A framework with cloud integration for CNN acceleration on FPGA devices

The recent years have seen a rapid diffusion of deep learning algorithms as Convolutional Neural Networks (CNNs), and as a consequence, an intensification of industrial and academic research focused on optimizing their implementation. Different computing architectures have been explored, and among all of them, FPGAs seem to be a very attractive choice, since they can deliver sustained performance with high power efficiency, as CNNs can be directly mapped onto hardware, and still offer flexibility thanks to their programmability. In this paper, we present an end-to-end framework to implement CNNs using a dataflow acceleration methodology. The resulting spatial accelerator can be scaled in size if enough resources are available and can exploit both intra- and inter-layers parallelism. We integrate the proposed framework with the deep learning engine Caffe, meaning that we are able to generate the accelerator starting from a Caffe model. We also provide cloud integration of such framework, enabling users to synthesize and deploy the accelerator on the Amazon F1 instances.