- Python 96.9%
- C 2%
- C++ 0.8%
- Cuda 0.2%
| Filename | Latest commit message | Latest commit date |
|---|---|---|
| code | ||
| data | ||
| LICENSE | ||
| README.md | ||
Readme.md
Overview
This repository contains the source code and data for our research on a highly energy-efficient multi-core neuromorphic architecture designed for training deep spiking neural networks. The project is divided into two main directories: code and data.
Code Structure
The code directory is organized into the following subdirectories, grouped by category:
Hardware Implementation Code
-
dc_synthesized_netlist: Gate-level netlists of Multi-Core Neuromorphic Architecture synthesized by Design Compiler. -
fpga_implementation_netlist: FPGA implementation netlist of the Multi-Core Neuromorphic Architecture for training deep spiking neural networks.
Software Toolchain
-
software_toolchain: SNN Many-Core Architecture Software Toolchain for building and deploying Spiking Neural Networks (SNNs). Supports compilation and deployment of trained SNN models to FPGA-based many-core chips, providing a complete SNN model building, training, optimization, and hardware deployment pipeline. Additionally, it supports statistical analysis of DRAM usage for SNN-ResNet18, SNN-ResNet50, and SNN-VGG16 models based on either our architecture or GPU. -
nir_converter: Neuromorphic Intermediate Representation (NIR) Conversion Framework. Used for processing neuromorphic computing network models, including model loading, serialization, IR/NIR format conversion, and other functions. Provides a complete framework for converting neural network models between proprietary IR (Intermediate Representation) and NIR (Neuromorphic Intermediate Representation) formats.
Analysis and Evaluation Tools
-
hardware_dse_simulator: S_ZigZag SNN Training Simulator, a hardware architecture mapping design space exploration (DSE) framework for Spiking Neural Network (SNN) accelerators. Bridges the gap between SNN training algorithms and their training costs on SNN-specific training accelerators through rapid and accurate hardware cost estimation. The framework can map SNN training algorithms onto hardware resources and memory, evaluating various indicators of the SNN accelerator hardware architecture. Additionally, it supports statistical analysis of FPS for SNN-ResNet18, SNN-ResNet50, and SNN-VGG16 models based on our architecture. -
gpu_performance_profiler: GPU-based SNN Model Performance Analyzer. Contains source code for analyzing the DRAM access and FPS performance of Spiking Neural Network (SNN) models including SNN-ResNet18, SNN-ResNet50, and SNN-VGG16 when training on GPU. -
memory_access_analyzer: Neuromorphic Architecture DRAM Access Analyzer. Contains source code for analyzing the DRAM access of Spiking Neural Network (SNN) models including SNN-ResNet18, SNN-ResNet50, and SNN-VGG16 when training on our neuromorphic architecture.
Usage
To use the source code, please refer to the documentation within each subdirectory for specific instructions and dependencies. The data files can be opened with any standard spreadsheet software and are intended to provide insights into the performance and efficiency of our neuromorphic architecture.
Data Structure
This directory contains the raw data associated with our research paper. The data is organized into two main folders:
maintext
This folder contains the source data for figures presented in the main text of the paper. The data is organized by figure numbers:
- figure3_d-f/ - Raw data for Figure 3(d-f)
- figure3_g-o.xlsx - Source data for Figure 3(g-o)
- figure3_p-q.xlsx - Source data for Figure 3(p-q)
- figure4_e-f.xlsx - Source data for Figure 4(e-f)
- figure5_c-d.xlsx - Source data for Figure 5(c-d)
- figure5_g-i.xlsx - Source data for Figure 5(g-i)
- figure5_l-o.xlsx - Source data for Figure 5(l-o)
supplementary
This folder contains the source data for figures presented in the supplementary materials:
- figure6_a-b.xlsx - Source data for Figure 6(a-b) in the supplementary materials
File Format
Most data files are provided in Microsoft Excel format (.xlsx) which can be opened with common spreadsheet applications like Microsoft Excel, LibreOffice Calc, or Google Sheets.
Usage
These data files can be used to reproduce the figures in the paper or to perform further analysis. Each file contains the raw numerical data that was used to generate the corresponding figure in the publication.