Convolutional neural networks (CNNs) have become the dominant neural networkarchitecture for solving many state-of-the-art (SOA) visual processing tasks.Even though Graphical Processing Units (GPUs) are most often used in trainingand deploying CNNs, their power consumption becomes a problem for real timemobile applications. We propose a flexible and efficient CNN acceleratorarchitecture which can support the implementation of SOA CNNs in low-power andlow-latency application scenarios. This architecture exploits the sparsity ofneuron activations in CNNs to accelerate the computation and reduce memoryrequirements. The flexible architecture allows high utilization of availablecomputing resources across a wide range of convolutional network kernel sizes;and numbers of input and output feature maps. We implemented the proposedarchitecture on an FPGA platform and present results showing how ourimplementation reduces external memory transfers and compute time in fivedifferent CNNs ranging from small ones up to the widely known large VGG16 andVGG19 CNNs. We show how in RTL simulations in a 28nm process with a clockfrequency of 500MHz, the NullHop core is able to reach over 450 GOp/s andefficiency of 368%, maintaining over 98% utilization of the MAC units andachieving a power efficiency of over 3TOp/s/W in a core area of 5.8mm2
展开▼
