Solution Architecture

In-house + open-source algorithm platform:The algorithm platform mainly includes physical-network transport, cluster management and scheduling, and an AI platform. For deep-learning model training, besides GPUs that provide powerful computing, PCIe transport bandwidth must be guaranteed; for multi-machine cases, network devices with better network bandwidth are needed to ensure overall data-transfer efficiency and reduce the impact of network data transfer; cluster management and scheduling must monitor and analyze overall cluster status and real-time node status and form real-time visual data reports; the AI platform must provide rapid deployment of deep-learning development environments and, for deep-learning development, split and distribute computing resources by training job.

Compute + AI chips:For different deep-learning scenarios, different types of GPU servers that can carry multiple GPUs become core compute units in the compute layer. Meanwhile, non-core compute parts such as cluster management and desktop services use general rack servers.

In-house storage system:The storage system is mainly used to store compute data. In HPC, tens or hundreds of compute nodes need unified-image shared storage; a parallel file system unifies all storage arrays into one large storage, and the parallel file system can meet this user need.

The deep-learning cluster solution applies to communications-industry customers’ business needs and exploration in smart operations, smart connectivity, smart services, intelligent marketing, intelligent decision-making, smart networks, intelligent IoT, intelligent customer service, interactive entertainment, and more.