MindSpore Realize decentralized autonomy through adaptive graph optimization driven by gradient data A11 Reduce, Gradient aggregation is in step, and calculation and communication are fully streamlined

Challenges of model execution under super chip computing power: Memory wall problems, high interaction overhead, and difficulty in data supply. Partly in Host Executed, partly in Device Execution, interaction

overhead is even much greater than execution overhead, resulting in low accelerator occupancy

MindSpore Through the chip-oriented depth map optimization technology, the synchronization wait is less, and the "data computing communication" is maximized. The parallelism of “trust”, compared with training

performance Host Side view scheduling method is flat

The challenge of distributed gradient aggregation under super chip computing power:ReslNet50 Single iteration 20ms Time will be generated The synchronization overhead of heart control and the communication

overhead of frequent synchronization. Traditional methods require 3 Synchronization completed A11 Reduce, Data-driven method autonomy A11 Reduce, No control overhead