Case Study: Worcester Polytechnic Institute achieves faster, cheaper deep learning training with Google for Education

Researchers at WPI use Google’s pre-emptible VMs to improve deep learning training

Worcester Polytechnic Institute (WPI) wanted to find a more efficient and affordable way to train deep learning models in the cloud, especially for research workloads that can take a long time and require significant compute resources. In partnership with Google for Education, the team explored whether Google Compute Engine pre-emptible VMs could be used for distributed training instead of relying only on on-demand servers.

Using Google for Education’s pre-emptible virtual machines with TensorFlow on Google Compute Engine, WPI tested multiple GPU configurations and compared training time, cost, and accuracy. The results showed that an eight-node K80 pre-emptible cluster was up to 7.7 times faster and 62.9% cheaper on average than a single K80 on-demand server, with revocations mainly affecting training time but not accuracy.

Worcester Polytechnic Institute

Robert Walls

Assistant Professor of Computer Science