AN02008: Reducing Memory Pressure in Neural Networks using Operator Splitting

Summary¶

This app note demonstrates how you can compile larger neural networks onto an XCORE device by using a feature called operator splitting. Operator splitting enables you to cut single large parts of a neural network into smaller ones, thereby reducing the amount of scratch memory (also known as tensor arena) required by the program.

Features¶

• Operator splitting demonstration

Known issues¶

• None

Required tools¶

• XMOS XTC Tools: 15.3.1

• xmos_ai_tools

Required libraries (dependencies)¶

• None

Related application notes¶

• None