Minimalistic Vector Extension for Image Detection on Edge Nodes Using ‘VGG16’

As Technological advancements are crossing the boundaries like never before, postulating the need for smarter algorithms such as NN’s and faster hardware to process them becomes evident in this era. The hardware available in the open market are general purpose, limiting the complete usability of hardware for the intended CNN application. This paper addresses the implementation of a custom 32bit vector extension for a complex neural network called VGG16. We have implemented a highly parallelised vector core that takes care of computationally intensive mathematical operations. Using four lanes, each lane with nine vector registers with three vector elements in each register. Vector changing techniques was used to optimise resource utilisation. An improvement of 30% was achieved with respect to throughput compared to existing designs which can process the VGG16 neural network model. As VGG16 is an inefficient model in the industry, the current design projects the improved performance of VGG16 using the vector core. This enables the implementation of image detections on edge IoT applications.