Memory for Rubin

Artificial intelligence servers spend much of their time moving data, not just calculating it, and SK hynix said on April 20 it began mass-producing a new memory module for Nvidia’s Vera Rubin systems. (news.skhynix.com) The product is a 192-gigabyte SOCAMM2 module, a server memory stick built from LPDDR5X, the low-power DRAM more commonly used in smartphones and laptops. SK hynix said it is using its sixth-generation 10-nanometer-class process, which it calls 1c nanometer. (news.skhynix.com) SK hynix said the module delivers about twice the bandwidth of a standard registered dual in-line memory module, or RDIMM, while using roughly 75% less power than the DDR5 RDIMM products now common in servers. The company also lists SOCAMM2 capacities from 32GB to 256GB and speeds up to 9,600 megabits per second on its product site. (news.skhynix.com; product.skhynix.com) That matters because Nvidia’s Vera Rubin platform is aimed at “agentic AI” and long-context reasoning, workloads that keep shuttling large amounts of data between processors and memory. Nvidia says Rubin is designed to remove bottlenecks in communication and memory movement and to produce more tokens per watt than the Blackwell generation. (nvidia.com) In plain terms, bandwidth is the width of the pipe carrying data, and power draw is the electric bill for keeping that pipe full. Faster, lower-power memory lets an AI rack feed chips with fewer delays and less heat, two constraints that shape how many systems a data center can deploy in one building. (nvidia.com; news.skhynix.com) SOCAMM2 also changes the shape of the hardware. SK hynix describes it as a small-form-factor server module, which means more memory can fit into tighter spaces than with bulkier RDIMMs, a useful trait in dense AI racks where cooling and board area are limited. (product.skhynix.com) Nvidia has been building Rubin as a rack-scale platform rather than a single chip launch. The company said on March 16 that seven new Rubin chips were in full production, including Vera Rubin NVL72 GPU racks and Vera CPU racks for large AI factory deployments. (nvidianews.nvidia.com) Memory suppliers have been racing to lock in that design cycle. The Korea Economic Daily reported in December 2025 that Samsung and SK hynix were both developing SOCAMM2 for Rubin-class systems, with Samsung moving quickly in customer sampling and SK hynix pushing high-capacity server DRAM in parallel. (en.sedaily.com) SK hynix had already been showing SOCAMM-family products at Nvidia events and CES before this production announcement. The shift from demos and samples to mass production means the memory piece of the Rubin stack is moving closer to volume server builds. (news.skhynix.com; news.skhynix.com; news.skhynix.com) For data-center operators waiting on Rubin, the bottleneck is no longer just the accelerator chip itself. The memory modules that keep those chips fed are now entering production too. (news.skhynix.com; nvidia.com)