NVIDIA and Partners Develop SOCAMM — The Future of Memory for AI Computers

NVIDIA has teamed up with leading memory manufacturers — SK Hynix, Samsung, and Micron — to create a new memory standard called the System On Chip Advanced Memory Module (SOCAMM). This module is poised to revolutionize AI devices by combining compact size, high performance, and energy efficiency.

SOCAMM is being developed as the successor to existing memory standards such as Low-Power Compression Attached Memory Modules (LPCAMM) and traditional DRAM modules. One of its key advantages is cost efficiency. Unlike SO-DIMM DRAM, SOCAMM integrates LPDDR5X memory directly onto the substrate, reducing manufacturing costs and improving energy efficiency.

Another distinguishing feature of SOCAMM is its increased number of input/output interfaces. The new module will support up to 694 ports — significantly more than LPCAMM (644 ports) and traditional DRAM solutions (260 ports). This will enable high bandwidth and faster data transfer rates, crucial for running resource-intensive AI models.

SOCAMM also stands out for its compact size. According to sources, the module will be roughly the size of an adult’s finger, making it considerably smaller than traditional DRAM modules. This not only increases overall memory capacity in devices but also opens up new possibilities for designing slimmer and lighter AI hardware. Additionally, SOCAMM will feature a removable mechanism, simplifying future hardware upgrades. Users will be able to upgrade their device’s memory without undergoing complex component replacements.

SOCAMM is expected to form the backbone of NVIDIA’s next-generation AI computers, including the successor to Project Digits, showcased at CES 2025. These systems are designed for demanding AI workloads that require substantial amounts of RAM. According to recent reports, NVIDIA and its partners are already exchanging SOCAMM prototypes for performance testing. Mass production of the new module could begin as early as the end of this year.