Overview
- Samsung says SOCAMM2 more than doubles bandwidth versus DDR5 RDIMMs while cutting power use by over 55 percent.
- The module packages LPDDR5X in a detachable compression-attached form factor to maintain serviceability and shrink board footprint.
- Samsung is working with Nvidia to validate the design for accelerated infrastructure, with broader uptake reported as likely alongside the Rubin platform targeted for 2026.
- The approach targets inference-heavy AI deployments as CPU-attached memory that complements, rather than replaces, HBM on accelerators.
- SOCAMM2 is being aligned with JEDEC’s CAMM2 (JESD328) standard as SK hynix and Micron pursue competing offerings, with thermal density, latency, and enterprise ECC still undergoing real-world validation.