Webinar Series: Profiling and optimizing your model on Habana Gaudi

Habana SynapseAI^® Software Platform makes it easy to build new and migrate existing AI models with highly cost-effective Gaudi AI Training processors. SynapseAI is integrated with TensorFlow and PyTorch, and we support a variety of CV and NLP models.

In this live four-session mini-series you’ll learn how to port and optimize your existing TensorFlow and PyTorch deep learning models for Habana Gaudi AI processors. You will be introduced to techniques to analyze your model and identify potential bottlenecks. A live Q&A will follow each session.

Agenda

In sessions 1 and 2, we will focus on the TensorFlow framework.

• Session 1, January 25^th, 2022 9AM to 10AM PST – View on-demand recording
  
  We will start with the basic steps to migrate your TensorFlow model and train on Gaudi® processors. We will cover the two options available for distributed training – Horovod and tf.distribute with Habana Processor Unit (HPU) strategy. We will introduce techniques to analyze the model and identify potential performance bottlenecks, including using TensorBoard and SynapseAI® profiler.

• Session 2, February 8^th, 2022 9AM to 10AM PST – View on-demand recording
  We will share tips for addressing common performance bottlenecks and optimization techniques such as mixed precision enabling. We will also share how to use the Tensor Processor Core (TPC) programming tools to implement custom TPC kernels with custom operators in the TensorFlow framework.

In sessions 3 and 4, we will focus on PyTorch framework.

• Session 3, February 22^nd, 2022 9AM to 10AM PST – View on-demand recording
  You will learn the basic steps to migrate your PyTorch model and train on Gaudi and we will show how to enable distributed training with PyTorch DDP and introduce techniques to analyze the model and identify potential performance bottlenecks.

• Session 4, March 8^th, 2022 9AM to 10AM PST – View on-demand recording
  This session will provide tips for addressing common performance bottlenecks and optimization techniques such as mixed precision enabling. The session will address how to use the TPC programming tools to implement custom TPC kernels with custom operators in the PyTorch framework.

We are experimenting with the mini-series webinar format. For those who are interested, you are welcome to make it an active learning exercise by porting and optimizing your model on a Gaudi-based AWS EC2 DL1 instance as we progress through the mini-series.