Abstract
The computational resources required for deep reinforcement learning (RL) model training are challenging for many applications with computational resource constraints. Low-precision methods are a promising avenue to alleviate computational bottlenecks and reduce the time, memory, and power required for model training. A notional drawback to low-precision training methods is that the benefits depend on model complexity and Graphical Processing Unit (GPU) hardware capabilities. In this work, we will address this notion and gap in the literature and investigate how the benefits of low-precision training methods scale with GPU capabilities and model complexity by benchmarking GPU resource usage using state-of-the-art system profiling methods for Soft Actor-Critic (SAC) deep RL agents trained on control environments for a range of model configurations and NVidia GPUs spanning from edge devices to data science grade desktop cards.