A new AWS Batch scheduling capability divides GPU capacity into quota shares per team or project, lends idle capacity across shares and preempts jobs so higher-priority SageMaker Training work runs first without manual coordination.
Read it at AWS Machine Learning →