Particle.news
Download on the App Store

SK hynix Starts Mass Production of 192GB SOCAMM2 for Nvidia’s Vera Rubin

The LPDDR5X-based module targets AI server memory limits with higher bandwidth and lower power for Nvidia’s next platform.

Overview

  • SK hynix, which began mass production Monday, is building a 192GB SOCAMM2 module tailored for Nvidia’s upcoming Vera Rubin AI platform.
  • The module uses sixth‑generation 10nm‑class LPDDR5X built on the 1cnm process and SK hynix says it delivers more than double the bandwidth with over 75% better power efficiency than DDR5 RDIMMs.
  • Design changes include vertically stacked LPDDR chips, a higher pin count for input and output, and data transfer speeds of 9.6 Gbps, up from 8.5 Gbps in the prior SOCAMM1.
  • SK hynix says SOCAMM2 is meant to relieve memory bottlenecks in training and running very large language models by adding a replaceable LPDDR tier between on‑package HBM and system DDR5 to cut power and cooling needs.
  • Investors welcomed the move as SK hynix shares rose about 3.4% in Seoul trading, and analysts say early volume production signals supply‑chain progress ahead of Vera Rubin shipments expected in the second half of 2026.