A Study on Resources Utilization of Deep Learning Workloads
In this research, researchers will study and improve an existing deep RL (reinforcement learning) network, developed by the Google Brain team, that will place deep learning operations of another deep learning network in the most efficient manner within a heterogeneous computing system or a multi-GPU system. Finding the most efficient placement will reduce the training time of a deep learning network, therefore reducing the energy input needed for a computing system. The goal is to reduce the training time of this existing RL network to effectively find better placements faster than its original implementation. This will improve overall training efficiency as deep learning becomes more widespread.