aie4ml is an end-to-end compiler that generates optimized AIE firmware automatically, which can be then built and simulated directly using AMD Vitis. It targets the AMD AI Engine (AIE) from model-level frontends and lowers supported operators into AIE graphs and kernels as a standalone AIE project.

• Current hardware targets: AIE-ML and AIE-MLv2 devices.
• Current frontend paths: ONNX for explicit operator graphs, and an optional hls4ml frontend path.

Current Support

aie4ml currently supports Dense/GEMM, dynamic MatMul, Elementwise Add, quantized LayerNorm, HCCS Softmax, final-two-axis Permute, Split/Slice, Concat, fanout, and fused ReLU across AIE-ML and AIE-MLv2 devices.

See the operator and feature support matrix for supported precisions, frontend coverage, tensor and transport contracts, and current limitations.

Prerequisites

• AMD Vitis 2025.2 and a valid AIE tools license.
• Python 3.10+.
• Optional: hls4ml if using the hls4ml frontend integration.

Frontend Compatibility

The ONNX path is the recommended route for operator-level compiler development and for models that already express quantized tensors and Q/DQ boundaries explicitly. The hls4ml path is intended for MLP-style pipelines at the moment.

Installation

pip install aie4ml

Install hls4ml only if you need the hls4ml frontend/backend integration:

pip install hls4ml

Documentation & Tutorials

Documentation and usage: https://github.com/dimdano/aie4ml

Tutorial 1: tutorials/tutorial_1.ipynb Tutorial 2: tutorials/tutorial_2.ipynb

General hls4ml concepts: https://fastmachinelearning.org/hls4ml

Maintainer

aie4ml is developed and maintained by Dimitrios Danopoulos.

Citation

If aie4ml contributes to your research, please cite the corresponding arXiv preprint:

@misc{danopoulos2025aie4mlendtoendframeworkcompiling,
      title={AIE4ML: An End-to-End Framework for Compiling Neural Networks for the Next Generation of AMD AI Engines},
      author={Dimitrios Danopoulos and Enrico Lupi and Chang Sun and Sebastian Dittmeier and Michael Kagan and Vladimir Loncar and Maurizio Pierini},
      year={2025},
      eprint={2512.15946},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2512.15946},
}