Particle.news
Download on the App Store

SK hynix Begins Mass Production of 192GB SOCAMM2 for Nvidia’s Vera Rubin AI Servers

The LPDDR5X design aims to ease AI memory bottlenecks with higher bandwidth at lower power.

Overview

  • SK hynix said it has started mass production of its 192GB SOCAMM2 memory module designed for Nvidia’s Vera Rubin platform.
  • The module repurposes LPDDR5X mobile memory for servers in a slim, replaceable format that the company says can act as primary system memory.
  • SK hynix claims more than double the bandwidth and over 75 percent better power efficiency than traditional RDIMM server modules.
  • The company says the gains target slowdowns that occur when training and running very large language models with hundreds of billions of parameters.
  • SK hynix reports stabilized production and supply to cloud providers, lists SOCAMM2 at 9.6 Gbps per pin versus 8.5 Gbps in SOCAMM1, and notes independent testing has not yet verified the claims.