Data Awakening

The groundbreaking control and data plane separation architecture facilitates the direct flow of data to disks, reducing CPU usage and delivering 24 million IOPS and 500 GB/s bandwidth per controller enclosure. Training-set loading efficiency is 8x higher than the next best in the industry. After a breakpoint, training can resume 4x faster than the next-best product in the industry. Near-storage computing accelerates data preprocessing. In addition, storage clusters can scale up to a maximum of 4,096 computing cards.

Fully-symmetric architecture supports scale-out to up to 512 controllers, delivering EB-level, on-demand capacity expansion. This enables large AI models to perform smooth evolution to ultra-large clusters with 10,000, or even 100,000 GPUs, multi-modality, and trillions of parameters.

An industry-leading intrinsic vector knowledge repository eliminates AI hallucinations and boosts vector retrieval speeds to more than 250,000 QPS, guaranteeing millisecond-level inference response.

The global file system uses standard interfaces to unify metadata management across regions, sites, and vendors, providing a truly global data view. This helps to implement fast file retrieval based on service requirements: retrieval from 50 billion files is finished in just 5s.

Omni-Dataverse supports automatic data tiering and classification, formulating policies based on the results of data analysis to achieve automatic mobility for hot, warm, and cold data between devices. Cold data, for example, can be moved from high-performance storage to low-cost storage, reducing the Total Cost of Ownership (TCO) by 20%.

Omni-Dataverse supports data access using different protocols such as S3, Network File System (NFS), Server Message Block (SMB), Distributed Parallel Client (DPC), and Hadoop Distributed File System (HDFS). In addition, it automatically loads data across sites to the cache at the site where the data access request was actually made, based on service requirements. Combined with DPC, this boosts data read efficiency by a factor of two.