CrossSim is a GPU-accelerated, Python-based crossbar simulator designed to model analog in-memory computing for neural networks and linear algebra applications. It is an accuracy simulator and co-design tool that was developed to address how analog hardware effects in resistive crossbars impact the quality of the algorithm solution. CrossSim provides a special interface to model analog accelerators for neural network inference and training, and also has an API that allows different algorithms to be built on resistive memory array building blocks.

CrossSim can model device and circuit non-idealities such as arbitrary programming errors, conductance drift, cycle-to-cycle read noise, and precision loss in analog-to-digital conversion (ADC). It also uses a fast, internal circuit simulator to model the effect of parasitic metal resistances on accuracy. For neural network inference, it can simulate accelerators with significant parameterizability at the system architecture level and can be used to explore how design choices such as weight bit slicing, negative number representation scheme, ADC ranges, and array size affect the sensitivity to these analog errors. CrossSim can be accelerated on CUDA GPUs, and inference simulations have been run on large-scale deep neural networks such as ResNet50 on ImageNet.

For neural network training accelerators, CrossSim can generate lookup tables of device behavior from experimental data. These lookup tables can realistically simulate the accuracy impact of arbitrarily complex conductance update characteristics, including write nonlinearity, write asymmetry, write stochasticity, and device-to-device variability.

CrossSim does not explicitly model the energy, area, or speed of analog accelerators. A full description of CrossSim's parameters and modeling assumptions can be found in the manual.

CrossSim has been tested on Ubuntu 18.04, and Windows 10 using Python 3.7.6.

CrossSim requires the following Python packages:

• TensorFlow 2.4.1 (includes Keras 2.4.0)
• Numpy 1.20.3
• SciPy 1.7.1
• Pandas 1.3.3
• MatPlotLib 3.4.3
• CuPy 8.3.0 with CUDA 10.2, or Cupy 10.3.1 with CUDA 11.2 (if GPU acceleration is enabled)

Several of the neural network models provided with CrossSim may require additional packages. ImageNet models may require:

• OpenCV 4.5.4
• Torchvision 0.11.1

Inference simulation of Larq quantized models requires:

• Larq 0.12.1

CrossSim has been tested with the version numbers above, but other versions may work as well.

CrossSim uses git submodules to distribute neural network datasets and device lookup tables for training, and neural network models. After cloning this repository, the following commands will fetch the submodules.

git submodule init
git submodule update --progress

This command downloads about 1.2GB of data for the two repositories (combined), which is then de-compressed. After cloning the submodules, you can test CrossSim Inference by running the following commands:

cd inference
python run_inference.py

This will run a CrossSim Inference simulation using a simple CNN and the MNIST dataset by default. The use of the run_inference.py script is explained in more detail in Chapter 3 of the manual.

Due to file size and copyright, this distribution does not include the ImageNet dataset. Section 4.3 of the manual explains how to add your own copy of the ImageNet test set to CrossSim.

CrossSim is primarily focused on neural network inference. Example scripts to perform inference are in the inference directory and described in Chapter 3 of the manual. If you are interested in benchmarking large neural networks, a CUDA-capable GPU is recommended.

If you would like to use the CrossSim array models for other analog MVM applications, the hardware models are located in the cross_sim/cross_sim/xbar_simulator directory of the repository.

Methods for applying device-specific models for programming errors, cycle-to-cycle read noise, and conductance drift are found in the corresponding files located inside the following directory:

/cross_sim/cross_sim/xbar_simulator/parameters/custom_device/

To implement a new device model for use in inference simulations, a new device model can be added as an option in these files. Please see Chapters 7.2-7.4 of the manual for more details.

To use your own pretrained models and datasets for use with CrossSim Inference, or your own device experimental data for use with CrossSim Training, modify the url field in the .gitmodules folder. If you have no yet initialized the submodules, use the commands above. To change the submodule url after the submodules have been initialized, use the following commands:

git submodule sync
git submodule update --init --remote

Instructions for adding new neural network models and devices into CrossSim for use during inference can be found in Chapters 5 and 7 of manual respectively.

If you would like to contribute your device data or models, create a pull request against this repository and the data or pre-trained models repositories.

If you use CrossSim for research please cite:

@article{CrossSim,
  author={T. Patrick Xiao and Christopher H. Bennett and Ben Feinberg and Matthew J. Marinella and Sapan Agarwal},
  title={CrossSim: accuracy simulation of analog in-memory computing},
  url = {https://github.com/sandialabs/cross-sim},
}

For questions, feature requests, bug reports, or suggestions, please submit a new issue through GitHub. The team will get back to you as soon as we can.