Aiming at the problem of data center resource island, that is, the architecture difference between different computing platforms is large, GPU resources cannot be shared and reused, and the elastic expansion is limited to the platform, a GPU resource scheduling system for heterogeneous computing platforms is proposed. The system takes GPU resources as general resources and schedules them among heterogeneous platforms to realize resource sharing and solve the problem of elastic expansion of node-level GPU resources in different application platforms. The specific architecture of the system will be elaborated in detail, and the data acquisition, resource scheduling and initialization of computing nodes of heterogeneous platforms will be discussed in detail.