Micron has begun sampling a 256GB DDR5 registered dual in-line memory module (RDIMM) to key server ecosystem partners. The module is built on the company's 1-gamma DRAM technology and uses 3D stacking (3DS) with through-silicon vias (TSVs) to connect multiple memory dies. Micron claims the module can reach speeds up to 9,200 megatransfers per second (MT/s), which is more than 40% faster than modules currently in volume production (compared to 6,400 MT/s).
What it does
The 256GB DDR5 RDIMM is designed for servers running large language models (LLMs), agentic AI, real-time inference, and high-core-count CPU workloads. By packing 256GB into a single module, it reduces operating power by more than 40% versus using two 128GB modules — Micron calculates 11.1W for the single 256GB module versus 19.4W total for two 128GB modules running at 9.7W each.
Ecosystem validation
Micron is collaborating with key server platform enablers to validate the module across current and next-generation platforms. This co-validation aims to ensure broad compatibility and accelerate production deployment for data center customers building AI and HPC infrastructure at scale.
Tradeoffs
While the module offers higher capacity and bandwidth per slot, it requires platform support for 3DS and TSV packaging. Not all existing server platforms may support the module's physical and electrical requirements. Additionally, the module is currently only sampling to ecosystem partners — general availability and pricing have not been announced.
When to use it
This module is aimed at hyperscale operators and enterprise data centers that need to maximize memory capacity per CPU socket while staying within thermal and power limits. It is particularly relevant for AI inference servers and high-performance computing clusters where memory bandwidth is a bottleneck.
Bottom line
Micron's 256GB DDR5 RDIMM on 1-gamma DRAM represents a meaningful step forward in server memory density and speed. The 40% power savings versus two 128GB modules is a practical advantage for data centers scaling AI workloads. However, real-world performance and platform compatibility will depend on ecosystem validation results, which are still in progress.