Nvidia wants the brain of its memory built somewhere specific. Seoul Economic Daily reported on Oct 18 that the three companies making NVHBM, Samsung Electronics, SK hynix and Micron, have been directed to produce the base die on TSMC’s 3nm logic process. For Samsung it is a first: the memory division has never depended on TSMC for an HBM part.
The base die sits at the bottom of every HBM stack and carries the circuits that shuttle signals between memory and processor. Under NVHBM, the memory controller that used to live on the GPU moves down into that base die. Nvidia claims bandwidth gains of up to 30 percent and 15 percent lower power against standard HBM4E, plus reclaimed area on the compute die.
That relocation is why the process node changed. Squeezing a controller into the floor of a memory tower takes fine geometry and, just as importantly, good yield. Industry estimates put TSMC’s 3nm output above 90 percent, while Samsung Foundry’s 2nm class process is judged near 60 percent. Nvidia also gets one verification path instead of three.
The squeeze is real. UBS sees HBM demand next year at roughly 61.5 billion gigabits, up 91 percent, against shipments of about 54.6 billion. TSMC is adding two 3nm lines at Fab 18 in Tainan to absorb base die orders.
Samsung keeps its turnkey pitch for catalog HBM, where its own 4nm base dies still do the job. NVHBM splits that stack in two, and the memory makers no longer own the half that differentiates.
