Linked Presentation: A Unified Architecture for Accelerating Distributed DNN Training in Heterogeneous GPU/CPU Clusters
