Improving Training Speed Of Neural Networks Using Input Pipeline

Authors

  • Muthumanikandan V, Bhuvaneswari A, Braveen M

Abstract

Neural Networks are used for solving all types of tasks and hence undergo a lot
of training time. GPUs radically help in reducing the time involved in executing a single
training step. However, there still lies a condition of idle CPU/GPU time which can be
significantly higher in case of larger models and datasets. In this paper, the HighPerformance Computing strategies like data pipelining on the Neural Networks to
significantly improve the training speed and lower the training time. If the idle time is
minimized or removed, then a better and lower training time can be achieved. To improve
and reduce the training speed of neural networks using the concept of data pipelines is
proposed. Without, pipelining the CPU and the Graphics Processing Unit (GPU) /Tensor
Processing Unit (TPU) will stay idle for more time but with pipeline idle time diminishes
significantly.

Published

2020-12-02

Issue

Section

Articles