NPU SDK Installation
NPU Overview
The Radxa Orion O6 features an NPU (Neural Processing Unit) with up to 28.8TOPs of computing power, supporting INT4, INT8, INT16, FP16, BF16, and TF32 acceleration types. This document explains how users can utilize the CIX P1 NPU SDK to run AI models and applications accelerated by the NPU. It includes explanations of the model compilation tools, toolchains, and common model cases.
CIX Early Bird Program
The NPU SDK includes the NOE Compiler and CIX AI Model Hub. To obtain the SDK, users must apply for the CIX Early Bird Program.
Registration link: CIX Early Bird Program
Install NPU SDK (x86 Linux Environment)
After applying through the CIX Early Bird Program, an email will be sent to the registered email address. Follow the email instructions to download the SDK.
tar -xvf nor.tag.gz
The CIX P1 NPU SDK is installed on an x86 host and requires Python version 3.8.
Users should prepare a Python 3.8 environment. It is recommended to use Anaconda.
Install NOE Compiler
The NOE Compiler is used for model compilation, converting ONNX framework models into formats that can be accelerated by the NPU.
cd noe
pip3 install -r requirements.txt
pip3 install ./CixBuilder-6.1.2958.1-py3-none-any.whl
Install NOE UMD (Orion O6)
The NOE UMD package is released in the form of a Debian installation package. If using the Python API, ensure to inherit system dependencies when creating a virtual environment:
python3 -m venv --system-site-packages .venv
After applying through the CIX Early Bird Program, an email will be sent to the registered email address. Follow the email instructions to download the UMD (User Mode Driver) Debian package.
bash sudo dpkg -i ./cix-noe-umd_0.01-1_arm64.deb ~
NOE UMD is already pre-installed in the OS. If you use the python api, please inherit the system dependencies when creating a python virtual environment: python3 -m venv --system-site-packages .venv to ensure that the libnoe module can be loaded correctly