Crossbar-Based Neural Network Inference Simulation

This project provides tools to simulate and evaluate neural network (NN) inference using crossbar arrays. The simulations focus on key performance metrics such as NN accuracy, chip-level power consumption, hardware inference time, and hardware area.

Quick Start Guide

Step 1: Neural Network Accuracy Simulation

Define your model architecture and dataset in ./models

Configure simulation parameters in inferenceConfig.py, note that this config will only effect inference!

Run the simulation script:

python inference_sim.py

It will generate NN traces (each layer's output), saved in "./layer_record_[NN]/", these traces will be used in NeuroPySim core program, to calculate power, latency and area.

Step 2: Chip-Level Simulation

build:

cd from_neurosim
mkdir build
cd build
cmake ..
make

This creates the FormulaBindings.so shared object file, which allows Python scripts to access NeuroSIM parameters and formulas.

Run chip sim

python Chip.py

This script calculates:

1. Chip Area
2. Inference Time
3. Energy Consumption
4. Performance Breakdown for crossbar arrays

NOTE:

The project defaults to RRAM, 1T1R, and conventional sequential crossbar configurations.

You can extend support for other memory types by modifying related files (see "Customization").

Customization Guide

Read before using your config:

1. ./from_neurosim/param.cpp: param defined by NeuroSIM.
2. ./from_neurosim/bindings.cpp: bind the NeuroSIM's formula.cpp to python.

Crossbar

Energy

1. Check Memory cell read energy in Array.py.
2. Check Memory cell write energy in Tile.py.

Latency

1. Check Memory cell read latency in Array.py.
2. Check Memory cell write latency in Tile.py.

Area

1. Check Array.py.

Peripheral Circuits

1. Check each class.

References

NeuroSIM v1.4 (inference)

NeuroSIM v2.0 (train)

Requirements

Python > 3.10
Pybind11
torch
numpy

Appendix

Data relations on simulation

1. Acc.: Related.
2. Read Power: Related, activity rows of input vectors only.
3. Write Energy: Related, Weight matrix before and after update.
4. Latency: Related, numbit of input only.
5. Area: Not related.