Reference Design Overview#
Vitis AI includes a reference design that integrates hardware-accelerated pre-processing and runs the machine learning inference on the Neural Processing Unit (NPU). This design showcases a complete, end-to-end pipeline from ingesting image or video inputs to producing and saving the inference results.
Supported Features:#
Hardware-accelerated pre-processing via the Image Processing PL kernel, including resizing, color space conversion, normalization, and cropping.
Inference execution on the NPU.
Software-based post-processing on the CPU for Resnet50 model
Visualization utilities to overlay bounding boxes and labels (text) on the input image based on inference results.
Reference Design Includes:#
Source code for the System Integrated Reference Design.
Sample applications that demonstrate ML inference accelerated on the AI Engines, hardware-accelerated pre-processing using the image_processing PL HLS(High Level Synthesis) kernel, and CPU-based post-processing.
Ready-to-test boot images that embed the application binary for immediate evaluation on target hardware.
Requirements#
To build the reference design and deploy it on board, the following setup are required:
System Requirements System Requirements
Apply the Vivado patch contained in vivado_patches.zip
Download Reference Design:#
Ensure the reference design has been extracted per the instructions in Software Setup. Download packages from the AMD Vitis™ AI lounge.
Integrated System Reference Design source code:
vitis_ai_2ve_source_v6.2.tarPrebuilt boot images (RevA):
vitis_ai_2ve_prebuilt_boot_images_v6.2_RevA.tarPrebuilt boot images (RevB):
vitis_ai_2ve_prebuilt_boot_images_v6.2_RevB.tarSDK to cross-compile applications on the host machine:
vitis_ai_2ve_sdk_v6.2.sh
The release package includes:
Boot images
ONNX model
VART-based and ONNX Runtime based applications
Source code for the VAIML Platform Reference Design
Deploying Pre-built Reference Design Images on the Board#
AMD provides a set of prebuilt C++ applications that you can run on the board for performance evaluation. For more details on running these pre-built applications see Reference Applications.
Building from Source#
To build the reference design yourself, see (Build the Reference Design).