SK Hynix Bets on EUV to Extend 6F² Architecture While CXMT Seeks Breakthrough with 4F²
nashnova research
The three largest DRAM makers have split on cell architecture — SK Hynix and Micron stick with 6F², while Samsung and China's CXMT bet on 4F². The real question: does shrinking the cell actually cut costs?
What are 6F² and 4F² — and why does this matter?
6F² and 4F² describe how much chip area a single DRAM storage cell occupies, where F is the smallest feature the manufacturing process can print. A 4F² cell is one-third smaller than 6F² at the same feature size.
This means → in theory, 4F² packs more storage cells onto the same wafer, raising bit density.
But SK Hynix flagged a critical caveat: a smaller cell does not automatically lower cost. Capacitor pitch — the physical gap between adjacent storage capacitors — process complexity, and manufacturing yield all shape the final effective density and cost.
In plain terms = drawing smaller boxes on the wafer does not guarantee every box works; and drawing smaller boxes is itself harder.
Why do SK Hynix and Micron refuse to go 4F²?
On its Q3 earnings call, SK Hynix stated plainly: it will extend 6F² using EUV lithography. 4F² and 3D DRAM stay on the roadmap but have no production timeline.
Micron's stance is even firmer. Technology chief Scott DeBoer said the company has studied 4F² for nearly twenty years and concluded the scaling gains are limited — DRAM density is fundamentally constrained by capacitor pitch, so a smaller cell layout may not deliver the expected overall density lift.
This means → both giants reached the same judgment: under current process conditions, 6F² plus EUV offers better cost-performance. The theoretical advantage of 4F² is not large enough to justify the process risk.
Micron's long-term plan is to skip 4F² entirely and move straight to 3D DRAM — stacking storage cells vertically rather than shrinking them on a flat plane.
Why is Samsung betting on 4F² instead?
Samsung CTO Jae-hyuk Song confirmed at SEDEX 2026 in Seoul: the company is developing a 4F² architecture.
According to The Elec, Samsung plans to complete its first 10a-node 4F² DRAM in 2026, run quality tests in 2027, and begin mass production in 2028 — though Samsung itself has not confirmed this timeline.
This reflects a strategic divergence: as the only top-three maker actively pushing 4F² toward production, Samsung may be betting that the first mover to commercialize 4F² captures a density edge.
SK Hynix, by contrast, reportedly plans to introduce 4F² only at the later 10b node — a notably more conservative pace.
How is CXMT's 4F² path different?
CXMT faces a hard constraint: export controls block its access to EUV lithography equipment.
This means → the mainstream 6F²-plus-EUV path is closed to CXMT. 4F² becomes a necessary alternative, not merely a preference.
CXMT president Mark Cao disclosed plans to launch a 4F²-based server DDR5 RDIMM product by the end of 2026, using an in-house vertical surrounding gate transistor (VSG — a transistor structure where current flows vertically), all built on existing DUV lithography — the generation below EUV.
In plain terms = others use the most advanced etching tool to make cells smaller; CXMT uses the previous-generation tool with a different cell design, aiming for comparable density.
Can CXMT's approach actually work?
CXMT has not yet disclosed verified commercial mass-production yields, customer qualification results, or volume shipment data.
This means → the technical blueprint is public, but the critical proof points between lab and factory floor remain blank.
Whether CXMT can ship on schedule and pass server-customer qualification is the hard gate — server memory demands far higher reliability than consumer-grade products.
What does this architecture split tell us?
Across the three majors plus CXMT, three distinct technology paths have emerged: 6F² plus EUV (SK Hynix / Micron), 4F² production (Samsung / CXMT), and a direct leap to 3D DRAM (Micron's long-term plan).
This reflects a rare inflection point in DRAM — no single path has been proven the definitive winner.
In plain terms = for decades, DRAM makers walked roughly the same road. Now the road forks three ways, and each company is betting its chosen fork leads somewhere first.
For CXMT specifically, 4F² is not just a technology choice — it is a survival strategy to stay competitive under an equipment blockade.
市场有风险，内容仅供研究参考，不构成投资建议。