Innodisk’s new DDR5 12800 MT/s MRDIMM aims to crush the AI “memory wall”

The bottleneck for training Large Language Models (LLMs) isn’t always the GPU; it’s often how fast the system can feed data to it. Innodisk tackled this “memory wall” head-on on July 30, introducing a staggering DDR5 12800 MT/s MRDIMM (Multiplexed Rank Dual Inline Memory Module).

Breaking bandwidth limits

Traditional high-end servers currently rely on DDR5 8000 MT/s RDIMMs. Innodisk’s new modules deliver a 60% increase in pure bandwidth. They achieve this by combining two DDR5 data words into a single, wider data word delivered to the host CPU, effectively doubling the data rate per clock cycle.

These modules are specifically validated for upcoming AI server platforms (such as those based on Intel’s Granite Rapids and AMD’s Turin architectures). As AI models scale into the trillions of parameters, memory speed has become the defining constraint on server performance, making MRDIMM technology essential for the 2026 enterprise upgrade cycle.