EdgeMap: An Optimized Mapping Toolchain for Spiking Neural Network in Edge Computing.

Area of Science:

• Artificial Intelligence
• Neuromorphic Computing
• Edge Computing

Background:

• Spiking Neural Networks (SNNs) offer intelligence and energy efficiency for edge computing.
• Current SNN mapping methods on neuromorphic hardware suffer from high latency, low throughput, and poor energy/connectivity management.
• These limitations hinder the practical deployment of SNNs in edge computing.

Purpose of the Study:

• To introduce EdgeMap, an optimized toolchain for deploying SNNs on edge devices.
• To overcome the performance and efficiency limitations of existing SNN mapping schemes.
• To enable high-performance, energy-efficient SNN applications in edge computing scenarios.

Main Methods:

• SNN graph partitioning using a streaming graph partition algorithm, clustering neurons based on hardware constraints.
• Multi-objective optimization algorithm for minimizing energy and communication costs during mapping.
• Evaluation of EdgeMap across four diverse SNN applications.

Main Results:

• EdgeMap reduced average latency by up to 19.8%, energy consumption by 57%, and communication costs by 58%.
• Execution time was improved by a factor of 1225.44×.
• Throughput was increased by up to 4.02× compared to state-of-the-art methods.

Conclusions:

• EdgeMap provides a highly efficient and effective solution for deploying SNNs on edge devices.
• The toolchain significantly enhances performance metrics crucial for edge computing.
• EdgeMap demonstrates strong utility for real-world SNN applications in resource-constrained edge environments.