Micron Unveils First 256GB SOCAMM2 LPDRAM Module

On March 3, 2026, Micron Technology announced that it has begun shipping customer samples of the industry's largest-capacity LPDRAM module — a 256GB SOCAMM2. The new module further strengthens Micron's position in low-power server memory and introduces a major step forward for AI data center architectures.

The milestone is enabled by the industry's first single-die 32Gb LPDDR5X design, allowing significantly higher memory capacity while maintaining low power consumption. With this innovation, the 256GB SOCAMM2 module helps unlock new system designs for AI inference, large language models, and other data-intensive computing workloads.

Higher Memory Capacity for AI Servers

The 256GB SOCAMM2 expands server memory capacity by one-third compared with the previous 192GB maximum module. In an 8-channel CPU platform, the configuration can deliver up to 2TB of LPDRAM per processor, enabling larger context windows and more complex inference workloads for AI applications.

Lower Power Consumption and Smaller Footprint

Compared with traditional RDIMM modules, SOCAMM2 uses roughly one-third of the power and occupies only one-third of the physical space. This efficiency allows higher rack density while helping reduce overall data center operating costs.

Faster AI Inference and Compute Performance

In unified memory architectures, the 256GB SOCAMM2 significantly improves inference responsiveness. When used for KV cache offloading, it can improve time-to-first-token (TTFT) by more than 2.3×.

According to Micron's internal testing using the Llama3 70B model with FP16 quantization, real-time inference was evaluated with a 500,000-token context length and 16 concurrent users. Systems configured with 2TB LPDRAM per CPU achieved a TTFT latency of about 0.12 seconds, compared with 0.28 seconds for systems with 1.5TB LPDRAM. Even with 1 million-token contexts, systems equipped with 2TB of LPDRAM maintained faster response times.

For standalone CPU-based computing workloads, LPDRAM also demonstrated more than three times better performance per watt compared with mainstream server memory modules.

Modular Design for Scalability and Serviceability

The SOCAMM2 modular architecture improves serviceability and supports modern liquid-cooled server designs. It also allows future capacity expansion as AI and high-performance computing workloads continue to demand larger memory pools.

Micron has started shipping 256GB SOCAMM2 customer samples and now offers one of the industry's broadest LPDRAM portfolios for data centers, ranging from 8GB to 64GB components and 48GB to 256GB SOCAMM2 modules.