Samsung Hands Nvidia HBM Base Dies to Rival TSMC 3nm Line
3 min readSamsung Electronics will let its biggest foundry rival build a critical piece of its next Nvidia memory product. According to the Seoul Economic Daily, Samsung will use TSMC’s 3 nanometer process for the base die of NVHBM, Nvidia’s custom high bandwidth memory. It is the first time the two companies have partnered at the chip manufacturing stage, and it says a lot about where the Samsung TSMC HBM race stands.
What NVHBM Is
High bandwidth memory stacks DRAM layers on top of a base die that handles the connection to the processor. Nvidia announced NVHBM in August as a custom take on seventh generation HBM4E. Its key change is moving the memory controller out of the GPU and into that base die, which turns the base layer into a logic chip. Nvidia claims up to 30 percent more bandwidth per stack and 15 percent lower power than standard HBM4E, with the Feynman GPU platform expected in 2028 as the first full scale user.
Why Samsung Is Giving Up the Base Die
Until now Samsung built both the DRAM layers and the base die of its HBM in house, a turnkey pitch it used against SK hynix, which already relies on TSMC for base dies. Nvidia has asked all three NVHBM suppliers, Samsung, SK hynix and Micron, to standardize the base die on TSMC 3nm and the DRAM layers on the 1c process, so that one design performs the same regardless of who built the stack.
Yield appears to be the deciding factor. Tech Times reports Samsung’s internal review concluded its SF2 process could not match TSMC’s yields at the scale Nvidia demands, with industry estimates putting TSMC 3nm above 90 percent and Samsung’s SF2 around 50 to 60 percent. Samsung will keep its own foundry as a backup in case NVHBM demand outruns TSMC capacity, and it will keep the turnkey model for standard HBM4 and HBM4E and for other custom customers such as AMD.
Why It Matters
Samsung has clawed its HBM share back to about a third of the market from 15 percent a year earlier, and NVHBM is the next prize. Taking Nvidia’s terms keeps Samsung in that race, but it concedes the most advanced logic layer to the rival it most wants to beat in foundry. UBS expects HBM demand to grow 91 percent next year with supply still roughly 10 percent short, so every qualified supplier will sell what it can make. The question for Samsung Foundry is whether it can prove 2nm yields in time for the 2027 commitment window, or watch the base die business settle permanently in Taiwan.