FPGA-implementation av ett neuralt nätverk

Abstract

Image recognition is a quickly growing field where convolutional neural networks, CNN, are in the bleeding edge. Today fast GPUs are used which consume a lot of power. Field programmable gate arrays, FPGAs, are more energy efficient per calculation. This report describes an architecture of a convolutional neural network implemented in a field programmable gate array. The main purpose is to design an architecture and demonstrate its functionality in regards to power, speed and resource usage. In order to achieve the architecture, the project has followed general guidelines for a convolutional neural network, with filters that extend over the entire depth of the image. The parameters of the design were adapted for the FPGA used in the project. The dimensions of the memory were adjusted to reduce the number of times each data has to be loaded for each calculation, due to max-pooling. The final architecture, however, resulted in a flexible enough design that is adaptable to other FPGAs. When implemented, the calculations used both data and filters from a limited read-only memory, ROM, the design could use data from the main processor. The computing capacity of the architecture is far below the theoretical capacity of the FPGA. However, there are multiple possibilities for improvements which would improve the computing potential dramatically. To utilize the increased potential, the summing tree used in the architecture can be modified which will potentially double the calculations per clock cycle and optimize the critical data path to further increase the clock speed. Despite these limitations, the current architecture has higher performance-to-power ratio than a GTX 1060.