General Purpose GPU Benchmarks for Neural Networks

Neural networks are a field of computing experiencing a rise in popularity in recent years due to the utilization of graphics processing units as their computational centerpiece. The lack of neural network benchmarks for the open-source Nyuzi architecture, a developing general-purpose processor with graphical processing capabilities, is the focus of this thesis. This work aims to determine whether Nyuziâ??s performance counters and traceable events suffice for performance tuning of neural network implementations. Given the mathematical intensity of neural networks, a strong emphasis is placed on events related to arithmetic instructions. Experimenting with neural network implementations in C and C++, existent functionality resulted sparse. Modification of performance counters allowed the observation that object-oriented implementations of a small neural network yield an order of magnitude improvement in execution time of arithmetic instructions when compared to a procedural implementation. Results suggest a need to expand Nyuziâ??s number of performance counters and diversity of traceable events to boost the development of efficient neural network applications.