Nvidia wants its three main High-Bandwidth Memory (HBM) suppliers to use the same manufacturing process at TSMC. Samsung is reportedly already agreed to this requirement. In doing so, the South Korean manufacturer is deviating from its usual approach for Nvidia, keeping memory chip production entirely in-house.
TNW reports this based on coverage by the South Korean newspaper Seoul Economic Daily. Sources say Nvidia has asked Samsung, SK Hynix, and Micron to use TSMC’s 3nm manufacturing process for its new NVHBM memory chips.
Specifically, this concerns the so-called “base,” the bottom logical layer of an HBM chip. This layer handles communication between the stacked memory chips and the processor. Samsung normally manufactures these components itself but, following an internal evaluation, is reportedly willing to make an exception for Nvidia.
Nvidia wants to minimize differences between suppliers
Nvidia unveiled NVHBM in August. This modified form of HBM is based on HBM4E and is intended to improve the performance of future AI processors. In addition to a common manufacturing process for the “base,” Nvidia wants suppliers to use the same DRAM generation, namely 1c.
With this standardization, Nvidia aims to prevent memory chips from different suppliers from delivering varying performance levels. It also allows the company to simplify the validation process.
With NVHBM, part of the memory controller is moved from the GPU to the base die. This frees up more space on the GPU for compute units. According to the Seoul Economic Daily, the design delivers up to 30 percent more memory bandwidth and 15 percent lower power consumption than standard HBM4E.
Samsung is keeping its own manufacturing process in reserve
The decision to partner with TSMC does not mean that Samsung is giving up its own production capacity. For standard HBM4 and HBM4E chips, the company will continue to use its own 4nm technology. This also applies to customized HBM memory for other customers, including AMD.
Samsung is also exploring whether it can potentially produce the HBM for Nvidia using its own 2nm process. This would provide an alternative should TSMC’s production capacity prove insufficient.
That capacity could become a bottleneck. According to UBS estimates, global demand for HBM will rise by 91 percent next year to approximately 61.5 billion gigabits. Expected production, at 54.6 billion gigabits, will lag behind that figure.
TSMC is therefore expanding its 3nm production capacity, including two additional production lines in Taiwan. These are scheduled to begin mass production in the first half of 2027 and 2028, respectively. The company also plans expansions in the United States and Japan.