Architectures for high-performance FPGA implementations of neural models

As the complexity of neural models continues to increase (larger populations, varied ionic conductances, more detailed morphologies, etc) traditional software-based models have difficulty scaling to reach the performance levels desired. This paper describes the use of FPGAs, or field programmable gate arrays, to easily implement a wide variety of neural models with the performance of custom analogue circuits or computer clusters, the reconfigurability of software, and at a cost rivalling personal computers. FPGAs reach this level of performance by enabling the design of neural models as parallel processed data paths. These architectures provide for a wide range of single-compartment, multi-compartment and population models to be readily converted to FPGA implementations. Generalized architectures are described for the efficient modelling of a first-order, nonlinear differential equation in throughput maximizing or latency minimizing data-path configurations. The homogeneity of population and multicompartment models is exploited to form deep pipelines for improved performance. Limitations of FPGA architectures and future research areas are explored.