Crossbar Parasitics Simulator

A comprehensive tool for modeling and analyzing parasitic effects in crossbar arrays. This repository provides two main functionalities:

1. Model Comparison – Benchmarks a variety of parasitic resistance models from the literature.
2. Hardware-Aware Training – Enables training and evaluation of neural networks using the models.

Table of Contents

1. Features
   1.1 Model Comparison
   1.2 Hardware-Aware Training
2. Installation
3. Usage
   3.1 Crossbar Simulation
   3.2 Parasitic-Aware Neural Network
4. Suggestions for Improvement
5. Citation

---

Features

Model Comparison

• Supported Models
  - Ideal, Jeong, DMR, αβ-matrix, CrossSim, Memtorch, ngSpice, LTSpice.
  - Easily switch the reference model (CrossSim by default).

• Parametric Analysis
  - Adjustable array sizes.
  - Configurable parasitic resistance values and memory window ratios (HRS/LRS).
  - Tunable sparsity levels (% of HRS devices).
  - Device variability for realistic crossbar simulation.

• Performance Metrics
  - Execution time benchmarking.
  - Current/voltage error measurements (relative error).
  - Robustness evaluations under varying crossbar parameters.

Hardware-Aware Training

• Memristive Neural Network Setup
  - Two-layer fully connected architecture for MNIST classification, with configurable hidden layer.
  - Flexible tile partitioning, linear mapping of weights to conductances and quantization.

• Training and Testing
  - Parasitic resistance incorporation into training, using parasitic models with automatic differentiation.
  - Weights, biases, and confusion matrix visualizations after training. 2D and 3D accuracy plots for evaluation.

---

Installation

1. Clone the Repository

   git clone https://github.com/alambertini01/Crossbar_Models_Comparison
   cd Crossbar_Models_Comparison

2. Create and Activate a Virtual Environment (Optional)
   Windows:

   python -m venv .venv
   .venv\Scripts\activate

   macOS/Linux:

   python -m venv .venv
   source .venv/bin/activate

3. Install the Package

   pip install -e .

4. (Optional) Installing NgSpice
   If you plan to use the NgSpice-based model:

   pyspice-post-installation --install-ngspice-dll
   pyspice-post-installation --check-install

---

Usage

1. Crossbar Simulation

1.1 Model Performance Comparison

python Crossbar_Simulator.py

• Ideal for in-depth current and voltage analysis on custom-sized arrays.
• Offers detailed parameter tuning (resistance ratios, device variability, etc.).

1.2 Robustness Analysis

python Crossbar_Simulator_Robustness.py

• Runs comprehensive simulation sweeps across multiple parameters.
• Generates robustness metrics (radar plots, 3D MSE plots).

2. Parasitic-Aware Neural Network

2.1 Training

python NN_Training.py

• Trains a two-layer MNIST classifier with PyTorch-based crossbar models (located in CrossbarModels/Crossbar_Models_pytorch.py).
• Command-line prompts allow setting parasitic resistance, HRS, tile size, etc.
• Additional parameters (LRS, quantization bits, training checkpoints) can be modified within the script.
• Trained models are saved in TrainedModels/ModelName.
• For GPU acceleration make sure to install pytorch based on your specific cuda version.

2.2 Testing

python NN_testing.py

• Loads a trained model and evaluates its accuracy.
• Different models can be used as reference, but it's recommended to use the CrossSim model for reliable validation.
• Further analysis on accuracy can be performed by sweeping parameters like R_HRS and bits.

---

Citation

@software{crossbar_simulator,
  author = {Lambertini, Alessandro},
  title = {Crossbar Parasitics Simulator},
  year = {2024},
  publisher = {GitHub},
  url = {https://github.com/alambertini01/Crossbar_Models_Comparison}
}