Intel
Overview
OAAX provides a runtime and conversion toolchain for running ONNX models on Intel CPUs, GPUs and NPUs on both Ubuntu and Windows.
It uses the ONNX runtime and OpenVINO™ toolkit under the hood to offload the computation to Intel hardware.
For more information about OpenVINO please visit the OpenVINO website.
Requirements
Hardware
Please consult the OpenVINO documentation for the most up-to-date hardware requirements.
Operating System
- Debian 11 or higher
- Ubuntu 20.04 or higher
- Windows 10, 11
Installation
To be able to use the Intel GPU and NPU, please make sure to install the latest drivers from Intel:
Usage
Runtime Library
The runtime library implements the OAAX's runtime interface for initializing, loading, running inference and destroying the runtime.
The initialization in particular can be done without providing a configuration by calling int runtime_initialization(); directly or by providing these parameters to int runtime_initialization_with_args(int length, char **keys, void **values);:
log_level(char *, default is "2"): The minimal log level for the runtime. This can be set to0for trace,1for debug,2for info,3for warnings,4for errors,5for critical and6to disable logging.log_file(char *, default is "runtime.log"): The file to which the runtime logs will be written. If not specified, logs will be written to stdout.num_threads(char *, default is "8"): The maximum number of threads that can be used by the runtime. The higher the number, the more CPU resources will be used, but the better the throughput.device_type(char *, default is "CPU"): The type of device to use for inference. This can be set toCPU,GPU, orNPU. The default isCPU.precision(char *, default is "FP32"): The precision to use for inference. This can be set toFP32,FP16, orACCURACY. The default isFP32.
You can check out the examples repository for more details on how to use the runtime library: OAAX Examples.
Conversion Toolchain
The conversion toolchain is used to validate, optimize and simplify the ONNX models. At the end of the process it produces a simplified ONNX model.
It can be used as follows:
docker run -v ./model:/model oaax-intel-toolchain:1.1.1 /model/model.onnx /model/output
The above command assumes that the model is located at ./model/model.onnx.
After a succesful conversion, the generated model will be saved in the ./model/output directory.
Download links and compatibility matrix
| OAAX versions | OS | Version | CPU architecture | Runtime library | Conversion toolchain |
|---|---|---|---|---|---|
| 1.1.1 | Ubuntu | 22.04, 24.04 | x86_64 | Download | Download |
| 1.1.1 | Windows | 10, 11 | x86_64 | Download | Download |