Accelerate AI Workloads with Intel® Optimization for PyTorch*

Deep learning is massively popular in scientific computing, with deep learning algorithms used by industries to solve complex, computationally sophisticated problems in real time. Find out how tools optimized with Intel oneAPI can boost training and inference performance of big models.

This workshop introduces Intel® Extension for PyTorch* (part of Intel® Optimization for PyTorch*), which extends stock PyTorch with optimizations for an extra performance boost on Intel architecture.

While most of these optimizations will eventually be included in stock Python*, the extension purpose is to deliver timely, up-to-date features and optimizations for PyTorch on Intel hardware.

This workshop provides:

• An introduction to the framework, including the optimizations Intel contributes to PyTorch
• Hands-on practice on Intel® Developer Cloud with a demonstration of using Intel Extension for PyTorch to accelerate a convolutional neural network (CNN) model on Intel platforms
• Instructions for installing Intel Extension for PyTorch and applying special techniques to achieve an AI performance boost

00:00 Introductions

1:53 Overview of the workshop

3:38 Intel Optimization for PyTorch

8:17 An overview of the AI Tools

14:21 The main ways to optimize PyTorch

15:55 An overview of PyTorch

21:06 What components does operator optimization include?

21:30 Vectorization

24:37 Parallelization

25:55 Memory layout

29:00 Low-precision optimization

30:12 How to fully take advantage of bfloat16 features

32:26 Low-precision optimization with int8

33:27 Quantization features to take advantage of

34:34 Quantization workflow and API

36:54 Introduction to graph optimization

37:00 Operator fusion

38:08 FP32 and bfloat16 fusion patterns

38:56 Constant folding

40:00 Introduction to runtime extension and motivation

41:40 Example of PyTorch that uses a runtime extension API

44:26 How a multistream module can reduce core traffic

45:03 Case study

45:23 Introduction to a BERT model

46:22 Performance boost results for the Intel Optimization for PyTorch

47:52 How to install Intel Extension for PyTorch

52:00 Hands-on session