Nvidia shifts NVHBM production to TSMC 3nm to boost bandwidth and cut power consumption
Nvidia mandates that Samsung, SK Hynix, and Micron use TSMC’s 3nm process for base dies in its custom high-bandwidth memory chips to ensure performance consistency.
Key points
- Nvidia has required all three NVHBM suppliers to manufacture base dies using TSMC’s 3-nanometre process, marking a shift from Samsung’s traditional turnkey approach.
- The move aims to standardize production, allowing Nvidia to validate chip performance uniformly across suppliers rather than testing each variant separately.
- NVHBM architecture moves the memory controller to the base die, enabling up to 30% more bandwidth and 15% less power consumption compared to standard HBM4E.
Nvidia has instructed its memory partners to use TSMC’s advanced 3-nanometre manufacturing process for the logic layer of its new custom high-bandwidth memory chips. According to a report by The Next Web, citing the Seoul Economic Daily, this requirement applies to all three of Nvidia’s NVHBM suppliers: Samsung, SK Hynix, and Micron. This decision breaks with previous industry norms where suppliers handled the entire production stack internally. The change ensures that the foundational logic component of these critical AI accelerators is built on a single, standardized process node, rather than varying by manufacturer.
In plain English
High-bandwidth memory, or HBM, is a specialized type of memory designed to feed data to graphics processors at extremely high speeds, which is essential for artificial intelligence workloads. An HBM module consists of multiple layers of memory stacked vertically. At the bottom of this stack sits the "base die," which acts as the logic layer, managing how data moves between the memory and the processor.
Traditionally, companies like Samsung manufactured both the memory layers and this base die themselves, a method known as a turnkey approach. Nvidia’s new directive forces suppliers to source this specific logic layer from TSMC, creating a hybrid manufacturing model for this specific product line.
The background
Samsung Electronics had previously produced both the base die and the memory dies for all its HBM products using its own manufacturing processes. For Nvidia’s custom NVHBM, unveiled in August, Samsung agreed to switch to TSMC’s 3nm process for the base die following an internal review. This marks the first time Samsung and TSMC have collaborated directly on the production of high-bandwidth memory components. The two companies did work together four years ago, when Samsung adjusted its HBM design to fit TSMC’s packaging requirements, but this is the first instance of joint component manufacturing.
What changes now
Nvidia’s requirement for a common 3nm base die and standardized 1c DRAM technology aims to eliminate performance inconsistencies that arise when suppliers use different manufacturing methods. By standardizing the process, Nvidia can validate the chips once, rather than testing each supplier’s unique combination of components separately. The NVHBM design also moves the memory controller from the GPU to the base die, freeing up space on the graphics processor for additional computing circuits. This architectural shift results in a physical layer that is 60% smaller than standard HBM, delivering up to 30% more bandwidth and 15% less power usage.
Key terms explained
Supply chain standardization involves aligning manufacturing processes across multiple vendors to reduce variability. In semiconductor production, using a single process node for critical components like base dies helps ensure that hardware behaves predictably, regardless of which supplier manufactures it. Teams usually monitor process node transitions closely, as shifts in manufacturing partners can impact yield rates and delivery timelines. Checking whether a vendor has validated compatibility with new process nodes is a standard step in procurement. This approach minimizes integration risks and simplifies quality assurance procedures for complex, multi-layered components.
Background: Supply chain attacks
Supply chain attacks compromise your systems indirectly. An adversary infiltrates a trusted vendor or software update. When you install that legitimate-looking update, the malware executes inside your network. This method bypasses perimeter defenses because the code carries a valid digital signature from a reputable source.
Read the full guide: What Is Supply Chain Attacks: How Indirect Compromise Works
What to do and how to stay safe: Nvidia
- Review current hardware procurement contracts to identify clauses related to manufacturing process nodes and supplier exclusivity.
- Monitor vendor announcements regarding changes in base die manufacturing partners, as these shifts can impact long-term availability and support.
- Evaluate the performance specifications of new memory modules against existing infrastructure to ensure compatibility with standardized components.
- Assess the impact of supply chain consolidation on risk management strategies, particularly regarding dependency on single-source process technologies.
General security guidance from the Vector Update newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
Why is Nvidia requiring TSMC's 3nm process for NVHBM base dies?
Nvidia requires the common process to ensure consistency in chip performance across all suppliers and to avoid validating each supplier's chips separately, according to the Seoul Economic Daily.
Does this change apply to all of Samsung's HBM products?
No, Samsung will keep its turnkey approach for standard HBM4 and HBM4E, as well as for custom HBM for other customers like AMD, using its own 4nm base dies.
How does NVHBM differ from standard HBM4E in terms of performance?
NVHBM offers up to 30% more bandwidth and 15% less power consumption than standard HBM4E, due to moving the memory controller to the base die and shrinking the physical layer.
