Optimising Temporary Memory Usage for Convolutions and Matmuls on the IPU

• 1. Introduction to the available memory proportion option
  • 1.1. The convolution planner
  • 1.2. Setting availableMemoryProportion to a very small value
• 2. Worked example
  • 2.1. Use case 1: fitting a model or pipeline stage
  • 2.2. Use case 2: Optimising for throughput
• 3. How to set the available memory proportion
  • 3.1. TensorFlow
  • 3.2. PyTorch
  • 3.3. PopART
• 4. Legal notices