KANELÉ: Kolmogorov–Arnold Networks for Efficient LUT-based Evaluation

Published in 34th ACM/SIGDA International Symposium on Field-Programmable Gate Arrays (FPGA' 2026) - Best Paper Nominee

Authors: Duc Hoang*, Aarush Gupta*, Philip Harris. MIT. *Equal Contributions.

Usage Guide

1. (Optional) Activate the Environment

We provided the environment in the environment.yml file, see Conda section for setting up. To activate the environment:

conda activate kanele

2. Train the model on a benchmark

cd benchmarks/<bench-mark-name>

There will be different versions of the training script in each directory, namely train_float.py, train_quant.py etc. To reproduce the results reported in the paper, you can run

python train_quant.py

Where benchmark names could be JSC_CERNBox, or MNIST. This will generate the trained model saved in benchmarks/<bench-mark-name>/models

3. Convert the trained model to FPGA hardware

In the same benchmark directory:

python convert.py

Each of the convert file will have a model_tag parameter, if you do train a new model you have to modify it. Otherwise we provide the trained models as a baseline.

4. Xsim

To run xsim with the generated RTL

cd models/<model-tag>/firmware/sim

To check for correctness:

vivado -mode batch -source sim.tcl

To measure latency:

vivado -mode batch -source sim_latency.tcl

5. Building bitfile or place and route in OOC (out-of-context) mode:

To build the bit-file, or compile the project out of context

cd models/<model-tag>/firmware/

vivado -mode batch -source vivado/build_full.tcl

or

vivado -mode batch -source vivado/build_ooc.tcl

Related FPGA benchmarks information

• Xilinx Virtex UltraScale+ FPGA: xcvu9p-flgb2104-2-i
• Zynq UltraScale+ MPSoC: xczu7ev-ffvc1156-2-e
• Vivado version: 2024.1

Conda

Create conda environment:

conda env create -f environment.yml

Activate the environment:

conda activate kanele

And then do whatever you want in this environment (edit files, open notebooks, etc.). To deactivate the environment:

conda deactivate

If you make any update for the environment, please edit the environment.yml file and run:

conda env update --file environment.yml  --prune

Reference on conda environment here: https://docs.conda.io/projects/conda/en/latest/user-guide/tasks/manage-environments.html Pytorch Note: Some people have encountered issue with Pytorch package installation within the environment, it's typically solved by installing Pytorch with the appropriate environment here: https://pytorch.org/

Citation

If you find the work useful to your research, you can cite us here:

@inproceedings{Hoang:2025sdb,
    author = "Hoang, Duc and Gupta, Aarush and Harris, Philip",
    title = "{KANEL{\'E}: Kolmogorov-Arnold Networks for Efficient LUT-based Evaluation}",
    eprint = "2512.12850",
    archivePrefix = "arXiv",
    primaryClass = "cs.AR",
    doi = "10.1145/3748173.3779202",
    month = "12",
    year = "2025"
}