Scalable Multi-Hierarchy Embedded Platform for Neural Population Simulations.

Area of Science:

• Computational Neuroscience
• Embedded Systems Engineering
• Brain-Inspired Computing

Background:

• Simulating large-scale, high-dimensional neural populations with complex nonlinearities presents significant computational challenges.
• Brain-inspired structured neural circuits are fundamental to both computational and perceived intelligence.

Purpose of the Study:

• To propose and validate an ARM-based scalable multi-hierarchy parallel computing platform (EmPaas) for real-time neural population simulations.
• To address the computational demands of simulating complex neural networks.

Main Methods:

• Developed EmPaas using 340 ARM Cortex-M4 microprocessors with a tree-two-dimensional grid-like hybrid topology.
• Deployed an optimized model of a biologically plausible basal ganglia-thalamus (BG-TH) network onto the EmPaas platform.
• Utilized distributed computations and embedded multi-cores for parallel processing.

Main Results:

• Simulated a 4000-neuron BG-TH network for 3000 ms at 168 MHz, achieving an actual computation time of 2748.57 ms.
• Demonstrated high-speed and high-accuracy parallel computing with low power consumption (56.565 mW per core).
• EmPaas supports real-time performance with up to 2000 Izhikevich neurons per Extended Community Unit (ECUnit).

Conclusions:

• The EmPaas platform significantly improves computational efficiency for large-scale neural network simulations.
• EmPaas offers a practical and efficient method for research in brain-inspired computing and large-scale brain network simulation.
• The parallel computing approach meets real-time performance requirements for complex neural models.