OpenViSTAI · Agents · Blockchain · Crypto
Submit a project
Menu
AI Infrastructure & Serving

Open AI Accelerator Exchange

Open AI Accelerator Exchange (OAAX) is an open standard for converting ONNX models into accelerator-specific formats and running them through a unified runtime interface. Its workflow connects models from frameworks such as TensorFlow, PyTorch and Hugging Face with compatible CPU and accelerator backends, including edge hardware.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

Open AI Accelerator Exchange — OpenViST project profile in AI Infrastructure & Serving.
OpenViST project illustration. See the linked sources for product details.
Understand the project

Practical overview

When it fits

Use OAAX when deploying an ONNX-exported model to an accelerator with an OAAX-compatible conversion toolchain and runtime, particularly when a shared host interface across hardware targets is useful.

Who it is for

  • AI developers integrating models with different accelerator platforms.
  • Teams targeting edge deployments across heterogeneous compute hardware.
  • Implementers creating or contributing OAAX-compatible conversion toolchains and runtime libraries.

Documented uses

  • Convert an ONNX model into a format optimized for a selected accelerator.
  • Load an accelerator-specific model through a shared runtime interface.
  • Send model inputs and retrieve outputs asynchronously from a host application.
  • Assess whether an accelerator has an OAAX implementation before planning a deployment.
Before you choose

Requirements & limits

  • The documented workflow starts with a model exported to ONNX.
  • Each accelerator requires a compatible OAAX implementation, including both conversion and runtime components.
  • Conversion creates an XPU-specific bundle or binary, so target hardware determines the conversion path.
  • The host application interacts with the runtime through initialization, model loading, asynchronous tensor exchange and cleanup.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Deploy an ONNX model through an OAAX runtime

Follow the documented flow for converting an ONNX model and using it from a host application.

Before you start

  • An AI model exported in ONNX format.
  • An OAAX-compliant XPU with its conversion toolchain and runtime library.
  • A host application able to use the runtime interface.

Steps

  1. Use the XPU’s conversion toolchain to convert the ONNX model into an XPU-specific OAAX bundle or binary.

  2. Load the runtime library appropriate to the XPU in the host application and initialize it with runtime_initialization().

  3. Load the converted model with runtime_model_loading("./model.oaax").

  4. Exchange data asynchronously using send_input(input_tensors) and receive_output(output_tensors_holder), then release resources with runtime_destruction().

Expected outcome

The host application can send inputs to the converted model and retrieve outputs through the OAAX runtime interface.

Guide limitations

  • The target accelerator must have an OAAX implementation that includes both a conversion toolchain and runtime library.
  • Conversion produces a format specific to the selected accelerator; the workflow does not make one binary universal across hardware.
Documentation 1 ↗
Evidence and transparency

Sources & review

oaax.org ↗Guide documentation ↗
Source record details
Source availability
Active
Sources checked
12 October 2026
Original website title
Open AI Accelerator Exchange | Bridging the Gap Between AI Models and XPUs · Official homepage title · retained in its source language
How we review sourcesSuggest a correction