SNN2Bitstream

Automated Model-to-FPGA HLS-Based Framework for Spiking Neural Network Hardware Accelerators

SNN2Bitstream compiles a Spiking Neural Network trained in snnTorch into an FPGA bitstream for the AMD/Xilinx ZCU104, based on HLS tools. It generates readable, HLS-friendly C++, lets you verify it with GCC before synthesis, and supports both Vitis HLS and the open-source Bambu HLS.

Python (tested with Python 3.10):

pip install -r requirements.txt

Hardware tools, required only to reach a bitstream and the board application:

Put them on PATH (source the Xilinx settings64.sh). The 2020.2 release also needs AMD's Y2K22 patch for HLS IP export (AMD Answer Record 76960); without it the IP export step fails. Other versions are untested.

# Software flow only: train → IR → C++ → GCC test
./snn2bitstream sw --toml configs/mnist_fcn_rate.toml

# Full flow to bitstream: SW + HLS + Vivado
./snn2bitstream full --toml configs/mnist_fcn_rate.toml

# Hardware only, on an already-generated project
./snn2bitstream hw <project_name>

./snn2bitstream forwards to script/run_all.sh; either name works.

Datasets download on demand, e.g. python tools/download_data.py mnist.

All options can be given on the command line or in the TOML [codegen] section (CLI wins).

Bambu is the open-source HLS tool of the PandA framework (Politecnico di Milano); see its repository for installation. Bambu tuning goes in the TOML: [codegen.bambu] with opt = "-O3" and extra = "...". The presence of that section also selects the backend, so --backend bambu on its own runs Bambu at its defaults.

Deploy your own pre-trained snnTorch model. Put the model class in a Python file, e.g. user_model/my_model.py, and pass it as a module path together with its weights:

# FCN (input dim auto-detected)
./snn2bitstream sw --custom user_model.my_model.MyFCN \
    --weights user_model/my_fcn.pt --project myfcn --timestep 10 --dataset nmnist

# CSNN (Conv2d models need --input-shape C,H,W)
./snn2bitstream sw --custom user_model.my_model.MyCSNN \
    --weights user_model/my_csnn.pt --project mycsnn --timestep 10 \
    --dataset nmnist --input-shape 2,34,34 --config SPQ

Support for custom models is currently limited.

./snn2bitstream full --toml configs/mnist_fcn_rate.toml --mul-impl-fabric

Target board is the AMD ZCU104 (XCZU7EV Zynq UltraScale+ MPSoC) at 100 MHz.

Generated output (all gitignored):

full and hw also build the application ELF. To measure accuracy on the FPGA, copy the test set to an SD card:

python tools/export_mnist_bin.py test_data/mnist --encoding rate --timesteps 10   # MNIST only
python tools/prep_sd_data.py <project_name> --sd <SD_ROOT>

Then open the Vitis workspace backend_projects/<project_name>/xilinx/vitis_<project_name>/ (bambu/ instead of xilinx/ for Bambu), run the ELF on the board over JTAG.

The article introducing SNN2Bitstream has been accepted for presentation at the IEEE International Conference on Artificial Intelligence Circuits and Systems (AICAS) 2026.

To cite our work, please use the following citation:

X. Zhang, I. Kaskampas, A. Pinna and H.-G. Stratigopoulos, "SNN2Bitstream: Automated Model-to-FPGA HLS-Based Framework for Spiking Neural Network Hardware Accelerators," IEEE International Conference on Artificial Intelligence Circuits and Systems (AICAS), 2026.

@inproceedings{zhang2026snn2bitstream,
      title={SNN2Bitstream: Automated Model-to-FPGA HLS-Based Framework for Spiking Neural Network Hardware Accelerators},
      author={Xindan Zhang and Ioannis Kaskampas and Andrea Pinna and Haralampos-G. Stratigopoulos},
      booktitle={IEEE International Conference on Artificial Intelligence Circuits and Systems (AICAS)},
      year={2026},
}

This work has been funded by the French National Research Agency (ANR) France 2030 AdaptING project (Grant No ANR-23-PEIA-0009).

Xindan Zhang is the Author of the Software
Copyright 2026 Sorbonne Université, Centre National de la Recherche Scientifique

SNN2Bitstream is free software: you can redistribute it and/or modify it under the terms of GNU General Public License version 3 as published by the Free Software Foundation.

SNN2Bitstream is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.

You will find in the LICENSE file a copy of the GNU General Public License version 3.