Starting with sixth-generation high-bandwidth memory (HBM4), the calculus of the three global DRAM makers is diverging. That is because the manufacturing process for the base die—the "brain" at the very bottom of the HBM stack that controls multiple DRAM layers—is shifting from a conventional DRAM process to an advanced logic process. As of the 10th, industry sources said Samsung Electronics is pursuing a strategy to strengthen its market position with "absolute performance," SK hynix is seeking a balance between performance and stability, and Micron is prioritizing stability but is seen as a step slower in the shift than its rivals.
HBM is a high-value memory that vertically stacks multiple DRAMs to maximize data processing speed (bandwidth), and it is a key component that determines the performance of artificial intelligence (AI) accelerators made by companies such as Nvidia. In previous generations, the base die at the very bottom of this stack was also made using processes similar to or more stable than those for general DRAM. But starting with HBM4, advanced processes used to make smartphone application processors (AP) or server central processing units (CPU) began to be applied to the base die. As how fast and efficiently data is controlled comes to decide overall HBM performance, the base die's completeness itself has emerged as a new battleground for the three memory makers.
The finer a semiconductor's line width—in other words, the smaller the nanometer figure—the faster it runs and the less power it consumes. However, manufacturing becomes that much trickier, bringing higher defect rates along with the benefits. Samsung Electronics chose a strategy to push performance to the limit by applying its finest 4-nanometer process to the base die. SK hynix is starting for now with the mature 12-nanometer process and plans to step down to the 3-nanometer class later, while Micron is focusing on stable supply rather than performance gains.
Samsung Electronics applied its own foundry's 4-nanometer process to the base die as is, choosing an architecture in which one company makes DRAM and logic from start to finish. The biggest advantage is considered the ability to coordinate everything in-house, from design and process to packaging. In fact, Samsung Electronics is highly likely to supply most of the high-end HBM4 products provided to Nvidia. However, risks are also cited because there is no precedent for a DRAM company raising an advanced logic process to its own Production yield at a smartphone AP level. The fact that Samsung Electronics' 4-nanometer foundry process has lagged TSMC in bids for major customers such as Apple and Qualcomm also supports such concerns.
SK hynix chose to secure stability by applying TSMC's 12-nanometer process to HBM4. A key strength is the ability to use a process already validated through numerous customers, including Apple, Nvidia and AMD, as is. The industry expects a shift to the 3-nanometer class starting with HBM4E (seventh-generation HBM), but it is not yet finalized. However, from TSMC's perspective, HBM base dies are relatively small in volume compared with logic chips for smartphones and AI accelerators, raising concerns that production schedules could be pushed back when large customer orders pile up.
Micron is understood to have stuck to its own low-cost process and only recently begun reviewing advanced logic processes. This amounts to a strategy that prioritizes cost, but the industry views the company as relatively behind Samsung Electronics and SK hynix in base die technology. Collaboration also takes a back seat compared with SK hynix, which moves in near "one-team" fashion with TSMC.
A semiconductor equipment industry source said, "In theory, Samsung Electronics' 4-nanometer base die can outperform SK hynix's 12-nanometer in speed and power efficiency, but because there is no precedent for a DRAM company raising an advanced logic process to its own Production yield, it will take more time for yield to stabilize at the actual mass production stage." The securities market noted, "SK hynix's strategy of starting with a validated process to lower risk is stable in the short term, but as the shift proceeds to HBM4E and HBM5 (eighth-generation HBM), the speed of transitioning to finer processes will be the key," adding, "The superiority of the nanometer strategy is more likely to be decided not in this generation but in the competition to secure volume in the next generation."