An FPGA-Based Massively Parallel Neuromorphic Cortex Simulator.

Area of Science:

• Neuroscience
• Computer Engineering
• Computational Neuroscience

Background:

• Simulating large-scale spiking neural networks requires significant computational resources.
• Existing methods face memory limitations for fully connected networks.
• Neuromorphic architectures offer potential for efficient neural network simulation.

Purpose of the Study:

• To develop a massively parallel and scalable neuromorphic cortex simulator.
• To enable real-time simulation of large, structurally connected spiking neural networks.
• To abstract neuromorphic architecture using neurobiology-inspired units (minicolumns and hypercolumns).

Main Methods:

• Novel architecture abstracting neuromorphic hardware into minicolumns and hypercolumns.
• On-chip memory storage for parameters and connections, avoiding large look-up tables.
• Hierarchical communication scheme for high neuron fan-out (up to 200k).
• Implementation on an Altera Stratix V FPGA for proof-of-concept.

Main Results:

• Real-time simulation of 20 million to 2.6 billion leaky-integrate-and-fire (LIF) neurons.
• Successful emulation of a simplified auditory cortex with 100 million neurons.
• Achieved low power dissipation of 1.62 μW per neuron.
• Demonstrated system reconfigurability for different neural network simulations.

Conclusions:

• The developed cortex simulator is a scalable and accessible tool for large-scale spiking neural network research.
• The novel architectural abstraction overcomes memory limitations in simulating complex neural networks.
• The system enables efficient, real-time simulation and analysis of brain-inspired computational models.
• Offers a low-power solution for advancing neuromorphic computing and neuroscience research.