Danny Weber
The new LPDDR6 standard focuses on higher density and efficiency, with future 512 GB modules and PIM features aimed at AI infrastructure.
The LPDDR6 memory standard is increasingly being described as one of the important building blocks for future data centers built around artificial intelligence. The reason is simple: agentic AI needs more and more system memory, while power consumption is becoming just as important as raw performance.
According to JEDEC, LPDDR6 will bring several architectural changes, with the focus not only on speed but also on memory density. One of the most notable steps could be modules with capacities of up to 512 GB. That is twice as much as current SOCAMM2 modules based on LPDDR5X, which top out at 256 GB.
The new standard is expected to use a narrower x6 interface per die and additional subchannels. This design should help memory makers increase the capacity of individual chips without a sharp rise in power consumption. At the same time, the bandwidth gain is expected to be relatively modest by early estimates — roughly 10–20% over the current generation.
LPDDR6 looks especially interesting for server platforms built for artificial intelligence. NVIDIA already uses LPDDR5X SOCAMM2 modules in Vera processors, while AMD has announced LPDDR5X support in Verano processors aimed at inference workloads. Moving to LPDDR6 could noticeably expand what these systems can do, especially where large memory capacity and energy efficiency matter most.
A separate track is LPDDR6-PIM technology. It is designed to run some calculations directly inside memory, without constantly moving data back and forth to the central processor. For AI workloads, that could be a serious advantage: fewer unnecessary data transfers, lower system load and higher overall efficiency.
Low power consumption remains one of the main strengths of LPDDR. For large data centers, this is especially important: AI infrastructure is growing quickly, and with it come rising demands for power, cooling and operating costs.
Earlier reports suggested that Samsung, SK Hynix and Micron could bring LPDDR6 to the commercial market between 2028 and 2029. But given the pace of AI infrastructure growth, adoption in some segments could begin sooner.
© RusPhotoBank