Don’t Miss the Train: A Case for Systems Research into Training on the Edge

Abstract

In this position paper, we explore the systems research challenges offered by DNN training on accelerated edge devices. Specifically, we examine the Nvidia Jetson devices, and explore system tuning opportunities offered by frequency and power controls, workload-specific hardware configurations, utilizing shared memory effectively, minimizing data stalls, etc. Our observations are based on first-hand experience with 20 Nvidia Jetson devices – Xavier AGX, Xavier NX, Nano, TX2 and TX1 development kits which have 2-8 ARM cores at 1.43-2.26 GHz, and a peak power of 10W–30W.