NeuroFlow: A General Purpose Spiking Neural Network Simulation Platform using Customizable Processors.

Area of Science:

• Computational Neuroscience
• High-Performance Computing
• Neuromorphic Engineering

Background:

• Spiking neural network (SNN) simulations are crucial for understanding brain function but computationally demanding.
• Existing hardware platforms like multi-core processors and ASICs have limitations in flexibility and performance for SNNs.
• Customizable hardware offers potential for optimized SNN simulation performance.

Purpose of the Study:

• To introduce NeuroFlow, a novel, scalable SNN simulation platform.
• To demonstrate NeuroFlow's reconfigurable architecture for optimized performance.
• To evaluate NeuroFlow's simulation capabilities and speedup over conventional hardware.

Main Methods:

• NeuroFlow utilizes Field-Programmable Gate Arrays (FPGAs) for its customizable processor architecture.
• The platform supports PyNN for configuring the processor and includes common neuronal models (e.g., integrate-and-fire, Izhikevich) and STDP learning.
• Performance was evaluated on a 6-FPGA system simulating large neural networks.

Main Results:

• A 6-FPGA system can simulate up to ~600,000 neurons, achieving real-time performance for 400,000 neurons.
• NeuroFlow on a single FPGA offers up to 33.6x speedup over an 8-core processor and 2.83x over GPU platforms.
• The reconfigurable architecture allows for tailored performance optimization.

Conclusions:

• NeuroFlow provides a highly flexible and high-throughput platform for large-scale SNN simulations.
• Its FPGA-based architecture offers significant performance advantages over traditional computing systems.
• NeuroFlow represents a viable environment for advancing computational neuroscience research through efficient simulation.