AUTOMATIC GPU PARALLELIZATION  WITH TensorFlow

 

Hannah Sechaga
Prof. Timothy Wiser, Faculty Mentor

Here, We aimed to show how TensorFlow can do computation much more efficiently than the commonly used Numpy across a range of tasks. TensorFlow is a machine learning system that operates at a large scale across multiple computing devices such as CPU, GPU, and TPU. In this presentation, I will mainly focus on the method that uses GPUs for physics calculations ranging from a simple force computation of a 1-dimensional simple harmonic motion to a 3 - dimensional gravitational force. We will also analyze how TensorFlow GPU uses parallelism through replication and parallel execution of the core model dataflow graph with many different computational devices to give us a faster run-time as compared to Numpy with only a slight modification in the algorithm. Our study revealed that TensorFlow running on GPU gives us a faster run-time as compared to TensorFlow CPU and Numpy for a large number of initial conditions.

 

Keywords: 

Topic(s):Physics

Presentation Type: Oral Presentation

Session: TBA
Location: TBA
Time: TBA