Compliant AI Accelerators
These accelerators have an OAAX 2.0 implementation: a runtime library and a conversion toolchain. Open a chip's page for its requirements, downloads, options and a complete example.
| Device class | Hardware | Page | Verified on |
|---|---|---|---|
| CPU | x86_64 with AVX2 (Intel, AMD), aarch64 with NEON (Arm) | CPU | x86_64 Linux. Not yet verified: aarch64, Windows |
| CPU | Intel CPUs with AVX2 | Intel | Yes |
| GPU | Intel GPUs | Intel | Integrated GPU |
| GPU | NVIDIA GPUs with CUDA 11, 12 or 13, Jetson, DGX Spark | NVIDIA | RTX A4000: ONNX Runtime and TensorRT 10. Not yet verified: Jetson, DGX Spark, Windows, TensorRT 11 |
| NPU | Intel NPU | Intel | Verified on Linux. Not supported on Windows yet |
| NPU | Hailo-8 | Hailo | Runtime on Hailo-8 hardware (HailoRT 4.20.0). Conversion needs Hailo's SDK |
| NPU | DEEPX DX-M1, DX-H1 | DEEPX | DX-M1: model conversion and inference (DX-RT 3.3.2) |
A listed chip means an OAAX 2.0 runtime and toolchain are published for it. The last column is what has been verified on real hardware; each page lists the details.
Fetch any runtime or toolchain with OAAX Manager, then follow the Hello World example.
Want your accelerator on this list? See the Contributing Guide and the Specification.