A real-time hybrid neuron network for highly parallel cognitive systems.

Area of Science:

• Computational Neuroscience
• Hardware Acceleration
• Biophysics

Background:

• Understanding neural communication requires advanced simulation tools capable of real-time performance.
• Existing methods face limitations in accurately mimicking neuron and neuron network behavior under real-time constraints.

Purpose of the Study:

• To propose a customizable, highly pipelined neuron network design for Field-Programmable Gate Arrays (FPGAs).
• To achieve optimal execution of floating-point operations for simulating a maximal number of biophysically plausible neurons.
• To enhance computational efficiency by using a single exponent instance for multiple neuron calculations, reducing resource requirements without impacting latency.

Main Methods:

• Developed a highly pipelined neuron network architecture optimized for FPGA implementation.
• Implemented optimally scheduled floating-point operations for neuron simulations.
• Utilized a single exponent instance to conserve resources across multiple neuron calculations.

Main Results:

• The proposed design allows simulation of up to 1188 neurons on a Virtex7 (XC7VX550T) FPGA.
• Achieved brain real-time simulation capabilities.
• Demonstrated a speed-up of x12.4 compared to state-of-the-art methods.

Conclusions:

• The novel FPGA design significantly advances the simulation of neuron networks in real-time.
• This approach offers a scalable and efficient solution for computational neuroscience research.
• The design provides a substantial improvement in simulation speed and neuron density over existing technologies.