[News] SOCAMM2 War Heats Up: Samsung Reportedly Delivers Samples to NVIDIA, Ramping Early 2026

While Micron confirmed at its earnings call that it has sampled its 192GB LP SOCAMM2, other memory leaders are ramping up their efforts. Yonhap News reports that Samsung has also developed SOCAMM2, an LPDDR-based server memory module, and has already delivered samples to AI powerhouse NVIDIA, with strong interest reportedly coming from other major customers.

The report highlights that SOCAMM2 is expected to power NVIDIA’s next-gen AI chip, Vera Rubin. Meanwhile, ZDNet suggests Samsung’s SOCAMM2 is expected to enter full-scale commercial production in early 2026, adding that it has developed the module based on 1b (5th-generation 10nm-class) DRAM and is currently conducting quality tests.

ZDNet explains that the SOCAMM2 (Small Outline Compression Attached Memory Module) packs four LPDDR chips per module, delivering over double the bandwidth of traditional RDIMMs while slashing power consumption by 55%. Its modular design enables easy replacement and upgrades—eliminating the need to solder LPDDR directly onto motherboards, the report adds.

As per the Korea Herald, while HBM stacks DRAM chips vertically, SOCAM combines multiple low-power DRAM (LPDDR) modules. The report notes that HBM attaches to NVIDIA GPUs to boost computational speed, whereas SOCAM integrates with CPUs to cut power consumption.

SK hynix, Micron Ramp up Efforts

In September, Etoday and DealSite indicate that NVIDIA has placed next-year SOCAM module orders with Samsung, SK hynix, and Micron, likely distributing allocations evenly across the three, alongside Rubin deployments. If confirmed, the shift would indicate NVIDIA’s intent to lean more heavily on Korean suppliers, the reports suggest.

According to ZDNet, SK hynix is gearing up to mass-produce SOCAMM2 based on 1C (6th-generation 10nm-class) DRAM starting in Q2 2026. With SOCAMM2 driving stronger LPDDR demand, a substantial volume of advanced LPDDR is expected to go to NVIDIA, potentially tightening supply for China’s smartphone makers, the report notes.

Micron, on the other hand, notes that it led the charge in bringing low-power DRAM to data centers, with its LP DRAM server modules consuming only a third of the power of traditional DDR DRAM modules. Building on this edge, the company has sampled its 192GB LP SOCAMM2, offering a 50% boost in capacity per module and enabling rack-scale LP DRAM densities exceeding 50TB.

Read more

• [News] NVIDIA’s SOCAMM2 Orders Likely Evenly Split Among Memory Makers in 2026, 1c Adoption in Focus
• [News] Samsung in Talks to Supply HBM4 to NVIDIA for 50K-GPU AI Mega Fab; HBM3e/GDDR7/SOCAMM2 Also Included

(Photo credit: NVIDIA)