Towards Energy-Efficient Neural Network Training on the Cloud for Effective Inference on IoT/Edge Devices

• Yasutaka, Wada

• Ken'ya, Onai

• Musashi, Aoto

抄録

IoT/Edge devices need to be low-power, and it is required to enhance their computational power by employing hardware accelerators like FPGAs and by offloading heavy workloads to the cloud side. However, maintaining the cloud environments at a low power is challenging because of their unstable workloads with virtualization. This paper explains our idea and strategy to realize energy-efficient deep learning computation on virtualized cloud platforms and IoT/edge devices. We propose to utilize cloud servers to provide sufficient computational resources for neural network training and its model optimizations. Then, IoT/edge devices can focus on inference tasks while accelerating the tasks with FPGAs. Based on this strategy, we are developing a framework to minimize the power consumption of virtualized cloud servers considering the difference in computational workloads between deep learning training tasks and High-Level Synthesis tasks.

IoT/Edge devices need to be low-power, and it is required to enhance their computational power by employing hardware accelerators like FPGAs and by offloading heavy workloads to the cloud side. However, maintaining the cloud environments at a low power is challenging because of their unstable workloads with virtualization. This paper explains our idea and strategy to realize energy-efficient deep learning computation on virtualized cloud platforms and IoT/edge devices. We propose to utilize cloud servers to provide sufficient computational resources for neural network training and its model optimizations. Then, IoT/edge devices can focus on inference tasks while accelerating the tasks with FPGAs. Based on this strategy, we are developing a framework to minimize the power consumption of virtualized cloud servers considering the difference in computational workloads between deep learning training tasks and High-Level Synthesis tasks.

収録刊行物

• Proceedings of Asia Pacific Conference on Robot IoT System Development and Platform

  Proceedings of Asia Pacific Conference on Robot IoT System Development and Platform 2019 38-39, 2020-02-25

詳細情報

• CRID: 1050011097117631744
• NII論文ID: 170000181732
• Web Site: http://id.nii.ac.jp/1001/00203366/
• 本文言語コード: en
• 資料種別: conference paper
• データソース種別:
  • IRDB
  • CiNii Articles