Micron and SK hynix introduce compact SOCAMM memory modules for Nvidia GB300 AI systems

Micron, Samsung and SK hynix have announced new SOCAMM memory modules based on stacked LPDDR5X chips, with capacities reaching up to 128 GB. The technology is aimed at low-power servers and systems using artificial intelligence. The first SOCAMM will appear in systems with the Nvidia Grace Blackwell Ultra GB300 Superchip.

Image source: Micron

Memory power consumption is one of the key factors affecting the overall energy efficiency of servers. In systems with terabytes of DDR5 memory per socket, DRAM power consumption can exceed CPU power consumption. Nvidia took this into account when designing its Grace processors, which were designed to use energy-efficient LPDDR5X memory. However, in the GB200 Grace Blackwell-based machines, the company had to use soldered-on LPDDR5X modules, as standard modules could not provide the required capacity.

Micron’s SOCAMM addresses this issue by offering a standard module solution that can accommodate four 16-layer LPDDR5X stacks, delivering high capacity. The company says its 128GB SOCAMM modules consume one-third less power than DDR5 R-DIMM modules of the same capacity. It’s unclear whether SOCAMM will become an industry standard supported by JEDEC or remain a proprietary solution developed by Micron, Samsung, SK hynix, and Nvidia for Grace and Vera-based servers.

Micron has already launched mass production of SOCAMM. They will appear in systems based on Nvidia GB300 Grace Blackwell Ultra Superchip in the second half of this year. It should be noted that the modular design simplifies the production and maintenance of servers, which will probably have a positive effect on the prices of systems, primarily for the needs of artificial intelligence.