A quantitative analysis of the performance of computing architectures used in neural simulations

Numerical solutions of neuron models are helping neuroscientists gain new insights into the behavior of neural systems. Although computing power is increasing, the complexity of the systems being simulated is also increasing. If the computation is not well matched to the computing hardware, simulations can take lengthy times to run, which can make it more difficult to draw inferences from those simulations and also to use them in feedback with living neurons such as in the dynamic clamp.

In this paper, we perform a quantitative analysis to get a better sense of how much impact the hardware architectures can have on simulation performance. Three different architectures are implemented on the same hardware platform and compared with respect to simulation time, error, and resources used.

The results indicate that a lookup table approach to evaluate functions can decrease simulation time by orders of magnitude with respect to the traditional approach of mathematical operations.

There are many different ways to implement a lookup table approach to evaluate a function. The method presented in this paper sacrifices some speed for greater generality and accuracy with respect to other published methods.

Lookup tables with 32 interpolation points can dramatically speed up computation time of neural simulations without adding significant error. In this paper linear interpolation was used, but higher order interpolation could be used to further reduce simulation time.