Against the background of the rapid expansion of agent AI (Agentic AI) workloads, the demand for high bandwidth, high energy efficiency, and large-capacity memory in data centers continues to soar. LPDDR memory is gradually becoming a key component of the new generation of AI servers. The latest trends in the industry show that the next generation LPDDR6 standard will increase the single module capacity to 512GB and support the SOCAMM2 packaging form, providing higher memory density and better energy efficiency for future AI training and inference.
According to the preliminary specifications of LPDDR6 published by JEDEC, LPDDR6 is positioned as the "main" memory standard for data centers and will continue to increase capacity and bandwidth while maintaining power consumption advantages. LPDDR6 adopts a narrower x6 single-chip interface design, supplemented by a non-binary interface and additional x6 sub-channels, allowing a single DRAM chip to be stacked to achieve higher capacity. The overall bandwidth is expected to increase by about 10%–20% compared to the existing LPDDR5 / LPDDR5X, but the design focus is more on density expansion rather than extreme bandwidth.
JEDEC stated that its goal is to increase the capacity of the LPDDR6 SOCAMM2 module to 512GB based on the existing LPDDR5X SOCAMM2 module of up to 256GB to meet the expanding memory requirements for AI training and inference. Today, as the computing power of data centers continues to increase, memory capacity has become one of the key bottlenecks affecting the training scale and inference throughput of large models. Higher-density LPDDR6 modules can stack up a larger effective memory pool within limited rack space.
From the perspective of the pace of industrial implementation, mainstream storage manufacturers such as Samsung, SK Hynix and Micron have previously set the commercial time window for LPDDR6 in 2028-2029. However, driven by the strong demand brought by agent-based AI, the industry generally expects that related products and platforms may be launched in advance. The LPDDR6 and SOCAMM2 technology paths disclosed by JEDEC are also regarded as important signals paving the way for a new generation of AI server platforms.
In terms of server platform ecology, NVIDIA has adopted the LPDDR5X SOCAMM2 solution on its Vera series CPUs, while AMD has introduced LPDDR5X in the Verano series CPUs optimized for inference scenarios to improve rack-level performance per watt. With the advancement of LPDDR6, the industry expects that manufacturers such as NVIDIA and AMD will upgrade to higher-density and more energy-efficient LPDDR6 SOCAMM2 modules to provide greater memory bandwidth and capacity redundancy for future AI factories and inference clusters.
It is worth noting that the LPDDR6-PIM (Processing/Programming In Memory) technology promoted by JEDEC at the same time will further expand the application boundaries of LPDDR6 in the data center field. By integrating the PIM controller on the memory side, the LPDDR6 module can perform some computing tasks locally and offload some data processing work from the CPU to the memory, which is expected to reduce data handling overhead and improve overall energy efficiency. In scenarios where AI training and inference highly rely on large-scale matrix operations and tensor operations, this type of "integrated storage and calculation" technology is regarded as an important development direction to deal with power consumption and bandwidth walls.
From the perspective of energy efficiency characteristics, the LPDDR standard has always been known for its low voltage and low power consumption. Compared with traditional server DDR memory, it can significantly reduce the overall energy consumption of the data center when deployed on a large scale. JEDEC pointed out that as AI and data center power consumption indicators are increasingly valued, LPDDR6 and its supporting SOCAMM package are rapidly attracting the attention of cloud service providers and hyperscale data center customers.
Overall, LPDDR6 is regarded as one of the core memory technologies for next-generation AI data centers and will see a peak in demand driven by agent-based AI. With a maximum module capacity of 512GB, better power consumption performance, and innovative features such as PIM, LPDDR6 is expected to become the mainstream memory standard for AI server platforms in the next few years, providing more efficient infrastructure support for large model training and inference.