Samsung adopts TSMC process for Nvidia’s custom HBM
Nvidia gains tighter control over memory supply consistency across suppliers
Overview
Nvidia is reportedly asking Samsung, SK hynix and Micron to use ’s for the logic in its custom memory. Each supplier would still use its own for the stacked core dies. The connects and manages the memory layers, while the core dies store data. Using the same could make performance more consistent and make Nvidia’s testing and approval easier.
Samsung has reportedly agreed to use a rival’s process while developing with Nvidia. It would still use its own processes for other HBM products. The plan could improve supply reliability, but it would make a common link in Nvidia’s AI-chip supply chain. Limited 3-nanometre capacity could restrict output and give more influence, despite plans to add production lines. The companies had not confirmed the terms publicly when this was published.
Timeline
The reported NVHBM plan made TSMC a common manufacturing dependency for Nvidia’s three memory partners
Samsung’s reported decision completed the picture of a common base-die strategy spanning Samsung, SK hynix and Micron. Nvidia would gain a more uniform logic layer and simpler , while the memory companies would remain responsible for producing the . The reported arrangement therefore shifts part of the AI-accelerator memory supply chain from independent memory-process choices toward Nvidia-led co-design and shared TSMC manufacturing.
Samsung reportedly kept its own processes for other HBM products
The reported TSMC arrangement applies to Samsung’s NVHBM work, not necessarily to its broader HBM manufacturing. Samsung was reported to be continuing in-house development of other HBM base dies, including work associated with a and . The contrast indicates that Samsung could use TSMC for Nvidia’s jointly specified product while retaining internal processes for other customers and products.
Samsung reportedly accepted TSMC’s 3-nanometre process for NVHBM base dies
Samsung was reported to have agreed to manufacture s using TSMC’s . The arrangement would make Samsung jointly develop with Nvidia while using a rival foundry’s process for the logic layer. The report was attributed to industry sources, and the cited coverage did not identify a public confirmation from Samsung, Nvidia or TSMC.
Nvidia reportedly tapped Samsung, SK hynix and Micron for its NVHBM program
Contemporary reporting identified Samsung, SK hynix and Micron as the three memory suppliers involved in Nvidia’s reported . The reports did not independently establish a formal production-partner selection or the program’s stated competition and standardization goals.
Micron reportedly planned to use TSMC’s 3-nanometre process for NVHBM base dies
Reporting said Micron planned to NVHBM base-die production to TSMC using its 3-nanometre process, alongside sixth-generation for the core dies. The available reports do not establish that Micron formally agreed before Samsung or confirm the stated October 1 timing.
Nvidia reportedly requested TSMC-made base dies for NVHBM
Reports said Nvidia asked Samsung, SK hynix and Micron to use TSMC’s 3-nanometre process for NVHBM base dies, with Samsung reportedly accepting the arrangement. The reports do not establish that every NVHBM stack will use TSMC-made dies or that capacity constraints will limit supply.
Nvidia reportedly asked all three memory suppliers to standardize NVHBM base dies on TSMC’s 3-nanometre process
Nvidia reportedly requested that Samsung, SK hynix and Micron use for the NVHBM logic base die. The same request reportedly specified for the stacked core dies. A shared logic process would reduce supplier-to-supplier variation and allow Nvidia to qualify a more consistent custom-memory design.
Nvidia separated the NVHBM logic base die from the stacked DRAM core dies
In an HBM stack, the is the bottom logic layer that connects and manages the vertically stacked DRAM, while the core dies hold the memory cells. Nvidia’s reported design allows the logic and to use different manufacturing processes. That separation lets Nvidia standardize a critical logic component while suppliers continue producing the memory layers.
Nvidia unveiled NVHBM as a custom memory program for AI accelerators
Nvidia introduced as a customized form of developed with memory suppliers rather than relying only on standard HBM products. The initiative gives Nvidia greater influence over the memory interface, logic and qualification process used alongside its AI accelerators. The three major HBM suppliers identified for the program were Samsung, SK hynix and Micron.