Optimized Deep Neural Networks for Real-Time Object Classification on Embedded GPUs

Convolution is the most computationally intensive task of the Convolutional Neural Network (CNN). It requires a lot of memory storage and computational power. There are different approaches to compute the solution of convolution and reduce its computational complexity. In this paper, a matrix multiplication-based convolution (ConvMM) approach is fully parallelized using concurrent resources of GPU (Graphics Processing Unit) and optimized, considerably improving the performance of the image classifiers and making them applicable to real-time embedded applications. The flow of this CUDA (Compute Unified Device Architecture)-based scheme is optimized using unified memory and hardware-dependent acceleration of matrix multiplication. Proposed flow is evaluated on two different embedded platforms: first on an Nvidia Jetson TX1 embedded board and then on a Tegra K1 GPU of an Nvidia Shield K1 Tablet. The performance of this optimized and accelerated convolutional layer is compared with its sequential and heterogeneous versions. Results show that the proposed scheme significantly improves the overall results including energy efficiency, storage requirement and inference performance. In particular, the proposed scheme on embedded GPUs is hundreds of times faster than the sequential version and delivers tens of times higher performance than the heterogeneous approach