Balancing Convolutional Neural Networks Pipeline in FPGAs

Convolutional Neural Networks (CNNs) have achieved excellent performance in image classification, being successfully applied in a wide range of domains. However, their processing power demand offers a challenge to their implementation in embedded real-time applications. To tackle this problem, we focused in this work on the FPGA acceleration of the convolutional layers, since they account for about 90% of the overall computational load. We implemented buffers to reduce the storage of feature maps and consequently, facilitating the allocation of the whole kernel weights in Block-RAMs (BRAMs). Moreover, we used 8-bits kernel weights, rounded from an already trained CNN, to further reduce the need for memory, storing them in multiple BRAMs to aid kernel loading throughput. To balance the pipeline of convolutions through the convolutional layers we manipulated the amount of parallel computation in the convolutional step in each convolutional layer. We adopted the AlexNet CNN architecture to run our experiments and compare the results. We were able to run the inference of the convolutional layers in 3.9 ms with maximum operation frequency of 76.9 MHz.