LPDDR6 memory capacity will double to 512 GB per module to support agentic AI datacenter workloads.

Agentic AI workloads are driving aggressive increases in memory capacity demand across the industry, positioning LPDDR6 as the next major component to address these requirements. LPDDR memory has become the gold standard for datacenters due to its higher memory capacities and superior efficiency characteristics, with LPDDR6 representing the next evolution in this technology.

Samsung, SK Hynix, and Micron are aiming for the commercialization of LPDDR6 by 2028-2029, though Agentic AI demand may accelerate this timeline. JEDEC has previewed the initial specification for LPDDR6, which will support 512 GB capacities in SOCAMM2 modules, clearing the path for next-generation AI servers. While the bandwidth improvement will be modest at 10-20% over the current generation, the higher densities are what matter most for these applications. According to JEDEC, datacenter customers are interested in LPDDR, and LPDDR6 will bring several new features addressing their needs.

The key technical advancement enabling higher capacities is a narrower x6 per-die interface combined with a non-binary interface and additional x6 sub-channel, allowing DRAM makers to achieve 512 GB density per integrated circuit. This represents a doubling of the current SOCAMM2 LPDDR5X maximum of 256 GB. As Nagashima stated, the goal is to reach 512GB density to address the ever-growing memory capacity requirements of AI training and inference workloads. LPDDR6 will continue using SOCAMM2 modules, the preferred choice for AI datacenters, following the path already taken by NVIDIA with its Vera CPUs and AMD with its inference-optimized Verano CPUs.

JEDEC is nearing completion of LPDDR6-PIM (Programming In Memory) technology, which further strengthens the LPDDR6 roadmap. The PIM controller will enable memory to perform calculations directly rather than simply storing data, offloading processing from the CPU to the memory itself. As JEDEC noted, "AI and data centers are really driving—or straining—the ecosystem right now." LPDDR standards are known for their superior power characteristics, which will contribute to more efficient datacenters as power consumption becomes an increasingly critical metric in AI infrastructure.