Design of a multi-FPGA system for biologically plausible neural networks based of heterogeneous computing

Abstract

Today neuroscience is vastly specialized such that computational neuroscience tries to bridge the gaps of knowledge between the theory and the experiments. In-silico experiments are computer simulations with complete control over the scenario; this techniques try to decode the functionality of the biological neural networks and the biophysical dynamics which this cells inherent. This work explores a way to improve biological-precise spiking neural networks simulations with FPGA acceleration. Our approach focuses with creating a hardware acceleration for one cell compartment using a system-on-chip, this serves a proof-of-concept to value how flexible is the platform to accelerate similar simulations using the hybrid hardware-software methods. The work described in this thesis is a implementation of the inferior olivary nuclei model implemented with a extended Hodgkin-Huxley neural model. The development platform was the Xilinx’s Zynq-7000 and the Vivado Hardware Design suite. Results obtained in this work shows that the hybrid computing is more performance efficient in using the FPGA resources. Also proves a more flexible platform unlike other authors similar work. Finally, the use of a shared DRAM between the CPU and FPGA fabric showed a bottleneck for the design, its noted that it would be preferable to separate if possible the main DRAM between both systems.