A GPU shared scheduling strategy provides single-device sharing of GPU resources and supports the sharing of up to 64 tasks per GPU. Allocation and isolation at any granularity are supported. Users can dynamically request GPU resources based on the GPU memory.
The strategies of "zero-copy" transmission, multi-thread fetch, incremental data update and affinity scheduling for training data greatly shorten the data cache cycle and improve efficiency of model development and training.
Supports the extension of distributed training through MPI in TensorFlow, PyTorch and other mainstream frameworks and provides standard UI operations, so that users can submit distributed training through simple GPU resources and training script configurations.
Provides fault tolerance for training tasks, enabling the platform to effectively ensure continuous training of tasks and reduce the recovery time in the event of a server crash or GPU failure.
The resnet50 benchmark test shows that the AI training efficiency using an AIStation data cache strategy significantly improved with an increase in the number of concurrent tasks. With 70 concurrent tasks, the efficiency of model training improved by 72%.
For distributed training for resnet50, with the increase of task concurrency, the GPU acceleration ratio of multi-card distributed training with AIStation can be increased by up to 90%.