The Pulse
Samsung Turns to TSMC’s 3nm Process for Nvidia HBM
Samsung plans to use TSMC’s 3-nanometer process for the base die in Nvidia’s custom NVHBM, according to Seoul Economic Daily. The reported arrangement would split Samsung’s memory production from base-die fabrication and standardize the des
AI.info Team · Written with AI ·
Samsung will use TSMC’s 3-nanometer process to make the base die for Nvidia’s custom high-bandwidth memory, Seoul Economic Daily reported on October 18, citing industry sources. The reported arrangement puts a rival foundry in the production path for Samsung’s NVHBM while Nvidia seeks to standardize the memory components it receives from several suppliers.
The base die is the logic layer beneath the stacked memory dies. Samsung has until now made both parts of its HBM products itself. For NVHBM, the South Korean company is set to supply the memory while TSMC fabricates the base die, according to the newspaper’s report.
Nvidia sets a shared design for three suppliers
Nvidia has asked Samsung, SK hynix and Micron to make NVHBM base dies using TSMC’s 3nm process, SED reported, citing unnamed industry sources. The company also wants the suppliers to use the same sixth-generation DRAM process, known as 1c, for the core dies. The report says Samsung’s Device Solutions division accepted Nvidia’s request after an internal review.
The requested uniformity addresses a practical problem: different manufacturing methods can make it harder to deliver consistent performance from a common design. Nvidia’s specification also means it can validate products from the suppliers against a more consistent set of manufacturing assumptions, rather than adapting its qualification work to each supplier’s process.
Nvidia introduced NVHBM in August as a memory design with its custom memory controller integrated into the HBM base die rather than the accelerator die. The company says the approach can provide up to 30% greater memory bandwidth and 15% lower HBM power consumption than standard HBM4E, while freeing up to 25% more area on the XPU compute die. Those are Nvidia’s stated comparisons, not performance results reported for Samsung’s planned product.
Samsung keeps its foundry option open
The reported shift applies to NVHBM, not Samsung’s entire HBM business. SED says Samsung plans to keep its integrated production approach for standard HBM products, using its own 4nm foundry process for the base dies in HBM4 and HBM4E. The company also intends to retain that approach for custom HBM made for other customers, including AMD.
Samsung has not ruled out making NVHBM base dies in-house as well. An unnamed industry official told SED that Samsung is checking whether its own foundry can produce the part, as a hedge if demand exceeds TSMC’s capacity. The report does not give a production schedule or disclose the expected volume.
The companies have worked together before on compatibility between Samsung HBM and TSMC’s packaging process, SED reported, but described the NVHBM arrangement as their first cooperation at the HBM manufacturing stage. The new split gives Nvidia a common base-die process across suppliers while leaving Samsung with a possible second production route.
Nvidia’s public pitch centers on standardization
Nvidia’s August announcement said, “NVIDIA is establishing a standard NVHBM implementation, available from multiple memory providers.” The company said the standard is intended to reduce the engineering and qualification work needed to integrate memory from different suppliers. Amazon’s Annapurna Labs was named as the first partner to work on NVHBM alongside NVLink Fusion.
That public announcement did not identify Samsung, SK hynix, Micron or TSMC as participants in the specific manufacturing arrangement described by SED. The October report relies on unnamed industry sources; it does not include public confirmation from Nvidia, Samsung or TSMC. The reported plan therefore outlines a proposed supplier configuration, not a publicly announced agreement among the three companies.
For Samsung, the arrangement trades some of its end-to-end manufacturing control for access to Nvidia’s custom-memory program. Its own foundry remains a possible fallback, but the report leaves unanswered whether that route will be qualified, how much NVHBM TSMC will produce, and when the first products will ship.