tutorials

Graphcore tutorials

These tutorials provide hands-on programming exercises to enable you to familiarise yourself with creating, running and profiling programs on the IPU. They are part of the Developer resources provided by Graphcore: https://www.graphcore.ai/developer.

Each of the tutorials contains its own README file with full instructions.

Poplar

Poplar is the underlying C++ framework for developing and executing code on the Graphcore IPU. It provides a generic interface on which the other frameworks are built.

• Tutorial 1: Programs and variables
• Tutorial 2: Using PopLibs
• Tutorial 3: Writing vertex code
• Tutorial 4: Profiling output
• Tutorial 5: Matrix-vector multiplication
• Tutorial 6: Matrix-vector multiplication optimisation

TensorFlow 2

Getting started with the IPU:

• Starter tutorial: MNIST training example
• TensorFlow 2 Keras: How to run on the IPU

Exchanging data between the host and the IPU:

• TensorFlow 2: How to use infeed/outfeed queues

Debugging and analysis:

• TensorFlow 2: How to use TensorBoard

PyTorch

Getting started with the IPU:

• Starter tutorial: MNIST training example
• From 0 to 1: Introduction to PopTorch

Exchanging data between the host and the IPU:

• Efficient data loading with PopTorch

Maximising compute on the IPU:

• Half and mixed precision in PopTorch

Using multiple IPUs and handling large models:

• Parallel execution using pipelining
• Fine-tuning BERT with HuggingFace and PopTorch

Debugging and analysis:

• Observing tensors in PopTorch

Running a Hugging Face model on the IPU:

• Fine-tuning a HuggingFace Vision Transformer (ViT) on the IPU using a local dataset

PopVision

PopVision is Graphcore's suite of graphical application analysis tools.

• Instrumenting applications and using the PopVision System Analyser
• Accessing profiling information with libpva
• Reading application instrumentation from PVTI files
• Profiling output with the PopVision Graph Analyser is currently included in Poplar Tutorial 4: profiling output

PopXL and popxl.addons

PopXL and popxl.addons are Graphcore frameworks which provide low level control of the IPU through an expressive Python interface designed for machine learning applications.

• Tutorial 1: Basic Concepts
• Tutorial 2: Custom Optimiser

Improving performance and optimising throughput:

• Tutorial 3: Data Parallelism
• Tutorial 4: Pipelining
• Tutorial 5: Remote Variables
• Tutorial 6: Phased Execution

Standard tools

In this folder you will find explanations of how to use standard deep learning tools with the Graphcore IPU. Guides included are:

• Using IPUs from Jupyter Notebooks
• Using VS Code with the Poplar SDK and IPUs