Micron announced that its HBM4 36GB 12H and 192GB SOCAMM2 memory modules, designed specifically for the NVIDIA Vera Rubin platform, as well as the world's first PCIe Gen6 data center SSD 9650, will all enter mass production in the first quarter of 2026. This not only makes Micron the first memory supplier to simultaneously provide three key components for the Rubin ecosystem, but also heralds a new era for AI infrastructure, moving from "single-point performance" to "platform-level collaboration."
HBM4 36GB 12H Mass Production: Bandwidth Exceeds 2.8 TB/s, Energy Efficiency Further Improved
As AI models continue to advance towards mega-scale parameters, GPUs' demands for memory bandwidth and capacity are endless. Micron's newly mass-produced HBM4 36GB 12H stack boasts data transfer rates exceeding 11 Gb/s and a total bandwidth surpassing 2.8 TB/s. Compared to the same capacity HBM3E, it offers up to 2.3 times the bandwidth and over 20% improvement in energy efficiency.
為了進一步滿足未來超大規模AI工廠的需求，美光也已開始向客戶送樣HBM4 48GB 16H，成功實現16層HBM晶片堆疊的先進封裝技術，使每顆HBM的配置容量較現有36GB 12H產品提升33%。
SOCAMM2 is designed specifically for Vera CPUs: a single CPU supports 2TB of memory.
Besides the HBM required by GPUs, the CPU also plays an increasingly important role in AI systems. With the explosive growth in demand for agent-based AI and real-time inference, CPUs need larger memory capacities to handle key-value caching, task scheduling, and tool invocation. Micron's SOCAMM2 (Small Outline Compression Attached Memory Module 2) was developed precisely for this purpose.
This low-power memory module, designed specifically for NVIDIA Vera Rubin NVL72 systems and discrete NVIDIA Vera CPU platforms, allows each CPU to support up to 2TB of memory and 1.2 TB/s bandwidth. Compared to traditional RDIMMs, SOCAMM2 consumes only one-third of the power and is also one-third the size, significantly improving rack density and reducing overall cost of ownership. In real-time inference scenarios involving long-context large language lines (LLM), it can accelerate the first token generation time by 2.3x when used for key-value cache offloading.
Micron's SOCAMM2 portfolio covers capacities from 48GB to 256GB, providing flexible options for AI deployments of different sizes.
The world's first mass-produced PCIe Gen6 SSD 9650: read speeds up to 28GB/s, optimized for the BlueField-4 STX storage architecture.
In the storage sector, Micron has also turned a new page in the industry. The Micron 9650 data center SSD has officially entered mass production, becoming the world's first PCIe Gen6 data center solid-state drive. Its sequential read throughput reaches up to 28 GB/s, and its random read speed reaches 550 million IOPS. Compared to the previously released PCIe Gen5 products, its read performance is doubled, and its performance per watt is improved by 100%.
The Micron 9650 SSD is optimized for energy efficiency and liquid cooling, providing high-speed, low-latency data access for AI training and inference phases under the NVIDIA BlueField-4 STX storage architecture. As AI workloads shift from simple model training to multi-stage inference and proxy tasks, storage device I/O bottlenecks have become a critical issue. The Micron 9650 SSD aims to bridge the gap between GPU computing power and data delivery.
Micron is simultaneously offering the 7600 and 9550 series PCIe Gen5 SSDs, providing customers with more diverse architectural design options when building AI infrastructure.
