According to the latest HBM market research from TrendForce, NVIDIA is planning to join more HBM suppliers for better and sound supply chain management.. Among them, Samsung’s HBM3 (24GB) is expected to complete verification at NVIDIA in December this year. Previously, NVIDIA's HBM was exclusively supplied by SK Hynix, but now Samsung and Micron will join in.This also means that all three storage leaders will supply HBM to Nvidia.
The progress of HBM3e is arranged according to the timeline as shown in the following table:
Micron has provided 8hi (24GB) NVIDIA samples at the end of July this year, SK Hynix has provided 8hi (24GB) samples in mid-August this year, and Samsung has provided 8hi (24GB) samples in early October this year.HBM3e product verification is expected to be completed in the first quarter of 2024.
HBM (HighBandwidthMemory) is high bandwidth memory, which is a type of graphics DDR memory., by using advanced packaging methods (such as TSV through silicon via technology) to vertically stack multiple DRAMs and interconnect them with the GPU through an interposer.
The advantage of HBM is that it breaks the memory bandwidth and power consumption bottlenecks. Moreover, the CPU handles more types of tasks, is more random, and is more sensitive to speed and delay. The HBM feature is more suitable for use with the GPU for intensive data processing operations. NVIDIA’s new generation of AI chips are all equipped with HBM memory.
Since the advent of HBM products, HBM technology has developed to the fourth generation.They are HBM, HBM2, HBM2e, and HBM3 respectively, and the fifth-generation HBM3e is already on the way.
As for the next generation HBM new product HBM4 (sixth generation), in addition to the existing 12hi (12 layers), the number of stack layers will also be developed to 16hi (16 layers).
It is expected thatHBM412hi will be launched in 2026, with 16hi products arriving in 2027.
It is reported that in HBM4, the bottom layer Logicdie (also known as Basedie) will be used for the first time using 12nm process wafers. This part will be provided by the wafer foundry, so that a single HBM product requires cooperation between the wafer foundry and the memory factory.