Download as PDF, PPTX
![OAAX from a user perspective
11
User trains and exports
model to ONNX
input_tensor = torch.rand((1,3,256,256));
model = MyModel();
torch.onnx.export(model, input_tensor,
"model.onnx", input_names=["input"])
Compile ONNX to the
XPU format
$ docker pull
onnx-to-xpu:latest
$ docker run
onnx-to-xpu:latest
input.onnx model.xpu
Run the model on the
XPU
int main() {
runtime = load_runtime("./xpu.so");
runtime.runtime_initialization();
runtime.runtime_load("./model.xpu");
while (true) {
runtime.send_input(input);
tensors_struct output;
runtime.receive_output(output);
}
runtime.runtime_destruction();
return 0;
}
© 2025 Network Optix](https://image.slidesharecdn.com/e1r04vanemdennetworkoptix2025-250602130816-fa7f49bf/75/OAAX-One-Standard-for-AI-Vision-on-Any-Compute-Platform-a-Presentation-from-Network-Optix-11-2048.jpg)
Uploaded byEdge AI and Vision Alliance
“OAAX: One Standard for AI Vision on Any Compute Platform,” a Presentation from Network Optix
For the full video of this presentation, please visit: https://www.edge-ai-vision.com/2025/06/oaax-one-standard-for-ai-vision-on-any-compute-platform-a-presentation-from-network-optix/ Robin van Emden, Senior Director of Data Science at Network Optix, presents the “OAAX: One Standard for AI Vision on Any Compute Platform” tutorial at the May 2025 Embedded Vision Summit. OAAX, an open accelerator compatibility standard being developed under the Linux Foundation AI & Data project, is key to making AI vision truly hardware-agnostic. In this talk, van Emden highlights recent progress and his company’s vision for seamless interoperability across CPUs, XPUs and GPUs. He also explores how developers and processor suppliers can benefit from and contribute to the standard. Most importantly, he discusses how contributors can help shape OAAX to ensure its growth and widespread adoption. By fostering a strong community, OAAX has the potential to become a widely accepted industry standard, enabling AI models to run efficiently across diverse hardware without vendor lock-in. As AI vision continues to evolve, open collaboration is essential for driving innovation, simplifying development and expanding accessibility. Join Network Optix in shaping the future of AI acceleration.
More Related Content
Similar to “OAAX: One Standard for AI Vision on Any Compute Platform,” a Presentation from Network Optix
![Vibe Coding vs. Spec-Driven Development [Free Meetup]](https://cdn.slidesharecdn.com/ss_thumbnails/vibecodingvsspecdrivendevelopment-251209105622-43f455e7-thumbnail.jpg?width=640&height=640&fit=bounds)
“OAAX: One Standard for AI Vision on Any Compute Platform,” a Presentation from Network Optix
- 1. OAAX: One Standardfor AI Vision on Any Compute Platform Robin van Emden Senior Director of Data Science Network Optix
- 2. OAAX © 2025 NetworkOptix 2 Open AI Accelerator eXchange
- 3. Outline • The problemOAAX addresses • Proposed solution • OAAX’s impact on chipmakers and users • Ecosystem and collaboration • OAAX in practice • Roadmap and future development • Get involved: Linux Foundation project & community resources © 2025 Network Optix 3
- 4. The value ofstandardized interfaces • The standard design of cars: • has consistent features like pedals, gears, dashboard, trunk... • boosts usability and mass adoption • eliminates the need for training or license for each car brand • Likewise, OAAX: • has a unified interface • boosts mass adoption of AI accelerators • allows ease of integration of any OAAX- compliant AI accelerator 4 © 2025 Network Optix
- 5. The problem thatgave rise to OAAX 5 Deploying on XPU X 1.Learn about X API & docs 2.Setup development env 3.Experiment with examples 4.Adapt example code 5.Integrate custom code 6.Compile code for target Deploying with OAAX $ docker run onnx-to-xpu:latest input.zip model.xpu runtime.load(“model.xpu”); runtime.run(tensors_in); Deploying on CPU/GPU $ pip install torch import torch model = torch.load(...) model = model.cuda() # model = model.cpu() Before Now Next Download runtime library © 2025 Network Optix
- 6. Solution using OAAX ©2025 Network Optix 6 OAAX proposes a standardized interface to compile a model and a process to run it. OAAX toolchain OAAX runtime car, truck, person $ docker run -v ... onnx-to-xpu:latest input.zip output.xpu send_input (tensors_struct *input_tensors); model_load (char *file_path); Edge Device tensors_struct **output_tensors Offline
- 7. Impact of OAAX 7 ©2025 Network Optix OAAX for AI chip makers ● Makes it easier to deploy AI models ● Reduces time, effort and money for deploying a trained model ● Allows them to focus on training models ● Enables smooth transition from one AI Accelerator to another ● Reduces time to market for companies ● Simplifies the process of showcasing their hardware, saving time and effort ● Enables new XPUs to compete on equal footing with existing ones ● Allows them to access broad segment of users ● Enables them to access the ecosystem of examples and tools already developed OAAX for users
