4 solution as a key driver of memory advancements for NVIDIA, especially regarding capacity and bandwidth enhancements.

Looking ahead, Samsung anticipates a threefold increase in its HBM revenue by 2025 and is setting its sights on launching HBM4E in the latter half of 2026.

Source & Images