Benchmarking Highly Parallel Hardware for Spiking Neural Networks in Robotics

Area of Science:

• Robotics and Artificial Intelligence
• Computational Neuroscience
• Neuromorphic Engineering

Background:

• Animal brains surpass machine performance in speed and efficiency.
• Significant advancements in robotic vision, motion, and path planning have been achieved.
• Spiking Neural Networks (SNNs) offer brain-inspired computational advantages for robotics.

Purpose of the Study:

• To benchmark the performance of SNNs on different hardware platforms relevant to robotics.
• To compare simulation time, energy consumption, and pathfinding accuracy for robotic use cases.
• To evaluate the suitability of neuromorphic hardware and GPUs for SNN simulations in robotics.

Main Methods:

• A neural Wavefront algorithm, representative of robotic path planning, was simulated using SNNs.
• The SNN model was developed in the simulator-independent language PyNN.
• Performance was evaluated across different backends: Nest (CPU), SpiNNaker (neuromorphic hardware), and GeNN (GPU).
• GeNN's performance was further analyzed across different hardware deployments.

Main Results:

• Comparative analysis of total simulation time, average energy consumption, and path length was conducted.
• Performance differences between serial CPU, neuromorphic hardware, and GPU implementations were identified.
• Variations in GeNN performance across different GPU hardware were investigated.

Conclusions:

• Insights into the performance of parallel hardware solutions for SNNs in robotics were gained.
• The study aims to guide the development of more efficient SNN implementations for robotic applications.
• Understanding hardware performance is crucial for optimizing SNNs in real-world robotic systems.