NeuFlow is a new concept of "dataflow" architecture, which is particularly well-suited for algorithms that perform a fixed set of operations on a stream of data (e.g. an image). NeuFlow is particularly efficient for such vision algorithms as Convolutional Networks. The NeuFlow architecture is currently instantiated on an FPGA board built around a Xilinx Virtex-6, which communicate with a laptop computer through a gigabit ethernet connection. It is capable of a sustained performance of 100 billion multiply-accumulate operations per second while consuming less than 15 Watts of power: about 100 times faster than on a conventional processor, and considerably faster than GPUs for a fraction of the power consumtion and a fraction of the volume. The system runs a number of real-time vision demos, such as a face detector, a general object recognition systems (trainable on-line), a pedestrian detector, and a vision system for off-road mobile robots that can classify obstacles from traversable areas. The system can also be trained on-line to recognize just about anything.