- 8. Current state ofOAAX 8 • OAAX simplifies AI deployment: It enables effortless deployment of AI models across various hardware platforms. • Proven in real-world use: As the first adopter, Network Optix uses OAAX to deploy AI models from the Nx Cloud to Nx supported servers and devices at the edge. • Open and collaborative: OAAX is a project under the Linux Foundation. • Broad hardware support: OAAX provides toolchains and runtimes at different levels of maturity for a growing list of AI accelerators, including: © 2025 Network Optix
- 9. Ecosystem of OAAX 9 OAAX AIChip Makers End Users: • AI enthusiasts • Researchers • Companies Companies: • Network Optix • Edge Impulse • Automobile AI Frameworks: • PyTorch • TensorFlow • Huggingface Users / Developers © 2025 Network Optix
- 10. OAAX in practice– general overview © 2025 Network Optix 10 Offline On Device Input preprocessing OAAX Runtime (.so shared lib) Output postprocessing Input Output • Resize • Pad • Normalize • Filter • Save • Etc OAAX conversion toolchain (Docker image) Optimized file deployed on hardware. Logs file
- 11. OAAX from auser perspective 11 User trains and exports model to ONNX input_tensor = torch.rand((1,3,256,256)); model = MyModel(); torch.onnx.export(model, input_tensor, "model.onnx", input_names=["input"]) Compile ONNX to the XPU format $ docker pull onnx-to-xpu:latest $ docker run onnx-to-xpu:latest input.onnx model.xpu Run the model on the XPU int main() { runtime = load_runtime("./xpu.so"); runtime.runtime_initialization(); runtime.runtime_load("./model.xpu"); while (true) { runtime.send_input(input); tensors_struct output; runtime.receive_output(output); } runtime.runtime_destruction(); return 0; } © 2025 Network Optix
- 12. Conversion toolchain interface ©2025 Network Optix 12 ● A process to go from a generic model specification (ONNX), to a specification that can be executed by the target specific runtime. ○ Implemented as a standalone Docker image: easy to distribute and to maintain, while being totally flexible regarding the implementation of the conversion process. ○ Usage: > docker run -v .:/app onnx-to-xpu:latest /app/input.onnx /app/output.xpu OAAX conversion toolchain (Docker image) Optimized file deployed on hardware. Logs file
- 13. Runtime interface -process 13 Runtime initialization Model loading Receiving inputs Returning outputs Runtime destruction int runtime_initialization(); int runtime_model_loading(const char *file_path); int send_input(tensors_struct *input_tensors); int receive_output(tensors_struct *output_tensors); int runtime_destruction(); © 2025 Network Optix
- 14. Runtime interface -definition © 2025 Network Optix 14 int runtime_initialization(); int runtime_initialization_with_args(int length, char** keys, void** values); int runtime_model_loading(const char* path); int send_input(tensors_struct* input_tensors); int receive_output(tensors_struct** tensors_out); int runtime_destruction(); typedef struct tensors_struct { size_t num_tensors; char** names; tensor_data_type* data_types; size_t* ranks; size_t** shapes; void** data; } tensors_struct; - Minimalist. - Abstract. - Works for any kind of AI model: vision, LLM, multimodal. The interface is:
- 15. OAAX roadmap 15 ● Supportmore AI chips: SiMa, Axelera, RockChip, etc. ● Improve the standard to accommodate for needs of both users & AI chip makers. ● Implement tools to make it easier for users to the toolchains & runtimes at scale. ○ Auto-detection of supported AI accelerators on a machine. ○ Download the correct runtime and toolchain programmatically. ○ Benchmark and profiling. © 2025 Network Optix
- 16. OAAX Linux FoundationProject 16 🌟 Support the Future of Open AI Acceleration! OAAX is now a Sandbox Project under the Linux Foundation AI & Data. We’re building an open, collaborative ecosystem for AI acceleration — and we invite developers, companies, and innovators like you to join us! 🔧 Contribute your expertise 🤝 Collaborate on groundbreaking solutions 🚀 Help shape the future of AI infrastructure 👉 Support OAAX — Let’s build the future, together! © 2025 Network Optix
- 17. OAAX Linux FoundationProject – Reach out 17 ● Sponsor: Network Optix https://networkoptix.com ● Tech lead: Ayoub Assis aassis@networkoptix.com ● OAAX Team: Robin van Emden rvanemden@networkoptix.com Josef Joubert jjoubert@networkoptix.com Maurits Kaptein mkaptein@networkoptix.com © 2025 Network Optix
- 18. Links to OAAXrelated resources: GitHub repository: https://github.com/OAAX-standard Project home: https://www.oaax.org/ Linux Foundation AI & Data project site: https://lfaidata.foundation/projects/oaax 2025 Embedded Vision Summit We warmly invite you to visit us at the Network Optix booth, #302. We’d love to connect and show you what we're working on! © 2025 Network Optix 18 OAAX Linux Foundation Project – Resources
- 19. OAAX: One Standardfor AI Vision on Any Compute Platform Robin van Emden