Skip to content

Home

1. Overview

The IPU Toolchain can help users rapidly deploy AI models onto IPU chips.

The IPU Toolchain provides a development suite for users to perform model conversion, inference, and performance evaluation on their computers.

OpenDLAModel is an open-source reference algorithm based on the IPU Toolchain.

OpenDLAModel encapsulates the conversion commands of the IPU Toolchain and provides a complete set of code for modifying open-source frameworks to deploying on the IPU board.

IPU Toolchain

2. Purpose

By using the tools provided by the IPU Toolchain and the Python interface, the following functions can be conveniently completed:

  1. Model Conversion: Support for converting models from ONNX, TensorFlow, TensorFlow Lite, Caffe, etc. to IPU models.

  2. Quantization Functionality: Support for quantizing floating-point models to fixed-point models, with mixed quantization support.

  3. Model Inference: Distribute the IPU model to designated devices for inference and obtain inference results; or simulate the model on a computer and obtain inference results.

  4. Quantization Accuracy Analysis: This feature will provide metrics on the inference results of each layer of the quantized model compared to the floating-point model inference results, making it easier to analyze how quantization errors occur, which can provide insights for improving the accuracy of the quantized model.