Session II: Mixed-Precision Training of Deep Neural Networks on AMD Vega GPUs

Today, the complexity and data set of deep neural networks grow exponentially. Using reduced precision is a promising solution, which can reduce both memory footprint and execution time. In this talk, we’ll first describe Vega GPU ISA extension and techniques in our MIOpen library for mixed-precision training on AMD GPUs. And then, we will talk about real-world training in mixed precision with the TensorFlow framework.

Download Presentation

Location: Grand Ballroom C Date: March 27, 2019 Time: 11:30 am - 11:50 am Jing Zhang, AMD