TLB 開始為 AI 伺服器開發 SOCAMM3 基板

TLB’s SOCAMM substrate. The upper image shows the front side where DRAM packages are mounted, while the lower image shows the rear contact pads that connect to the server board. Measuring 90mm by 14mm, it is considerably smaller than a conventional server registered dual in-line memory module (RDIMM), which measures 133.35mm by 31.25mm. (Photo: TLB)

TLB, a specialist in memory substrates, has begun developing substrates for SOCAMM3 (Small Outline Compression Attached Memory Module 3), a next-generation low-power memory module standard for artificial intelligence (AI) servers, in collaboration with two major memory manufacturers in South Korea and overseas, according to industry sources.

The company is participating in development of the successor standard from an early stage as sales of its SOCAMM2 substrates, which entered mass production early in 2026, surge. Industry observers expect TLB to maintain its position in the specialized AI server memory substrate market into the next generation.

According to industry sources on Oct. 22, TLB recently began discussions with two existing memory customers on SOCAMM3 substrate design and manufacturing processes. The standard has not yet been finalized. Memory manufacturers are expected to develop substrate designs with suppliers before incorporating them into the standard. Mass production of the substrates is expected in 2028.

During development, memory manufacturers specify not only the basic substrate design but also the suppliers of copper-clad laminates (CCLs). Samsung Electronics, SK hynix and Micron select CCLs that meet their respective specifications from Doosan, Taiwan’s EMC and ITEQ, finalizing their choices during the development stage.

SOCAMM is a standard developed to enable low-power double data rate (LPDDR) memory, primarily used in mobile devices such as smartphones, to be installed in AI servers as removable modules. Nvidia initially developed the format for its AI servers. SOCAMM2 subsequently became a standard under semiconductor standards organization JEDEC. Nvidia’s Vera CPUs currently use eight LPDDR5X-based SOCAMM2 modules per CPU, providing up to 1.5TB of memory capacity and 1.2TB per second of bandwidth.

SOCAMM2 substrates undergo five buildup cycles, in which circuit layers are stacked and connected through microscopic holes. SOCAMM3 substrates are expected to require more buildup cycles than SOCAMM2. Additional layers increase manufacturing complexity, particularly in processes such as forming microvias that connect the layers.

“SOCAMM3 will likely be more expensive because it requires more buildup layers than the previous generation,” an industry official said. “For substrate manufacturers, this creates an opportunity to increase revenue.”

The systems that will use SOCAMM3 have not yet been disclosed. However, the expected 2028 production timeline aligns with Nvidia’s roadmap. At its GTC conference in March 2026, Nvidia announced plans to introduce its Feynman graphics processing unit (GPU) and a new central processing unit (CPU), Rosa, in 2028. SOCAMM3 could be adopted when the CPU changes in 2028.

Supply in the SOCAMM market is struggling to keep pace with demand. Market research firm TrendForce reported in June 2026 that Nvidia had halved the originally planned memory capacity per SOCAMM module for Vera Rubin. The three major memory manufacturers had allocated LPDDR5X supplies equivalent to only about 60% of Nvidia’s requirements. TrendForce expects shortages to persist into 2027.

SOCAMM2 substrates are already significantly boosting TLB’s financial performance. Production volume and revenue more than tripled in the second quarter from the previous quarter. Because SOCAMM2 substrates require repeated buildup processes, their production time is more than twice that of conventional memory module substrates. Their selling prices are also approximately double.

Industry estimates put TLB’s SOCAMM substrate revenue at around 70 billion won in 2026, with projections of approximately 140 billion won in 2027. Securities firms expect TLB’s total revenue to increase about 40% year-on-year to between 350 billion won and 370 billion won in 2026. Revenue could rise to around 450 billion won in 2027, driven by SOCAMM growth and capacity expansion.

Anticipating rising demand, TLB began investing in buildup equipment at its second Ansan plant and Vietnam facility in the third quarter of 2025. The investments came approximately one year before competitors recently began expanding their related production capacity.

The company plans to begin construction of a second factory in November at the Yen Phong Industrial Park in Bac Ninh Province, Vietnam. The new facility will be built next to its existing first plant and will be twice its size. Full-scale operations are targeted for the fourth quarter of 2027. The second plant will house equipment for hole drilling, copper plating, circuit formation and surface treatment, processes not available at the first plant. Previously, the first plant handled downstream processes such as profiling, inspection and packaging, along with buildup processes added in 2026, while intermediate processes were performed at TLB’s Ansan headquarters. Once the second plant begins operating, the company will be able to handle the entire manufacturing process in Vietnam.

With the second plant operational, TLB’s effective company-wide production capacity will double from 20,000 square meters to 40,000 square meters per month. By 2028, the company will have a production base theoretically capable of generating between 700 billion won and 800 billion won in annual revenue.

However, SOCAMM substrates are expected to remain in production exclusively at TLB’s Ansan headquarters for the time being because of their manufacturing complexity and stringent customer qualification requirements.