Samsung Electronics plans to use TSMC’s 3-nanometre process for base dies in Nvidia’s custom high-bandwidth memory, Reuters reported on October 18, citing Seoul Economic Daily. The report describes cooperation between chipmaking rivals on NVHBM, a memory design intended to ease bottlenecks in artificial intelligence computing.
According to the newspaper’s industry sources, Nvidia is working with Samsung and TSMC on the base die, the logic layer beneath stacked memory chips that controls signal exchanges. Reuters reported that Nvidia asked Samsung, SK Hynix and Micron to use TSMC’s process for this component and sixth-generation DRAM technology for the memory layers.
Seoul Economic Daily said the design moves the memory controller from the graphics processor to the memory base die. This would free processor space for additional computing circuits. The newspaper presents the arrangement as a response to limits in both memory performance and manufacturing capacity; it is reporting unnamed sources, not a publicly disclosed contract.
Samsung’s existing products provide context for the reported change. In its February HBM4 announcement, the company said it combined sixth-generation 10nm-class DRAM with its own 4nm logic base die. It described bandwidth of up to 3.3 terabytes per second per stack and said commercial shipments had begun. These are Samsung’s product claims, not specifications for the reported NVHBM design.
The distinction matters because manufacturing the memory layers and fabricating the underlying logic are separate parts of the same product. The reported plan would bring TSMC’s process into Samsung’s custom offering for Nvidia. It does not establish that Samsung is abandoning its own base-die technology across its memory portfolio.
The companies’ detailed production commitments and delivery schedule remain unconfirmed in the material reviewed. Further disclosures would be needed to establish when the reported arrangement reaches commercial supply. For now, the October 18 reporting points to closer coordination between memory suppliers and a foundry as Nvidia pursues a customised memory architecture.
Comments