Methodology and Design Flow for Assisted Neural-Model Implementations in FPGAs

摘要

Field programmable gate arrays (FPGAs) have previously been shown as high-performance platforms for neural-modeling applications. Implementations have traditionally been time-consuming and error-prone, requiring the neural modeler to work outside of their expert domain. This paper demonstrates a new approach to the development of neural models using an auto-generation toolkit. This design tool enables model construction-level alterations (e.g., adjustment of model population size or insertion/deletion of an ionic conductance) within hours and parameter changes on-the-fly. The approach is validated on a 40-neuron pre-Boumltzinger complex population model consisting of Hodgkin-Huxley style conductances and fully interconnected synapses. A total of 1880 parameters are on-the-fly user tunable on a free-running model. The resulting implemented model performs at a theoretical 8.7times real-time utilizing 90% of logic elements within a Xilinx Virtex-4 XC4VSX35-fg676-10FPGA