Micron Unveils SOCAMM Memory Modules for NVIDIA GB300 Series at GTC 2025
March 19, 2025 – At the GTC 2025 conference, Micron Technology officially introduced its SOCAMM (Stacked-On-Carrier Attached Memory Module) solution, which will be deployed in NVIDIA’s newly launched Blackwell Ultra GB300 product line.
March 19, 2025 – At the GTC 2025 conference, Micron Technology officially introduced its SOCAMM (Stacked-On-Carrier Attached Memory Module) solution, which will be deployed in NVIDIA’s newly launched Blackwell Ultra GB300 product line.
Micron confirmed that SOCAMM has now entered mass production, offering the world’s fastest, most compact, lowest-power, and highest-capacity memory module. Designed to support the growing demand for AI workloads, this advanced memory solution significantly enhances memory bandwidth, power efficiency, and scalability, making it an ideal choice for AI servers and data-intensive applications.
SOCAMM vs. RDIMM: A Leap in Performance and Efficiency
Micron’s SOCAMM solution outperforms traditional RDIMM in multiple key aspects, offering:
2.5x the bandwidth of RDIMM, enabling faster access to massive AI training datasets and more complex models while improving inference throughput.
Compact size (14 x 90 mm)—only one-third the size of standard RDIMM modules—allowing for space-efficient and high-density server designs.
Superior power efficiency, leveraging LPDDR5X memory, reducing power consumption to just one-third of standard DDR5 RDIMMs—an essential factor for AI-driven data center architectures.
Next-Generation Memory Architecture for AI Acceleration
Micron’s SOCAMM memory modules integrate:
Four 16-die stacked 16Gb LPDDR5X chips, achieving an impressive 128GB capacity.
128-bit memory bus at 8533 MT/s, facilitating faster AI model training and supporting higher concurrent inference workloads.
Enhanced scalability and maintenance features, including advanced error correction capabilities, optimized for data center reliability.
Optimized for AI Server Deployment & Liquid Cooling
Unlike LPCAMM2, Micron’s SOCAMM modules:
Feature a single-sided four-chip design, secured with three mounting screw holes for better stability.
Eliminate the protruding trapezoidal structure on top, reducing overall height to enhance compatibility with liquid-cooled server architectures.
By integrating SOCAMM into NVIDIA’s Blackwell Ultra GB300 lineup, Micron is setting a new standard in AI memory technology, delivering a high-bandwidth, power-efficient solution tailored for the next generation of AI servers and cloud infrastructures.
