Efficient FPGA Acceleration of Convolutional Deep Neural Networks

Department of Computer EngineeringDeep Convolutional Neural Networks (CNNs) are a powerful model for visual recognition tasks, but due to their very high computational requirement, acceleration is highly desired. FPGA accelerators for CNNs are typically built around one large MAC (multiply-accumulate) array, which is repeatedly used to perform the computation of all convolution layers, which can be quite diverse and complex. Thus a key challenge is how to design a common architecture that can perform well for all convolutional layers. In this paper we present a highly optimized and cost-effective 3D neuron array architecture that is a natural FFt for convolutional layers, along with a parameter selection framework to optimize its parameters for any given CNN model. We show through theoretical as well as empirical analyses that structuring compute elements in a 3D rather than a 2D topology can lead to higher performance through an improved utilization of key FPGA resources. Our experimental results targeting a Virtex-7 FPGA demonstrate that our proposed technique can generate CNN accelerators that can outperform the state-of-the-art solution, by 1.80x to maximum 4.05x for 32-bit ??floating-point, and 16-bit fixed-point MAC implementation respectively for different CNN models. Additionally, our proposed technique can generate designs that are far more scalable in terms of compute resources. We also report on the energy consumption of our accelerator in comparison with a GPGPU implementation.ope