Solution Architecture

Solution technical architecture

Solution physical architecture

Solution composition

SothisAI software platform:Sugon SothisAI is a containerized enterprise distributed deep-learning platform, providing efficient, fast AI solutions and one-stop deep-learning solutions. It helps users with multi-group resource allocation, rapid development-environment setup, and flexible application migration. SothisAI supports mainstream deep-learning frameworks, provides graphical, SSH, Jupyter, and other access methods, and is supported by Slurm and Kubernetes dual scheduling engines, meeting characteristics of different application scenarios.

GPU-based heterogeneous computing cluster:For deep-learning application characteristics, use 4U 8-GPU high-density in-house servers X780 and X795 with mainstream AI heterogeneous accelerator cards to provide the cluster with strong computing support. Meanwhile, the cluster’s high-bandwidth, low-latency InfiniBand network can meet high PCIe transport-bandwidth requirements of multi-machine, multi-GPU network-model training and ensure overall system data-transfer efficiency, reducing the impact of network data transfer.

ParaStor storage system:ParaStor is Sugon’s independently developed distributed parallel storage system; the latest version is ParaStor300, using multi-replica, N+M erasure coding, and other data-protection technologies and fully redundant design, supporting a single storage namespace, massive capacity expansion, and linear performance scaling, fully meeting deep-learning application needs such as frequent dataset read/write, concurrent multi-user access, and frequent data interaction during training.