Samsung Reportedly Developing Custom 8-Layer HBM for NVIDIA with 17-18Gbps Speeds
Samsung Electronics is reportedly developing a custom 8-layer High Bandwidth Memory (HBM) product, referred to as HBM4E or NVHBM, for NVIDIA. This memory features a low-stack design and supports high pin speeds of 17 to 18Gbps, surpassing Samsung's previous HBM4E sample speed of 16Gbps. This development aims to mitigate the "memory wall" bottleneck in AI accelerators by increasing throughput while reducing manufacturing costs and improving production yields. It also highlights Samsung's competitive advantage in offering full-process custom HBM development due to its integrated memory and foundry businesses. The custom memory will utilize a low-stack design to reduce DRAM die usage. Additionally, NVIDIA's NVHBM and standard HBM4E are expected to feature distinct base die designs.
## BACKGROUND
High Bandwidth Memory (HBM) is a specialized 3D-stacked DRAM architecture used to provide high-speed data access for AI accelerators. The "memory wall" refers to the performance bottleneck where data transfer speeds between the processor and memory cannot keep up with the processor's computational power. Custom HBM designs often utilize a specialized base die at the bottom of the stack to control interfaces and optimize performance.