CUTIE, the Completely Unrolled Ternary Inference Engine, is a ternary neural network accelerator targetting high-efficiency and low-power applications. This repository contains the hardware model of CUTIE, as well as the infrastructure to test it by deploying networks to CUTIE.
This repository is structured as follows:
rtlcontains the RTL code and Verilog testbench for CUTIEconfcontains the implementation configuration of the acceleratorstimulicontains the python-based stimuli generators for CUTIEsimcontains the files required to start the simulation
This repo uses Bender (https://github.com/pulp-platform/bender) to manage its dependencies and generate compilation scripts. For this reason, the build process of this project will download a current version of the Bender binary.
The included simulation flow is based on a file-based testbench and a golden model implementation in PyTorch / python. To install the required python dependencies, run
pip install -r requirements.txtGenerating stimuli was tested with python 3.7, other versions may or may not work.
Currently, Modelsim is the only supported simulation platform.
To compile and simulate the project after installing all dependencies, you may run
make gen simwhich will download bender if not done already, fetch the RTL dependencies, generate random test stimuli and start ModelSim.
CUTIE is designed to be parametrizable in many of its fundamental aspects. If you would
like to parametrize CUTIE differently than the given example, you can edit conf/cutie_conf.sv and conf/cutie_config.py
which are the packaged parameter for the RTL implementation and stimuli generation, respectively.
Please be aware that not all combinations of parameters have been tested or are anticipated to work without changes to RTL/Stimuli generators. The main parameters that have been ensured to be modifiable are N_I and N_O, the number of input, and output channels.
CUTIE is released under permissive open source licenses. CUTIE's source code is released under the Solderpad v0.51 (SHL-0.51) license see LICENSE. The code in stimuli is released under the Apache License 2.0 (Apache-2.0) see stimuli/LICENSE.
If you find CUTIE useful in your research, you can cite us:
@InProceedings{CUTIE2022,
author={Scherer, Moritz and Rutishauser, Georg and Cavigelli, Lukas and Benini, Luca},
journal={IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems},
title={CUTIE: Beyond PetaOp/s/W Ternary DNN Inference Acceleration With Better-Than-Binary Energy Efficiency},
year={2022},
volume={41},
number={4},
pages={1020-1033},
doi={10.1109/TCAD.2021.3075420}
}
This paper is also available at arXiv, at the following link: arXiv:2011.01713 [cs.AR].