Samsung will reportedly use TSMC's 3 nanometer process for the base die of Nvidia's next custom high bandwidth memory

Context

This is secondhand: Polymarket is relaying a Seoul Economic Daily report that cites industry sources, which Reuters also carried. The report says TSMC will make the base die while Samsung keeps making the memory itself, a narrower arrangement than Polymarket's teaming up framing, and none of the companies has confirmed it.

Deep dive

On October 18, Polymarket posted that Samsung and TSMC are reportedly teaming up to produce Nvidia's next generation high bandwidth memory chips.

The Seoul Economic Daily reported that day, citing industry sources, that Samsung Electronics will use TSMC's 3 nanometer process to make the base die for NVHBM, the custom high bandwidth memory Nvidia unveiled in August. The base die sits under the stacked memory and works as a logic chip. The paper said Nvidia asked all three of its NVHBM suppliers, Samsung, SK hynix and Micron, to make the base die on TSMC's process and to standardize the memory dies on sixth generation DRAM, known as 1c.

According to the report, Nvidia made the request because different manufacturing methods across suppliers make it hard to get consistent performance from one design, and standardizing reduces the work of validating each supplier's product. It said Samsung's chip division accepted after an internal review, and that this would be the first time Samsung and TSMC have partnered at the HBM manufacturing stage.

Samsung has so far made both the memory and base dies for all of its HBM, using its own 4 nanometer foundry process for HBM4. The paper said it will keep that approach for standard HBM and for custom memory for other customers such as AMD, and is checking whether its own foundry could also make NVHBM base dies in case demand exceeds TSMC's capacity. Reuters carried the report, attributing it to the Seoul newspaper.

What to expect next

Watch for confirmation from Samsung, TSMC or Nvidia, and whether SK hynix and Micron also move their NVHBM base dies to TSMC as the report says Nvidia asked.

Also watch whether TSMC's 3 nanometer capacity becomes a new bottleneck for AI memory, given that the report says Samsung is preparing its own foundry as a backup. Memory prices, already driving up costs for phones and PCs, could be affected by how quickly NVHBM ramps.