This work presents a parametrizable design of a neural network on an FPGA, being trained previously in the CPU using backpropagation. We test different configurations and report delay, power and area.