Under the hood
The VoyagerTM SDK builds on industry-standard APIs and open-source frameworks, enhanced with advanced capabilities by Axelera® R&D team. Voyager’s Machine Learning compiler, built on the Apache TVM compiler framework, automates the compilation and optimization of models for the Metis AI Processing Unit. The compiler inputs models pretrained in industry-standard frameworks such as PyTorch, and outputs code tuned for Metis hardware. During compilation, the compiler quantizes the model using proprietary, state-of-the-art algorithms and partitions the model for optimal execution on Metis AIPU. Without manual intervention, the compiler generates code with an accuracy practically indistinguishable from the original model. Similarly, pipelines use the open source framework GStreamer and whilst the vast majority of users need not understand any of the internals of generated pipelines, the open nature of the stack allows expert users to customize the generated code for a specific use case.
Applications deployed with Voyager onto the Metis AIPU have the highest performance and energy efficiency, while retaining equivalent accuracy of the original FP32 model.
Read on
AI at the Edge: A fast, accurate and effortless journey with Voyager SDK
February 27, 2023
| Development machine: Python project with native framework support for PyTorch/Tensor Flow | ||||||||||||||||||||
| Input form dataloader | Normalize | Scale & crop | Convert to tensor | Model | NMS | Calculate metric and visualise stream |
||||||||||||||
| ↓ Deploy to the Edge | ||||||||||||||||||||
| Inference host CPU: C++ optimized code with streamer and dma-buf for efficient buffer sharing and synchronization between computing elements | ||||||||||||||||||||
| Input from camera or socket (h264 decode) | YUV-to-RGB | Normalize | Scale & crop | Convert to tensor | Model | NMS | C++/Python callback | Visualize | Output to display or socket | |||||||||||
| ↑ Optimize at the Edge | ↑ | ↑ | ↑ | |||||||||||||||||
| Host Media Accelerator: Input from camera, YUV-to-RGB convert, scale and crop |
Axelera AI: Optimized Model |
Customer application code |
Host GPU: OpenGL kernel |
|||||||||||||||||