Release Notes

Release Notes#

Version 6.3#

This is the public release of the AMD Vitis AI 6.3 User Guide for Versal AI Edge Series Gen2.

Key features#

  • Tool versions: Updated support for Vivado 2026.1 and Vitis 2026.1.

  • Windows boot flow: Support for Windows host-based OSPI and SD Card boot flows.

  • Advanced custom operator (NPU): Enhanced guidance for developing and integrating advanced AIE-based custom operators.

  • NPU resource reporting: Compilation-time reporting of NPU resource allocation, including all configurations and AIE tiles consumed.

  • Power measurement: Added section describing method and recommendations for measuring power.

  • Runtime DDR bandwidth profiling: Support and guidance for live memory profiling to help analyze and optimize memory usage.

  • VEK385 MIPI streaming AI reference design: Added end-to-end camera stream-based inference design, including complete source code for MIPI camera capture, ISP (Image Signal Processor), hardware-accelerated streaming preprocessing via PL kernels, NPU inference, and CPU-based postprocessing, with output displayed via the DisplayPort interface (supported only on Rev-B board).

  • Streaming preprocessing: Added streaming PL preprocess IP supporting resize, color space conversion, normalization, and cropping, with an AI Layout Formatter for writing preprocessed data to memory in multiple layout formats and data types.

  • Example applications and models: Added ready-to-run GStreamer pipeline for the camera stream-based inference design, demonstrating object detection (YOLOx) and scene classification (ResNet-50) pipelines using camera input.

  • Renamed Integrated System Design: Renamed Integrated System Design to VEK385 AI Reference Design.

  • Custom operators: Implement your own operators to extend what the NPU can run, or to replace a compiler-supported operator with a more performant one. You supply the AIE kernel, tiler, and configuration; Vitis AI compiles, partitions, and executes it as it would any built-in operator. See Custom operators.

  • AI-assisted workflows: Early access release of three optional AI skills that automate custom operator development, quantization, and compiler option tuning. Requires a language model that you supply, and an AI coding assistant to run the skills in. See AI-Assisted Workflows Reference.

Deployment runtimes#

Limitations#

Model and operator constraints#

  • ONNX opset: Opset support map can be seen here: System Requirements

  • Operator constraints: Certain ONNX operators are not supported and cause models to fall back to CPU execution. Refer to the ONNX Operators section in the user guide for detailed operator compatibility.

Model compilation constraints#

  • Large ONNX models: When a model exceeds 2 GB, it is stored using the ONNX external data format (model.onnx plus model.onnx.data). The full file path to model.onnx must be specified at runtime to ensure the companion .data file is correctly resolved and loaded.

  • Write permissions: The cache directory must have write permissions enabled during compilation. This allows the compiler to store generated artifacts necessary for the build process.

  • Docker stack size configuration: When launching Docker containers, use the --ulimit stack=-1:-1 option to allocate unlimited stack memory. This configuration is essential for compiling large models. - Compilation of large models may fail inside the Docker container: Large models can exceed the container’s default memory limit during Vitis AI compilation. The failure surfaces as an aiecompiler error and does not identify memory as the cause. Start the container with a raised memory limit, for example docker run --memory 64g.

Known system issues#

  • Permission requirements: Must run sudo -i on the target board to avoid permission issues when creating a hardware context.

Contents