Overview of the AI Data Center Storage Chip Adaptation Solution

Focusing on the core demands of high bandwidth, large capacity and low latency in AI data centers, we carefully select flagship storage products from leading brands for cooperation. The computing end is equipped with SK Hynix HBM3E (with a bandwidth of 1.23TB/s) and the mass-produced HBM4 (with a bandwidth of 2TB/s), and is paired with Samsung and Micron DDR5 (up to 6400MT/s, supporting ECC error correction), providing support for large model training and massive data processing. The storage end adopts Yangtze Memory Technologies Co., LTD. 's Xtacking 4.0 (with over 300 layers stacked), Kioxia's 10th generation BiCS FLASH (with a 59% increase in storage density), and Micron's 176-layer 3D NAND (with sequential read and write speeds exceeding 10GB/s), ensuring high-speed storage throughput. The entire solution is compatible with PCIe 5.0/UFS4.1 interfaces, balancing heat dissipation efficiency and stability. It meets the 7× 24-hour high-load operation requirements of AI clusters, supercomputers, and data processing nodes, providing a high-performance and highly reliable full-stack storage solution.