Samsung Adopts Wafer Bonding for HBM5, Outlines Triple-Track DRAM Roadmap
nashnova research
Samsung has confirmed three parallel DRAM development tracks beyond D1c, with HBM5 set to use wafer-to-wafer hybrid bonding for the first time. The three-track split is fundamentally a risk hedge — breaking apart the hardest technology changes so they don't all have to land at once.
What does each track do?
D0A (targeting HBM5): bonds a memory-cell wafer built on D1b process to a peripheral-circuit wafer built on D1d process using hybrid bonding — a technique that presses two wafers copper-to-copper so they work as a single chip. This means → memory and control circuitry can each use the process best suited to it, instead of compromising on one.
D1d (targeting mainstream mobile memory): keeps the existing 6F² cell design with memory and circuitry on a single wafer, aimed at LPDDR and other low-power products. In plain terms = this is the steady, low-risk track that protects Samsung's volume shipments.
D0AV (targeting future high-performance products): uses a smaller 4F² cell and introduces vertical-channel transistors (VCT — rotating the transistor channel from horizontal to vertical to handle shrinking dimensions), with a 2028 target node.
Why not combine wafer bonding and 4F² in one step?
Samsung considered doing exactly that, but stacking multiple major technology changes raised engineering complexity to unacceptable levels. This means → a failure in any single element would stall the entire production line.
The final strategy: land 6F² wafer bonding first (D0A, H2 2027), then branch 4F² and VCT onto a separate track (D0AV, 2028). In plain terms = prove the new packaging with a mature cell design first, then load the new cell onto the proven package — risk in two steps, not one.
At ISSCC 2026 in February, Samsung researchers demonstrated a 16 Gb test chip combining 4F²+VCT+hybrid copper bonding. The paper reported a >20% increase in die yield per wafer under the same design rules. This reflects early proof of technical feasibility, though the gap from lab to mass production remains.
What are Micron and CXMT doing differently?
Micron: President and CTO Scott DeBoer said on October 21 that Micron evaluated 4F² for over twenty years and concluded the gains do not justify the added complexity and cost. Micron plans to bring its 1δ process to mass production in H2 2027, with wafer bonding introduced in the subsequent 1ε generation.
CXMT (长鑫存储): chose 4F² as its lead architecture for the server market, announcing on October 30 that it completed engineering-sample validation on its sixth-generation DRAM platform. This means → the Chinese player has taken a path opposite to Micron's and partially overlapping with Samsung's.
In plain terms = three major memory makers faced the same fork and gave three different answers — Samsung wants both tracks, Micron wants bonding but not 4F², and CXMT bets on 4F² first.
Why is H2 2027 the critical window?
Samsung's D0A and D1d, plus Micron's 1δ process, all target H2 2027 for production ramp. This means → all three submit their answers almost simultaneously; whoever hits yield and volume targets first gains the upper hand in HBM and high-end DRAM.
For Samsung specifically, whether the HBM5 wafer-bonding track delivers on schedule will determine if it can close the gap with SK Hynix in HBM.
This reflects a broader shift: the memory industry is moving from a single race on process shrink to a three-dimensional competition across process + packaging + cell architecture, and the H2 2027 production ramp is the first real proving ground.
市场有风险,内容仅供研究参考,不构成投资建议。
