SK Hynix rolls out SOCAMM2 memory

SK hynix has started mass production of a 192-gigabyte server memory module called SOCAMM2, built for Nvidia’s next Vera Rubin artificial intelligence systems. (news.skhynix.com) Memory is the working space an artificial intelligence server uses while it answers prompts, ranks tokens, and moves data between chips. SK hynix said SOCAMM2 uses LPDDR5X low-power DRAM in a server module, instead of relying only on the larger registered DIMMs, or RDIMMs, common in today’s servers. (news.skhynix.com) The company said the new module delivers more than double the bandwidth of conventional server RDIMM and improves power efficiency by more than 75%. SK hynix said the product is based on its sixth-generation 10-nanometer-class, or 1c, DRAM process. (news.skhynix.com) SK hynix said SOCAMM2 is designed for Nvidia’s Vera Rubin platform, which Nvidia introduced in March 2025 as the successor to Blackwell for large-scale artificial intelligence computing. Nvidia said Vera Rubin systems will pair new graphics processors with a custom central processor called Vera. (news.skhynix.com) (nvidia.com) The pitch is not that SOCAMM2 replaces high bandwidth memory, or HBM, the ultra-fast memory stacked next to an artificial intelligence chip. The pitch is that it fills the gap between HBM, which is fast and expensive, and DDR5 server memory, which is cheaper but slower for some inference workloads. (trendforce.com) (news.skhynix.com) That matters as data center spending shifts from training giant models to serving them. Nvidia said inference token generation demands high memory bandwidth and capacity, and hyperscale operators have been reworking server designs around power and throughput limits. (nvidia.com 1) (nvidia.com 2) SK hynix said the module uses a compression connector to improve signal integrity and make replacement easier. The company also said the design is thinner than standard server memory, which lets system builders fit more modules into a server. (news.skhynix.com) The rollout also widens the memory fight around Nvidia’s next platform. TrendForce said Samsung and Micron are also pursuing SOCAMM products, turning a niche server module into a new contest beyond the high bandwidth memory market that SK hynix already leads. (trendforce.com) (reuters.com) For now, the clearest fact is simpler: SK hynix is moving low-power mobile-style memory into artificial intelligence servers at production scale, and tying that bet to Nvidia’s Rubin generation. (news.skhynix.com)