When reading semiconductor news, a curious question often arises: Why do foundry and SoC news (TSMC, Samsung Logic) use straightforward nanometer numbers like “3nm mass production” or “entering 2nm”, while memory (DRAM) news uses confusing letter combinations like “1a, 1b, 1c process transition”?
Why are the naming systems different? And how wide is the actual technology gap with China’s CXMT (ChangXin Memory Technologies), which is rapidly chasing the Korean memory giants?
Here is a simple, clear breakdown.
1. Logic Uses 3nm/2nm, so Why Does DRAM Use 1a, 1b, 1c…?
① Structural Limits: “Transistors vs. Capacitors”
- Logic (SoC/CPU/GPU): Primarily consists of transistors serving as switches. To improve gate control, chipmakers changed the 3D structure (FinFET, GAA/Gate-All-Around), allowing them to continuously shrink the physical feature size.
- DRAM: A single DRAM cell consists of 1 Transistor + 1 Capacitor (charge storage). As feature size shrinks, the capacitor becomes smaller, making it harder to retain electrical charge. Engineers must build the capacitor taller (increasing the aspect ratio), but due to tall-structure limits and current leakage, shrinking DRAM physically by 2–3nm per generation became nearly impossible below 20nm.
② “Marketing Metric” vs. “Actual 10nm-Class Steps”
In reality, ‘3nm’ or ‘2nm’ in Logic node naming does not mean the physical gate length is literally 3 nanometers. It is a marketing metric signaling “this node offers performance and power efficiency equivalent to a generational leap.”
Conversely, as DRAM entered the sub-20nm era, naming precise nanometer steps became ambiguous. The industry adopted the 10nm-class (10nm) label and assigned alphabetical suffixes for each sub-generation.
| DRAM Generation | Meaning / Term | Estimated Feature Size |
| 1x (1st Gen) | 10nm-class Step 1 | ~18 – 19nm |
| 1y (2nd Gen) | 10nm-class Step 2 | ~16 – 17nm |
| 1z (3rd Gen) | 10nm-class Step 3 | ~14 – 15nm |
| 1a / 1-alpha (4th Gen) | 10nm-class Step 4 (EUV fully introduced) | ~13nm |
| 1b / 1-beta (5th Gen) | 10nm-class Step 5 (Core node for HBM3E) | ~12nm |
| 1c / 1-gamma (6th Gen) | 10nm-class Step 6 | ~10 – 11nm |
| 1d / 1-delta (7th Gen) | 10nm-class Step 7 (Preparing sub-10nm transition) | Sub-10nm range |
2. China’s CXMT vs. Samsung & SK Hynix: How Wide Is the Gap?
China is pushing hard for semiconductor self-reliance, with CXMT expanding rapidly. However, compared to industry leaders (Samsung Electronics & SK Hynix), a 2 to 3 generation gap (approx. 3–4+ years) remains in mass-production technology.
[DRAM Mass Production Node Comparison]
Samsung / SK Hynix : [1a nm] ──► [1b nm (HBM3E Core)] ──► [1c nm (Mass Prod.)]
▲
CXMT (China) : [1x / 1y nm] ──► [1z nm / 1a Challenge] ───┘ (Blocked by EUV Sanctions)
① Mainstream Production Node Gap
- Samsung & SK Hynix: Volume production is centered around 1a and 1b nm, with ramp-up and mass production of 1c nm active across leading lines.
- CXMT: Volume output remains heavily reliant on 1x and 1y (16–18nm range) legacy nodes. While CXMT is attempting 1z and 1a-class yield optimization, it lags significantly in die efficiency and defect density.
② The Decisive Barrier: EUV Equipment Sanctions
Transitioning DRAM to sub-1a (4th gen) nodes requires ASML’s Extreme Ultraviolet (EUV) lithography tools. US export controls restrict EUV shipments to China, creating a structural wall for CXMT.
CXMT relies on multi-patterning (DUV DPT/QPT) to achieve fine lines, but multi-patterning sharply increases process steps, driving up wafer manufacturing costs and severely lowering yields.
③ AI Era Bottleneck: HBM & Advanced Packaging
High-Bandwidth Memory (HBM3E, HBM4) requires cutting-edge 1a/1b base dies integrated with Through-Silicon Via (TSV) and advanced packaging/hybrid bonding techniques.
While CXMT poses volume pressure in legacy commodity markets (DDR4/DDR5 & LPDDR4) backed by state subsidies, it lacks the leading-edge nodes and advanced packaging ecosystem needed to compete in the high-value HBM market.
📌 Summary
- Naming Difference: Logic uses performance/power marketing metrics (3nm, 2nm), whereas DRAM uses 10nm-class alphabetical nodes (1x, 1y… 1a, 1b, 1c) due to physical capacitor scaling limitations.
- CXMT’s Challenge: CXMT is aggressive in legacy memory, but faces a hard technological ceiling due to the inability to import EUV equipment.
- Tech Gap: In cutting-edge nodes (1b, 1c) and AI memory (HBM), Samsung Electronics and SK Hynix maintain a solid 2 to 3 generation moat.