Heterogeneous Computing Management & Domestic AI Computing Platform
Unified Scheduling of CPU, GPU, NPU Heterogeneous Computing Resources, Building Kilocard to Myriacard Domestic AI Computing Clusters
Functional Overview
Unified scheduling of CPU, GPU, NPU, and other heterogeneous computing resources, building kilocard to myriacard domestic AI computing clusters, supporting large-model training/inference and business-scale applications.
Customer Pain Points
Computing resources (CPU/GPU/NPU) are scattered, utilization is low, and unified scheduling is impossible.
Domestic computing power (Huawei Ascend, Pingtouge, etc.) is difficult to adapt, and model inference performance is poor.
Large-model training costs are high, and there is a lack of planning for scaling from kilocards to myriacards.
Data center energy consumption is high, PUE optimization is insufficient, and non-real-time computing is not coordinated with real-time computing.
Solution Advantages
Full-stack localized kilocard cluster: built based on Huawei Ascend, with end-to-end joint optimization from hardware procurement to model tuning.
Heterogeneous computing unified scheduling: the cloud platform uniformly manages computing power across three cities and five centers in Shanghai, Hefei, and Horinger, achieving "real-time + non-real-time" collaboration.
Model lightweighting capability: supports collaboration among large, small, and multimodal models, with model distillation reducing deployment costs.
Myriacard evolution planning: expand to myriacard scale to support complex AI scenarios.
Green data center: implement "East Data West Computing," PUE optimization, AIOps intelligent fault prediction, and energy management.
