Market jitters persist six days after CXMT announcement

Industry calls claim 'declaratory' but concerns linger

Samsung, SK hynix already developing 4F² technology

'They're building an Avante while we build a G90'

China announced a launch preview, not mass production

DUV equipment use expected to drag down yields

The headquarters of CXMT, a Chinese memory chip manufacturer, in Hefei, Anhui Province, eastern China. [Xinhua]
The headquarters of CXMT, a Chinese memory chip manufacturer, in Hefei, Anhui Province, eastern China. [Xinhua]

Aftershocks from CXMT's announcement that it plans to launch a next-generation server DRAM chip based on 4F² (4F-squared) architecture before year's end have now stretched into their sixth day.

Semiconductor industry insiders and academics have largely dismissed the announcement as "declaratory," yet the shock waves rippling through global financial markets show little sign of fading.

Experts warn that characterizing CXMT as having leaped "two years ahead" of Samsung Electronics and SK hynix conflates a product launch preview with actual mass production — a confusion they say has fueled an outsized sense of crisis. Significant technical hurdles remain before stable mass production becomes a reality, they say.

Still, concerns are mounting that the Chinese chip offensive is advancing faster than expected and could close the gap with Korean semiconductor makers more quickly than anticipated. Analysts say the market was particularly rattled by the implication that China may be able to challenge the industry leaders even without extreme ultraviolet (EUV) lithography equipment — the key tool for shrinking chip geometries.

According to industry sources, Samsung Electronics had already unveiled its "4F²-based vertical cell transistor (VCT) with hybrid bonding" technology in February at the International Solid-State Circuits Conference (ISSCC), organized by the Institute of Electrical and Electronics Engineers (IEEE) and considered the world's most prestigious forum for solid-state circuit research. The presentation outlined a path to overcoming DRAM scaling limits and realizing next-generation memory.

Song Jae-hyuk, president and chief technology officer of Samsung Electronics' semiconductor division, reaffirmed Wednesday that the company is developing 4F² technology, speaking at the 28th Semiconductor Exhibition (SEDEX 2026) held at COEX in Gangnam-gu, Seoul.

SK hynix had also publicly shared its 4F² development roadmap at the VLSI Symposium 2025, organized by IEEE, in the United States last June.

The "F" in 4F² stands for feature size — the minimum line width of a semiconductor. In DRAM, data is stored in individual cells, and the area each cell occupies is expressed as a multiple of F².

Today's standard 6F² DRAM uses a rectangular cell measuring 3F by 2F, while a 4F² cell is a square measuring 2F on each side. The 33 percent smaller footprint allows 30 to 50 percent more cells to be packed onto the same chip without shrinking the line width — a productivity and cost advantage that has made 4F² one of the most closely watched next-generation memory technologies.

A CXMT booth at the 21st China International Semiconductor Expo held in Beijing on Nov. 20, 2024. [AP]
A CXMT booth at the 21st China International Semiconductor Expo held in Beijing on Nov. 20, 2024. [AP]

Even though Samsung Electronics and SK hynix have both officially confirmed they are developing 4F², markets reacted with unusual sensitivity to the announcement from CXMT, the late-coming challenger, sending shares sharply lower.

Experts say CXMT's mention of a "year-end launch" signals that the company has moved beyond the research stage into product development — but they stress that any proper assessment must wait until the actual chip is in hand and can be rigorously tested.

Jeon Byung-seo, director of the China Economy and Finance Research Institute, put it in automotive terms. "What we make is a Genesis G90; what CXMT says it will make is an Avante," he said. "It has not even been proven that CXMT actually possesses the technology. This is nothing more than a declared roadmap for what they intend to do by 2028."

CXMT President Cao Kanyu said at the Fourth Chiplet and Heterogeneous Integration Conference in Shanghai on Saturday only that the company would "unveil" a next-generation server memory product — a DDR5 RDIMM incorporating 4F² and hybrid bonding — by year's end. He used neither the words "mass production" nor "sample shipment," and disclosed no specific specifications or release date.

Experts say the announcement should be read as a launch preview, not a declaration of mass production of 4F²-based DRAM. They estimate it will take at least two more years before the product can be manufactured at scale.

While 4F² is widely regarded as a technology that could dramatically boost semiconductor industry productivity, its process complexity is equally formidable — Samsung Electronics and SK hynix themselves do not expect to adopt it until after 2028. Doubts are also surfacing over whether CXMT has the technical capability to reach actual mass production.

Industry observers expect yields — the share of defect-free chips produced — to be low for any CXMT 4F²-based DRAM, given that China continues to struggle to obtain advanced semiconductor manufacturing equipment under US export controls.

Kim Hyung-jun, director of the Next-Generation Intelligent Semiconductor Project and an emeritus professor at Seoul National University, said Chinese companies are forced to use deep ultraviolet (DUV) equipment, which requires multiple patterning passes to etch circuits. "That inevitably drives yields down," he said, adding that the resulting chips would be less competitive than the 6F²-based DRAM that Samsung Electronics and SK hynix currently produce.


joze@heraldcorp.com
jeongwan@heraldcorp.com