Abstract:
Samsung Electronics is developing an 8-layer HBM4E product that meets NVIDIA's requirements and strives to play a key role in the supply of its customized high-bandwidth memory NVHBM. The product will reduce the stack height from the originally planned 12- and 16-layer versions, and Nvidia's target speed specification is 17-18Gbps, which is about 20% higher than the 14.4Gbps speed of Samsung's original HBM4E sample.
According to ChainCatcher reports, the 8-layer design is specially built for NVIDIA's disclosed NVLink-optimized NVHBM. This move is seen as a way to reduce manufacturing difficulty and yield pressure, and is more feasible than the traditional approach of increasing capacity by increasing the number of stacking layers. At the same time, this is also seen as part of Nvidia's strategy to deal with memory shortages.
NVHBM is expected to be used in Nvidia’s next-generation AI GPU Rubin Ultra, which is scheduled to be launched next year. By then, the number of interconnections between GPUs will be expanded from a maximum of 72 to 576. Industry sources said that Samsung has demonstrated the highest speed level in the industry in HBM4 verification and may have an advantage in the speed competition.
In the customized HBM market, Samsung is considered to be more competitive than SK Hynix and Micron, because NVHBM requires both DRAM and logic chip-based base die production capabilities, and Samsung can provide these capabilities in an integrated manner.
Comments