Huawei AI Data Infrastructure Suite for Modern Data Centers Including High Density Data Lakes and Model Engineering Automation
Vice President Yuan Yuan announced during the Huawei Innovative Data Infrastructure Forum a full stack data infrastructure suite, enabling organizations to facilitate the move to an AI data center. In his keynote, titled
"Data Awakening Infra Evolving"he discussed the physical and digital architecture that would be necessary for the sheer quantity of token consumption by enterprise AI systems. Yuan stated that modern IT infrastructure needs to adapt in six areas: Data lakes, AI data platforms, computing power, model engine, agent frameworks and data resilience platforms.
The new architecture begins with the requirement of high density storage to facilitate tremendous quantities of training data. Huawei released the OceanStor Pacific Scale Out Storage system which offers 11 PB of data in 2U of server space. The system is designed to maximize the total cost of ownership by providing increased data capacity and reduced space and cooling demands. To assist with global data access, the DME Omni Dataverse unified space solution was introduced. This software platform facilitates multi modal and real time data ingestion across multiple physical locations and enables researchers to search data of billions of 1000 dimensional vectors within seconds.
To enable inference on a large scale, Huawei created the Context Memory Storage system, the first available heterogeneous computing system. This memory solution boasts a 90% improvement in the time to the first token by leveraging key value semantic direct pass technology, or offloading semantic tasks to a specialized data processing unit. The key value pool can scale up to PBs and be accessed at higher speeds across large clusters of servers.
For typical enterprise inference environments, a 3+1 AI data platform was revealed. The system includes a key value cache, a knowledge base with 95% retrieval accuracy and a dynamic memory bank, which are synchronized by the Unified Cache Manager software tool. This combined platform enables 30% greater AI inference accuracy due to better scheduling of cached data.
The ModelEngine platform, a zero code adaptation and single click deployment for machine learning model use was announced. The system utilizes an intelligent resource partitioning feature for an up to 1:10 split between compute resources. This means that one graphics card will support several isolated operations which will dramatically reduce downtime of expensive hardware.
ModelEngine Nexent is an agent platform to allow for the deployment of digital employees. Developers will be able to prompt for the creation of digital agents through natural language input, reducing the deployment timeline by 80%. The Nexent platform will automatically improve prompting and memory paths for greater accuracy over time.
The presentation concluded with a discussion of how to secure digital assets from increasingly prevalent security threats. Huawei showed off an end to end security platform that is designed to protect AI data centers against tool poisoning, data poisoning, unauthorized modification and ransomware by monitoring the entire data cycle from ingest to model training.
