GTC ON-DEMAND

 
SEARCH SESSIONS
SEARCH SESSIONS

Search All
 
Refine Results:
 
Year(s)

SOCIAL MEDIA

EMAIL SUBSCRIPTION

 
 

GTC ON-DEMAND

Deep Learning & AI Frameworks
Presentation
Media
Node-Level Deep Learning Input Pipeline Optimization on GPGPU-Accelerated HPC Systems
Abstract:
Learn how to implement and analyze a simple deep learning input pipeline pattern that prevents slowdowns from input queue exhaustion on accelerated HPC systems with limited impact to model performance. Queue exhaustion occurs because the throughput-driven dequeue rate is greater than the enqueue rate, which is bound by storage access bandwidth. In this session we will describe a technique that prevents queue exhaustion by artificially slowing the effective dequeue rate, without sacrificing substantial validation set performance. An example using TensorFlow is presented, and the resultant optimization step speedup and model performance are analyzed across several HPC resource configurations.
 
Topics:
Deep Learning & AI Frameworks, HPC and Supercomputing
Type:
Talk
Event:
GTC Silicon Valley
Year:
2018
Session ID:
S8674
Streaming:
Download:
Share: