This repository provides a Python library for interfacing with Daffodil, a flexible-interface platform designed for research and development of two-terminal resistive memory and selector devices. The original design and vision for Daffodil was described in this paper by Hoskins et al., though it has evolved considerably since then. Daffodil mainly targets 18kb two-terminal analog memory device arrays that resulted from the Nanotechnology Xccelerator project.
The code provided here should be regarded as an alpha version. It has been demonstrated to work in a restricted experimental setup, first demonstrated in a paper by Yousuf et al. who use it to measure and operate a memristive neural network. The current release is intended to provide an initial virtual interface to research groups interested in using a Daffodil system who wish to start developing the appropriate experimental control codes before receiving a Daffodil board.
The core feature provided by this library is the definition of a pure abstract virtual class Daffodil_Base, as well as two implementations of this class: Daffodil_Phys, and Daffodil_Sim. The Daffodil_Sim class mimics the behavior of an ideal control board. The Daffodil_Phys class operates the physical Daffodil board, which currently exists in a pre-release version.
Operating Daffodil_Phys requires fabricating a printed circuit board according to certain specifications, integrating it as a daughterboard to an FPGA, and creating an interface such as that provided by the daffodil-fpga library. Eventually, we plan to support this full-stack setup, but at the moment this option should be regarded as experimental and pre-release. If you are interested acquiring a physical experimental setup for use with the Daffodil_Base class, or otherwise would like to contact the developers, you can either submit an issue or email NIST's Alternative Computing Group at [email protected].
This library depends on several other Python libraries, which are listed in requirements.txt. The recommended installation method is to create a lightweight virtual environment with the python -m venv env command, and then install the dependencies and daffodil-lib using pip. For instance,
$ python -m venv .env
$ source .env/bin/activate
(.env) $ pip install -r requirements.txt
(.env) $ pip install daffodillibThe library can then be imported like any other Python library. The script examples/infer_wine.py shows a basic example of using the Daffodil_Sim class to emulate the operation of a neural network inference engine based on a memristive device array.
To compile the API documentation, run make html from within the docs directory. The current API documentation will be hosted online soon.
If you use this library, please cite this repository according to the information in CITATION.cff.
Depending on your usage, it may also be appropriate to cite the following papers:
- A system for validating resistive neural network prototypes (doi:10.1145/3477145.3477260), which first proposed the Daffodil platform;
- Layer Ensemble Averaging for Improving Memristor-Based Artificial Neural Network Performance (doi:10.1038/s41467-025-56319-6), the first published work to utilize a fully-integrated Daffodil system
This library was developed by multiple parties as part of a collaboration between the National Institute of Standards and Technology (NIST), The George Washington University (GWU), and Western Digital Research. This library as a whole is licensed under the BSD-3 license (see LICENSE); however, specific contributions and modifications made by employees of NIST are provided as a public service and are not subject to copyright protection within the United States (see LICENSE-NIST).
Research groups interested in collaborating on this project are encouraged to reach out to either the Alternative Computing Group at NIST or to the Adaptive Devices and Microsystems (ADAM) group at GWU.
NIST Contact:
Dr. Matthew W. Daniels
[email protected]
Alternative Computing Group
Nanoscale Device Characterization Division
Physical Measurement Laboratory
GWU Contact:
Prof. Gina Adam
[email protected]
Adaptive Devices and Microsystems Group
Department of Electrical and Computer Engineering