Date of Award


Document Type


Degree Name

Master of Science (MS)

Legacy Department

Computer Engineering


Smith, Melissa C

Committee Member

Birchfield , Stanley

Committee Member

Gowdy , John


There has been a strong interest in modeling a mammalian brain in order to study the architectural and functional principles of the brain and offer tools to neuroscientists and medical researchers for related studies. Artificial Neural Networks (ANNs) are compute models that try to simulate the structure and/or the functional behavior of neurons and process information using the connectionist approach to computation. Hence, the ANNs are the viable options for such studies.
Of many classes of ANNs, Spiking Neuron Network models (SNNs) have been employed to simulate mammalian brain, capturing its functionality and inference capabilities. In this class of neuron models, some of the biologically accurate models are the Hodgkin Huxley (HH) model, Morris Lecar (ML) model, Wilson model, and the Izhikevich model. The HH model is the oldest, most biologically accurate and the most compute intensive of the listed models. The Izhikevich model, a more recent development, is sufficiently accurate and involves the least computations. Accurate modeling of the neurons calls for compute intensive models and hence single core processors are not suitable for large scale SNN simulations due to their serial computation and low memory bandwidth. Graphical Processing Units have been used for general purpose computing as they offer raw computing power, with a majority of logic solely dedicated for computing purpose.
The work presented in this thesis implements two-level character recognition networks using the four previously mentioned SNN models in Nvidia's Tesla C870 card and investigates performance improvements over the equivalent software implementation on a 2.66 GHz Intel Core 2 Quad. The work probes some of the important parameters such as the kernel time, memory transfer time and flops offered by the GPU device for the implementations. In this work, we report speed-ups as high as 576x on a single GPU device for the most compute-intensive, highly biologically realistic Hodgkin Huxley model. These results demonstrate the potential of GPUs for large-scale, accurate modeling of the mammalian brain. The research in this thesis also presents several optimization techniques and strategies, and discusses the major bottlenecks that must be avoided in order to achieve maximum performance benefits for applications involving complex computations. The research also investigates an initial multi-GPU implementation to study the problem partitioning for simulating biological-scale neuron networks on a cluster of GPU devices.