Samsung turns to TSMC for part of Nvidia’s custom memory

Nvidia unveiled its own design, NVHBM, in August. It has now asked its three suppliers, Micron, SK hynix and Samsung, to make the bottom logic layer of each stack (the base die) on TSMC’s 3-nanometre process. The memory layers above must use the same sixth-generation DRAM process. One shared process gives Nvidia matching parts from all three suppliers.

For Samsung this is a first. It has always made both layers itself, and until now any HBM work with TSMC stopped at packaging, the paper says. For standard HBM and other customers’ custom chips, such as AMD’s, Samsung still does everything itself.

Nvidia says NVHBM moves up to 30 per cent more data than standard HBM4E memory, Wccftech reports. It also uses up to 15 per cent less power. If TSMC cannot make enough base dies, Samsung is checking whether its own factories could fill the gap.

Memory is one of the scarcest parts of an AI server, and this decides how Nvidia’s next custom supply of it gets built.